Method and system for marine catastrophe risk prediction based on machine reinforcement learning

Abstract

This invention discloses a method and system for predicting marine disaster risks based on machine reinforcement learning. The method includes collecting historical and real-time marine monitoring data and aligning them spatiotemporally; calculating the forced field gradient; dividing the historical marine monitoring data into first and second data sets according to a random ratio; correcting the energy flux using the forced field gradient in the first data set; constructing a disaster-causing factor; building an XGBoost prediction model; processing the second data set and inputting it into the XGBoost prediction model to obtain classification result probabilities and calculate a weighted anomaly index; constructing a stratified threshold set; processing real-time marine monitoring data and inputting it into the XGBoost prediction model to obtain real-time classification result probabilities; calculating a real-time weighted anomaly index; and comparing the stratified threshold set with the real-time weighted anomaly index to obtain real-time stratified early warning signals. This method can provide faster and more accurate marine disaster prediction results, while also having good interpretability, and can be directly applied to marine disaster risk prediction systems.

Description

Technical Field

[0001] This invention relates to the field of marine disaster early warning and risk prediction technology, and in particular to a marine disaster risk prediction method and system based on machine reinforcement learning. Background Technology

[0002] Marine disasters pose a serious threat to the lives and property of people in coastal areas and to marine economic activities. Accurate and efficient marine disaster risk prediction is the core foundation of disaster prevention and mitigation efforts. Existing marine disaster early warning methods mainly rely on empirical statistics or single-model extrapolation based on single types of data sources, lacking the ability to systematically and jointly represent multi-interface physical processes. Purely data-driven machine learning models cannot guarantee that the prediction results conform to the basic laws of ocean dynamics, while the computational efficiency of purely physical models cannot meet the needs of real-time early warning, resulting in insufficient reliability of early warning results under extreme sea conditions.

[0003] In existing disaster early warning technologies, the construction of disaster-causing factors generally adopts linear weighting or empirical threshold combination methods, which fail to fully consider the non-uniform modulation effect of the forced field on energy flux, and also make it difficult to simultaneously couple multiple physical field variables such as sea surface state and seabed state, thus limiting the accuracy of the characterization of disaster evolution process; after the disaster type is identified, a unified prediction architecture is adopted, which lacks the ability to make differentiated predictions for different disaster physical mechanisms; the setting of early warning thresholds mostly depends on historical statistical quantiles or expert experience, which cannot be adaptively optimized and is difficult to quantify the asymmetric cost difference between missed and false alarms. Therefore, this invention proposes a marine disaster risk prediction method and system based on machine reinforcement learning. This method uses historical disaster values ​​corresponding to a given state as the true intensity and predicted output values ​​corresponding to candidate warning actions as the predicted intensity. A regression penalty term is constructed based on the fact that the true intensity is higher than the predicted intensity. A weighted anomaly index penalty term is constructed based on the weighted anomaly index corresponding to the state and a preset weighted anomaly index. Based on the asymmetric classification error cost, this invention constructs an energy flux coupling method with forced field gradient correction, designs a physically constrained embedded XGBoost prediction model architecture, extracts energy fingerprint feature sets, and constructs a hierarchical threshold set based on Markov decision back mapping, achieving a deep integration of physical mechanisms and data-driven approaches. This method combines multi-source physical field collaborative representation with decision optimization, solving the industry pain points of empirical construction of disaster-causing factors and static, non-adaptive warning thresholds in existing technologies. It is of great significance for improving the accuracy, timeliness, and engineering practicality of marine disaster warnings. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting marine disaster risks based on machine reinforcement learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Historical and real-time ocean monitoring data were collected and spatiotemporally aligned. The gradient of the forcing field was calculated, and the historical ocean monitoring data was divided into first data and second data according to a random ratio.

[0008] In the first data, forced field gradient correction of energy flux is used, and disaster-causing factors are constructed by coupling sea surface state and seabed state.

[0009] Construct an XGBoost prediction model, process the second data and input it into the XGBoost prediction model to obtain the classification result probability and calculate the weighted anomaly index, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set.

[0010] Real-time marine monitoring data is processed and input into the XGBoost prediction model to obtain the probability of real-time classification results. The real-time weighted anomaly index is calculated, and the stratified threshold set is compared with the real-time weighted anomaly index to obtain real-time early warning stratified signals.

[0011] The forced field gradient includes the wind energy gradient, the pressure gradient, and the tsunami wave ray path density gradient.

[0012] The energy fluxes include wind energy flux, latent heat flux, and tsunami wave energy flux;

[0013] The disaster-causing factors include a first disaster-causing factor, a second disaster-causing factor, and a third disaster-causing factor;

[0014] The historical data includes wind field, pressure field, geological source information, sea surface condition, and seabed condition; the sea surface condition includes sea surface roughness, eddy current, divergence, and sea surface temperature; the seabed condition includes seabed roughness, turbulent dissipation rate, bottom boundary layer flow regime, and seabed temperature.

[0015] Furthermore, the method for constructing the hazard-causing factor includes:

[0016] The wind speed gradient is used as the wind stress enhancement factor, and the wind energy flux is scaled and adjusted by a product. The wind speed gradient and pressure gradient are used together as thermodynamic adjustment factors, and the latent heat flux is scaled and adjusted by a two-factor product. The tsunami wave ray path density gradient is used as the diffusion adjustment factor, and the tsunami wave energy flux is scaled and adjusted by an exponential decay correction.

[0017] The sea surface roughness and absolute value of eddy current are harmonic averaged to obtain the basic wind stress quantity. The divergence is used as a gated symbol variable, and the basic wind stress quantity is asymmetrically modulated by a sigmoid-type gated function to construct the first disaster-causing factor.

[0018] The sea surface temperature and the absolute value of divergence are inversely proportionally saturated and coupled to obtain the dynamic convergence enhancement coefficient. A vorticity inverse suppression term is constructed from the vorticity square term. The dynamic convergence enhancement coefficient is then adjusted inversely to construct the second disaster factor.

[0019] By saturating and mapping the bed surface roughness to the bottom boundary flow regime, a basic term for bottom boundary energy dissipation is obtained. A secondary acceleration dissipation term is constructed using the turbulent dissipation rate. The basic term for bottom boundary energy dissipation and the secondary acceleration dissipation term are multiplied to construct a triple product exponential decay kernel. The seabed temperature is used as a compensation term to adjust the logarithmic gain of the decay result and construct a third disaster factor.

[0020] Furthermore, the method for constructing the XGBoost prediction model includes:

[0021] The disaster-causing factors were randomly divided into a training set and a test set in a 7:3 ratio to train and test the XGBoost prediction model; the XGBoost prediction model includes an input layer, a discriminant layer, a prediction layer and an output layer;

[0022] The discriminant layer unfolds the disaster-causing factors in chronological order and extracts the energy fingerprint feature set, which is then input into the XGBoost multi-classifier to output the disaster type discrimination result.

[0023] The prediction layer allocates disaster-causing factors to preset prediction sub-modules according to the disaster type discrimination results, and outputs disaster risk prediction results; the prediction sub-modules include typhoon prediction sub-module, tsunami prediction sub-module, storm surge prediction sub-module, typhoon-storm prediction sub-module and calm period prediction sub-module.

[0024] The XGBoost prediction model improves prediction accuracy through mean squared error loss term and physical constraint loss term; the XGBoost prediction model uses an adaptive momentum estimation algorithm to optimize model hyperparameters;

[0025] The physical constraint loss term uses the gradient wind balance equation to constrain the typhoon prediction submodule, the tidal theory equation to constrain the tsunami prediction submodule, the tidal-storm surge nonlinear equation to constrain the storm surge prediction submodule, and the wind stress curl coupled water level response equation to constrain the typhoon-storm surge prediction submodule.

[0026] Furthermore, the method for constructing the energy fingerprint feature set includes:

[0027] The energy fingerprint feature set includes frequency domain fingerprint, evolution rate fingerprint, wave morphology fingerprint, and coupling correlation fingerprint;

[0028] The divergence gating term is used to calculate the physical gated energy concentration ratio, construct the physical gated frequency scale, and calculate the physical gated multi-scale entropy rate. The minimum point of the physical gated multi-scale entropy rate and the physical gated energy concentration ratio together constitute the frequency domain fingerprint vector.

[0029] The recursive matrix is ​​obtained by reconstructing the phase space of the time series. The physical damping coefficient is calculated by using the vortex inverse suppression term, and the physical damping evolution weighted index is constructed. The physical damping weighted determinism is calculated based on the recursive matrix. The evolution rate fingerprint vector is formed by the physical damping weighted evolution index and the physical damping weighted determinism.

[0030] Hilbert transforms are performed on the time series of the first and second disaster-causing factors respectively, and the physical coupling index and non-uniform physical phase coupling strength of the disaster-causing factors are calculated simultaneously to determine the coupling correlation fingerprint vector.

[0031] Homology analysis is performed on the time series to obtain the persistence of connected components. The physical decay weighted persistence energy entropy is calculated. The reciprocal of the triple product exponential decay kernel is used as the physical decay time scale. The time position corresponding to the maximum persistence is divided by the negative value of the physical decay time scale as the decay weight of the exponent. The exponent is then multiplied by the maximum persistence and the total energy of the time-domain waveform to determine the physical decay topology-energy coupling index and obtain the waveform morphology fingerprint vector.

[0032] Furthermore, the method for constructing the hierarchical threshold set includes:

[0033] The range of the disaster-causing factor is used as the basic dimension of the state space. The continuous range of the disaster-causing factor is discretized into multiple level intervals and combined to construct a high-dimensional joint state space and an asymmetric cost function.

[0034] Using a high-dimensional joint state space as the state space and a preset set of warning level actions as the action space, the optimal strategy is solved.

[0035] The optimal strategy is back-mapped to the value space of the disaster-causing factors, the state corresponding to the optimal action is marked, the level interval boundary values ​​of each disaster-causing factor in the state are extracted, the physical quantity thresholds corresponding to each disaster-causing factor under different warning levels are determined, and a hierarchical threshold set is constructed.

[0036] Furthermore, the method for constructing the asymmetric cost function is as follows:

[0037] Obtain the disaster discrimination probability output by the XGBoost prediction model, and calculate the weighted anomaly index based on the disaster-causing factors and the disaster discrimination probability;

[0038] The disaster category in the high-dimensional joint state space is taken as the state, and the warning level in the preset warning level action set is taken as the candidate warning action to construct an asymmetric classification error cost term.

[0039] The historical disaster values ​​corresponding to the state are used as the true intensity, and the predicted output values ​​corresponding to the candidate warning actions are used as the predicted intensity to construct a regression penalty term.

[0040] Based on the weighted anomaly index and the preset weighted anomaly index, a weighted anomaly index penalty term is constructed;

[0041] An asymmetric cost function is constructed based on the asymmetric classification error cost matrix, regression penalty term, and weighted anomaly index penalty term.

[0042] Secondly, a marine disaster risk prediction system based on machine reinforcement learning includes:

[0043] Data processing module: used to collect historical and real-time marine monitoring data, perform spatiotemporal alignment, calculate the forced field gradient, and divide the historical marine monitoring data into first data and second data according to a random ratio;

[0044] Disaster Factor Module: Used to correct energy flux using forced field gradient in the first data and to construct disaster factors by coupling sea surface state and seabed state;

[0045] Prediction Model Module: Used to build XGBoost prediction models, process the second data input to the XGBoost prediction model, and obtain the model output results;

[0046] The hierarchical threshold set module is used to calculate the weighted anomaly index based on the classification result probability obtained from the model output, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set.

[0047] Early warning output module: It is used to process real-time marine monitoring data, input it into the XGBoost prediction model, obtain the probability of real-time classification results, calculate the real-time weighted anomaly index, compare the stratified threshold set with the real-time weighted anomaly index, and obtain real-time early warning stratified signals.

[0048] The beneficial effects of this invention are:

[0049] This invention relates to a method and system for predicting marine disaster risks based on machine reinforcement learning. Compared with existing technologies, this invention has the following technical advantages:

[0050] This invention overcomes the problem of missing multi-physics field coordination by constructing a disaster-causing factor by forcing the energy flux through the gradient correction of the field and coupling the sea surface and seabed states, and realizes the systematic characterization of multi-interface physical processes.

[0051] This invention enhances the model's ability to capture differentiated features of different disaster states and improves the sensitivity of disaster identification by extracting an energy fingerprint feature set that integrates physical gating and topological features.

[0052] This invention constructs a hierarchical threshold set based on Markov decision inverse mapping, and quantifies the cost difference between false alarms and missed alarms using an asymmetric cost function, thereby achieving dynamic optimization and hierarchical accurate output of early warning thresholds. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the steps of the marine disaster risk prediction method based on machine reinforcement learning according to the present invention. Detailed Implementation

[0054] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0055] The present invention provides a method and system for predicting marine disaster risks based on machine reinforcement learning, comprising the following steps:

[0056] like Figure 1 As shown, this embodiment includes the following steps:

[0057] Historical and real-time ocean monitoring data were collected and spatiotemporally aligned. The gradient of the forcing field was calculated, and the historical ocean monitoring data was divided into first data and second data according to a random ratio.

[0058] In the first data, forced field gradient correction of energy flux is used, and disaster-causing factors are constructed by coupling sea surface state and seabed state.

[0059] Construct an XGBoost prediction model, process the second data and input it into the XGBoost prediction model to obtain the classification result probability and calculate the weighted anomaly index, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set.

[0060] Real-time marine monitoring data is processed and input into the XGBoost prediction model to obtain the probability of real-time classification results. The real-time weighted anomaly index is calculated, and the stratified threshold set is compared with the real-time weighted anomaly index to obtain real-time early warning stratified signals.

[0061] The forced field gradient includes the wind energy gradient, the pressure gradient, and the tsunami wave ray path density gradient.

[0062] The energy fluxes include wind energy flux, latent heat flux, and tsunami wave energy flux;

[0063] The disaster-causing factors include a first disaster-causing factor, a second disaster-causing factor, and a third disaster-causing factor;

[0064] The historical data includes wind field, pressure field, geological source information, sea surface condition, and seabed condition; the sea surface condition includes sea surface roughness, eddy current, divergence, and sea surface temperature; the seabed condition includes seabed roughness, turbulent dissipation rate, bottom boundary layer flow regime, and seabed temperature.

[0065] In this embodiment, the method for constructing the hazard factor includes:

[0066] The wind speed gradient is used as the wind stress enhancement factor, and the wind energy flux is scaled and adjusted by a product. The wind speed gradient and pressure gradient are used together as thermodynamic adjustment factors, and the latent heat flux is scaled and adjusted by a two-factor product. The tsunami wave ray path density gradient is used as the diffusion adjustment factor, and the tsunami wave energy flux is scaled and adjusted by an exponential decay correction.

[0067] The sea surface roughness and absolute value of eddy current are harmonic averaged to obtain the basic wind stress quantity. The divergence is used as a gated symbol variable, and the basic wind stress quantity is asymmetrically modulated by a sigmoid-type gated function to construct the first disaster-causing factor.

[0068] The sea surface temperature and the absolute value of divergence are inversely proportionally saturated and coupled to obtain the dynamic convergence enhancement coefficient. A vorticity inverse suppression term is constructed from the vorticity square term. The dynamic convergence enhancement coefficient is then adjusted inversely to construct the second disaster factor.

[0069] By saturating and mapping the bed surface roughness to the bottom boundary flow regime, the basic term of bottom boundary energy dissipation is obtained. The secondary acceleration dissipation term is constructed using the turbulent dissipation rate. The triple product exponential decay kernel is constructed by multiplying the basic term of bottom boundary energy dissipation and the secondary acceleration dissipation term. The logarithmic gain of the decay result is adjusted by using seabed temperature as a compensation term to construct the third disaster factor.

[0070] In actual assessments, the expression for the first hazard-causing factor is:

[0071]

[0072] in, As the primary cause of disaster, This is the corrected wind energy flux. This is the wind stress adjustment coefficient. For sea surface roughness, vorticity, is the kurtosis coefficient of the gated function. Let be the divergence, and satisfy . ;

[0073] The expression for the second hazard factor is:

[0074]

[0075] in, As the second hazard factor, This is the corrected latent heat flux. The positive adjustment enhancement coefficient, Sea surface temperature, To adjust the inhibition coefficient in the reverse direction, and satisfy... ;

[0076] The expression for the third hazard factor is:

[0077]

[0078] in, As the third hazard factor, This is the corrected tsunami wave energy flux. The coefficient of friction is adjusted to the lowest level. For bed surface roughness, For characteristic length, The Reynolds number of the bottom boundary layer. The dissipation acceleration adjustment coefficient, For turbulent dissipation rate, The baseline value for turbulent dissipation rate, This is the temperature compensation coefficient. For seabed temperature, Characteristic temperature;

[0079] When the seabed is rough, the seabed flow is turbulent, and the dissipation rate increases, the absolute value of the exponential decay kernel increases, and the energy flux of the tsunami wave will be rapidly reduced, causing the third catastrophic factor to significantly approach 0.

[0080] In this embodiment, the method for constructing the XGBoost prediction model includes:

[0081] The disaster-causing factors were randomly divided into a training set and a test set in a 7:3 ratio to train and test the XGBoost prediction model; the XGBoost prediction model includes an input layer, a discriminant layer, a prediction layer and an output layer;

[0082] The discriminant layer unfolds the disaster-causing factors in chronological order and extracts the energy fingerprint feature set, which is then input into the XGBoost multi-classifier to output the disaster type discrimination result.

[0083] The prediction layer allocates disaster-causing factors to preset prediction sub-modules according to the disaster type discrimination results, and outputs disaster risk prediction results; the prediction sub-modules include typhoon prediction sub-module, tsunami prediction sub-module, storm surge prediction sub-module, typhoon-storm prediction sub-module and calm period prediction sub-module.

[0084] The XGBoost prediction model improves prediction accuracy through mean squared error loss term and physical constraint loss term; the XGBoost prediction model uses an adaptive momentum estimation algorithm to optimize model hyperparameters;

[0085] The physical constraint loss term uses the gradient wind balance equation to constrain the typhoon prediction submodule, the tidal theory equation to constrain the tsunami prediction submodule, the tidal-storm surge nonlinear equation to constrain the storm surge prediction submodule, and the wind stress curl coupled water level response equation to constrain the typhoon-storm surge prediction submodule.

[0086] In actual evaluation, the physical constraint loss term of the typhoon prediction submodule is:

[0087]

[0088] in, This is the physical constraint loss term for the typhoon forecasting submodule. To predict the tangential wind speed of a typhoon, This is the radius distance from the center of the typhoon. Coriolis parameters, Atmospheric density, To predict atmospheric pressure;

[0089] The physical constraint loss term of the tsunami prediction submodule is:

[0090]

[0091] in, This is the physical constraint loss term for the tsunami prediction submodule. To predict tsunami wave amplitude, The initial tsunami wave amplitude, The wave height attenuation coefficient is... The equivalent bottom friction coefficient, The distance the tsunami traveled. To predict the water depth at the location;

[0092] The physical constraint loss term of the storm surge prediction submodule is:

[0093]

[0094] in, This is the physical constraint loss term for the storm surge prediction submodule. To predict the total water level, For time variables, The radiation stress-water level coupling coefficient is... Let be the wave radiation stress tensor. This is the nonlinear interaction term between tides and storm surges;

[0095] The physical constraint loss term of the typhoon-storm surge prediction submodule is:

[0096]

[0097] in, This is the physical constraint loss term for the typhoon-storm surge prediction submodule. To predict storm surge, It is a time-varying transfer function. The wind stress vector For curl, The wind stress curl rate of change - water level coupling coefficient.

[0098] In this embodiment, the method for constructing an energy fingerprint feature set includes:

[0099] The energy fingerprint feature set includes frequency domain fingerprint, evolution rate fingerprint, wave morphology fingerprint, and coupling correlation fingerprint;

[0100] The divergence gating term is used to calculate the physical gated energy concentration ratio, construct the physical gated frequency scale, and calculate the physical gated multi-scale entropy rate. The minimum point of the physical gated multi-scale entropy rate and the physical gated energy concentration ratio together constitute the frequency domain fingerprint vector.

[0101] The recursive matrix is ​​obtained by reconstructing the phase space of the time series. The physical damping coefficient is calculated by using the vortex inverse suppression term, and the physical damping evolution weighted index is constructed. The physical damping weighted determinism is calculated based on the recursive matrix. The evolution rate fingerprint vector is formed by the physical damping weighted evolution index and the physical damping weighted determinism.

[0102] Hilbert transforms are performed on the time series of the first and second disaster-causing factors respectively, and the physical coupling index and non-uniform physical phase coupling strength of the disaster-causing factors are calculated simultaneously to determine the coupling correlation fingerprint vector.

[0103] Homology analysis is performed on the time series to obtain the persistence of connected components. The physical decay weighted persistence energy entropy is calculated. The reciprocal of the triple product exponential decay kernel is used as the physical decay time scale. The time position corresponding to the maximum persistence is divided by the negative value of the physical decay time scale as the decay weight of the exponent. The exponent is then multiplied by the maximum persistence and the total energy of the time-domain waveform to determine the physical decay topology-energy coupling index and obtain the waveform morphology fingerprint vector.

[0104] In practical evaluation, the expression for the physical gating energy concentration ratio is:

[0105]

[0106] in, The concentration ratio of physically gated energy. The energy concentration ratio is the physical modulation coefficient. For frequency scale, The frequency scale is a physically gated correction. For frequency sub-band index, To correct the frequency scale of the first The energy percentage of each sub-band;

[0107] The expression for the physical damping weighted evolution index is:

[0108]

[0109] in, The physical damping weighted evolution index. The physical damping coefficient, For discrete-time sampling point index, This represents the time-series value of the current disaster-causing factor being processed. The average asymmetric index during historical periods of calm;

[0110] The frequency domain fingerprint vector, evolution rate fingerprint vector, coupling correlation fingerprint vector, and waveform morphology fingerprint vector are concatenated to construct an energy fingerprint feature set.

[0111] In this embodiment, the method for constructing a hierarchical threshold set includes:

[0112] The range of the disaster-causing factor is used as the basic dimension of the state space. The continuous range of the disaster-causing factor is discretized into multiple level intervals and combined to construct a high-dimensional joint state space and an asymmetric cost function.

[0113] Using a high-dimensional joint state space as the state space and a preset set of warning level actions as the action space, the optimal strategy is solved.

[0114] The optimal strategy is back-mapped to the value space of the disaster-causing factors, the state corresponding to the optimal action is marked, the level interval boundary value of each disaster-causing factor in the state is extracted, the physical quantity threshold corresponding to each disaster-causing factor under different warning levels is determined, and a hierarchical threshold set is constructed.

[0115] In actual assessment, the first, second and third disaster-causing factors are divided into multiple level intervals according to preset energy levels and then combined to obtain a high-dimensional joint state space.

[0116] The high-dimensional joint state space is used as the state space, the pre-set set of warning level actions is used as the action space, and the negative value of the asymmetric cost function is used as the reward function. The optimal strategy is solved iteratively using the Bellman optimal equation.

[0117] The expression for the optimal strategy is:

[0118]

[0119] in, For state The optimal early warning strategy is as follows: For candidate early warning actions, For the action space, As a discount factor, For the state at the next moment, For the transition state probability, For state The optimal value function under the following conditions This is the function that returns the minimum value.

[0120] In this embodiment, the method for constructing the asymmetric cost function is as follows:

[0121] Obtain the disaster discrimination probability output by the XGBoost prediction model, and calculate the weighted anomaly index based on the disaster-causing factors and the disaster discrimination probability;

[0122] The disaster category in the high-dimensional joint state space is taken as the state, and the warning level in the preset warning level action set is taken as the candidate warning action to construct an asymmetric classification error cost term.

[0123] The historical disaster values ​​corresponding to the state are used as the true intensity, and the predicted output values ​​corresponding to the candidate warning actions are used as the predicted intensity to construct a regression penalty term.

[0124] Based on the weighted anomaly index and the preset weighted anomaly index, a weighted anomaly index penalty term is constructed;

[0125] Construct an asymmetric cost function based on the asymmetric classification error cost matrix, regression penalty term, and weighted outlier index penalty term;

[0126] In actual assessment, the disaster discrimination probability output by the XGBoost prediction model is obtained, and combined with the disaster-causing factor combination function of various disaster events, the weight coefficients of various disaster events are used as weighting factors. The products of the discrimination probability of various disaster events and the corresponding disaster-causing factor combination function are weighted and summed to obtain the weighted anomaly index.

[0127] The disaster category in the high-dimensional joint state space is taken as the state, the warning level in the preset warning level action set is taken as the candidate warning action, and an asymmetric classification error cost term is constructed based on the fact that the false negative rate is much greater than the false positive rate. The historical disaster value corresponding to the state is taken as the true intensity, and the predicted output value corresponding to the candidate warning action is taken as the predicted intensity. A regression penalty term is constructed based on the fact that the true intensity is higher than the predicted intensity. A weighted anomaly index penalty term is constructed based on the weighted anomaly index corresponding to the state and the preset weighted anomaly index. An asymmetric cost function is constructed based on the asymmetric classification error cost term, the regression penalty term, and the weighted anomaly index penalty term.

[0128] The formula for calculating the asymmetric cost function is as follows:

[0129]

[0130] in, For state Take action below The asymmetric cost function, The weighting coefficient for the cost of classification errors. To regress the underestimation of the penalty weight coefficient, The weighted abnormality index penalty coefficient is the coefficient. The cost of asymmetric classification error, To represent the true intensity of the disaster, Take an action in state s The disaster intensity prediction value output by the corresponding prediction submodule, The weighted anomaly index warning threshold. For state The weighted anomaly index.

[0131] In addition, this embodiment takes a certain sea area in location A as the monitoring target area, and selects the period from 00:00 on July 15, 2024 to 00:00 on July 16, 2024 as the monitoring period;

[0132] Wind and pressure field data were provided by three buoy-type meteorological observation stations deployed in the target sea area and satellite remote sensing wind field inversion data covering the area. Geological source information was supported by source parameter data released in real time by the China Earthquake Networks Center. During this period, a submarine earthquake with a magnitude of 7.2 occurred about 800 km southeast of the target sea area, with a focal depth of 15 km. Sea surface state data were collected in real time by wave meter, acoustic Doppler current profiler and infrared temperature sensor carried on the buoy. Seabed state data were obtained by the bottom boundary layer observation system deployed on the seabed.

[0133] Spatiotemporal alignment is performed using the main shaft synchronization time reference, resampling is performed to a uniform time resolution, and quality control and outlier removal are carried out. The preprocessed historical marine monitoring data is randomly divided into first data and second data. In this embodiment, the monitoring time point of 08:00 on July 15, 2024 is used as an example, and the first data is used to calculate the specific construction of disaster-causing factors.

[0134] Using the center point of the target sea area as the center, wind speed vector data from four surrounding grid points were selected. The central difference method was used to calculate the horizontal wind speed gradient and pressure gradient, respectively, yielding a horizontal wind speed gradient of 2.5 × and pressure gradient 1.2× Based on earthquake source parameters and measured water depth data of the target sea area, the ray tracing method was used to simulate the ray paths of tsunami waves propagating from the source outwards. The number of ray paths around the center point of the target sea area was counted as 42, with a ray path density of [missing value]. The tsunami wave ray path density gradient was calculated using the central difference method as follows: ;

[0135] The raw wind energy flux at the monitoring time was obtained from the monitoring data. Based on historical calibration experiments in the target sea area, the wind stress enhancement adjustment coefficient was set to 8.0. The wind speed gradient was used as the wind stress enhancement factor, and a product-based scaling adjustment was performed to obtain the corrected wind energy flux. ;

[0136] The raw latent heat flux obtained from the monitoring data is 120 at the monitoring time. Based on historical calibration experiments in the target sea area, the thermodynamic coefficient is taken as 8.0. Using both wind speed and pressure gradients as thermodynamic adjustment factors, and employing a two-factor product approach for scaling, the corrected latent heat flux is obtained as 122.88. ;

[0137] The raw tsunami wave energy flux at the monitoring time was 150, obtained from the monitoring data. The diffusion regulation coefficient for the target sea area was set to 1.0 based on historical calibration experiments. Using the tsunami wave ray path density gradient as a diffusion adjustment factor, and employing an exponential decay correction method for scaling adjustment, the corrected tsunami wave energy flux was obtained as 33.47. ;

[0138] Based on the sea surface state data observed by the buoy at the stated time: the sea surface roughness was obtained as 0.035 m from wave spectrum inversion, and the vorticity and divergence were obtained as 1.2 m and 1.2 m respectively from the acoustic Doppler current profiler. and -3.0 By harmonic averaging the absolute values ​​of sea surface roughness and eddy current, the basic wind stress value was obtained as 2.392. Based on historical calibration experiments of the target sea area, the wind stress adjustment coefficient was set to 0.01, and the first disaster-causing factor was found to be 0.659.

[0139] Based on the sea surface condition data observed by the buoy at that moment, the sea surface temperature was found to be 28.5 degrees Celsius, and the vorticity and divergence were both 1.2. and -3.0 By inversely proportionally saturating and coupling sea surface temperature with the absolute value of divergence, a dynamic convergence enhancement coefficient of 0.0285 was obtained. Based on historical calibration experiments in the target sea area, a reverse regulation inhibition coefficient of 5.0 was determined. The second hazard factor was found to be 3.502.

[0140] Based on the seabed condition data observed by the buoy at the monitoring time, the surface roughness was found to be 0.2, and the turbulent dissipation rate was... With a seabed temperature of 4.5 degrees Celsius, the surface roughness and bottom boundary flow regime are coupled by saturation mapping to obtain a basic term of bottom boundary energy dissipation of 0.998, a secondary acceleration dissipation term of 2.2, and a triple product exponential decay kernel of 0.1108. The decay result is adjusted logarithmically by using seabed temperature as a state compensation term, and the third disaster factor is obtained as 4.035.

[0141] In the discriminant layer, the time series data of disaster-causing factors output from the input layer are expanded in chronological order to extract the energy fingerprint feature set. The extracted energy fingerprint feature set is then input into the XGBoost multi-classifier. Through ensemble learning of multiple decision trees, the energy fingerprint feature set is subjected to nonlinear transformation and classification decision. The category with the highest probability is taken as the preliminary disaster type discrimination result.

[0142] The XGBoost prediction model uses an adaptive momentum estimation algorithm to optimize the model's hyperparameters. By calculating the first and second moment estimates of the gradient, it adaptively adjusts the learning rate of each hyperparameter, enabling the model to converge to the optimal solution efficiently during training.

[0143] The divergence gating term is used to physically gate the time series of disaster-causing factors. The divergence value at each time point is mapped to the interval between 0 and 1 through the Sigmoid gating function to obtain the physical gating weight. The gating weight is less than 0.5 under convergence conditions to suppress the energy ratio, and greater than 0.5 under divergence conditions to enhance the energy ratio, thereby achieving asymmetric modulation under different physical conditions.

[0144] Wavelet packet decomposition is performed on the time series of disaster-causing factors to obtain the energy proportion of different frequency sub-bands. The energy proportion of each sub-band is calculated by wavelet packet coefficient energy normalization.

[0145] The physical gated energy concentration ratio was calculated using the divergence gating term. The physical modulation coefficient of the energy concentration ratio was taken as 0.3 based on the historical calibration experiment of this sea area, and the calculated physical gated energy concentration ratio was approximately 1.42.

[0146] The time series was coarse-grained under different scale factors (values ​​from 1 to 20), and the sample entropy value at each scale was calculated. The entropy rate was weighted and modulated using a divergence gating term to obtain the physical gated multi-scale entropy rate. The time scale corresponding to the global minimum point was calculated to be 3, indicating that the signal has the lowest entropy rate at time scale 3, which corresponds to a specific periodicity feature.

[0147] The minimum point value of the entropy rate of the physical gated multiscale system and the value of the energy concentration ratio of the physical gated system are used to construct the frequency domain fingerprint vector, which is equal to [3, 1.42].

[0148] For the evolution rate fingerprint vector, the phase space is first reconstructed. Using Takens' embedding theorem, the embedding dimension is set to 3 and the time delay is set to 1. The one-dimensional time sequence of 60 sampling points is reconstructed into 58 three-dimensional phase points, and the recursion matrix is ​​calculated to obtain a 58-by-58 recursion matrix.

[0149] The physical damping coefficient is calculated using the vortex reverse suppression term. The physical damping coefficient is equal to 1 minus the reverse regulation suppression coefficient multiplied by the square of the vortex, and the calculated physical damping coefficient is approximately 0.98.

[0150] The physical damping weighted evolution index was constructed, and the algebraic sum of the third-order differences of the first-order differences of the time series was calculated and divided by the sum of the absolute values ​​of the third-order differences. The calculated ratio was 0.38. The average asymmetric index of the historical quiet period was statistically taken as 0.12 based on historical data. The physical damping weighted evolution index was calculated to be approximately equal to 0.3800.

[0151] The physical damping weighted determinism is calculated based on the recursion matrix. The total number of line segments with a diagonal length greater than or equal to 2 in the recursion matrix is ​​156, and the sum of all elements in the recursion matrix is ​​1240. Considering the histogram weighting of the diagonal distribution, the physical damping weighted determinism is calculated to be 0.682.

[0152] The physical damping weighted evolution index and the physical damping weighted determinism together constitute the evolution rate fingerprint vector, which is equal to [0.3800, 0.682].

[0153] For the coupled fingerprint vector, take a 1-hour time series segment of the first and second disaster-causing factors, perform Hilbert transform on the two sequences respectively to obtain their respective analytical signals, and extract the instantaneous amplitude and instantaneous phase;

[0154] The amplitude correlation coefficient is obtained by dividing the time average of the product of the instantaneous amplitudes of the two sequences by the square root of the product of the squared time averages of their respective instantaneous amplitudes, and the calculated amplitude correlation coefficient is 0.72; the phase synchronization index is the absolute value of the average of the phase difference indexes of the two sequences at each time point, and the calculated synchronization index is 0.65; the physical coupling index is the average of the amplitude correlation coefficient 0.72 and the phase synchronization index 0.65, which is 0.685.

[0155] Mapping the phase difference to arrive Within the interval, the phase difference distribution was statistically analyzed and the circular entropy was calculated. The circular entropy was found to be 1.28, the maximum entropy value was approximately 1.838, and the non-uniform physical phase coupling strength was 0.58.

[0156] The physical coupling index and the non-uniform physical phase coupling strength together constitute the coupling correlation fingerprint vector, which is equal to [0.685, 0.58].

[0157] For waveform morphology fingerprint vectors, taking the time series of the third catastrophic factor as an example, the upper level set method is used for homology analysis to record the occurrence and merging times of connected components in the waveform. A total of 8 connected components are identified, and the persistence of the largest connected component is 23 (sampling point interval, corresponding to a duration of 23 seconds). The birth time is the 18th sampling point.

[0158] The reciprocal of the triple product exponential decay kernel is used as the physical decay time scale. The physical decay weighted persistence probability distribution is normalized by multiplying each persistence by the exponential decay term to obtain the physical decay weighted persistence energy entropy, which is approximately 1.84. The total energy of the time-domain waveform is the sum of the squares of the values ​​at each time step of the time series, which is calculated to be 0.0729.

[0159] The waveform morphology fingerprint vector is constructed by combining the physical decay weighted persistence energy entropy and the physical decay topology-energy coupling index. The waveform morphology fingerprint vector is equal to [1.84, 0.0729].

[0160] The above fingerprint vectors are concatenated to construct a complete energy fingerprint feature set. The concatenated feature set is an 8-dimensional feature vector, and the energy fingerprint feature set is equal to [3, 1.42, 0.3800, 0.682, 0.685, 0.58, 1.84, 0.0729].

[0161] Taking 08:00 on July 15, 2024 as an example, the probability of various disasters output by the XGBoost multi-classifier is as follows: typhoon probability 0.35, tsunami probability 0.28, storm surge probability 0.22, typhoon-storm surge combined probability 0.10, and calm period probability 0.05;

[0162] The probability of various disasters output by the XGBoost prediction model is multiplied by the combined function value of the causative factors corresponding to each disaster event (the combined function of the causative factors of each disaster is composed of the first causative factor, the second causative factor, and the third causative factor, which are weighted according to the weights obtained by the least squares method based on historical disaster data). The sum of the products is divided by the sum of the probabilities to obtain the weighted anomaly index. The weighted anomaly index at this moment is calculated to be 2.5761.

[0163] The disaster category in the high-dimensional joint state space is taken as state s, and the warning level in the preset warning level action set is taken as candidate warning action a. Based on the prior principle that the false negative rate is much greater than the false positive rate, an asymmetric classification error cost term is constructed.

[0164] Taking a specific state s (disaster type is typhoon, and the states of disaster-causing factors are H1=0.659, H2=3.502, and H3=4.03) as an example, the asymmetric classification error cost is set as follows when different early warning actions are taken:

[0165] When the actual disaster type is a typhoon, if action a1 (normal, no warning) is taken, it is a serious underreporting situation, and the cost of underreporting is set at 10.0. If action a2 (blue warning) is taken, it is a minor underreporting situation (the warning level is insufficient), and the cost is set at 6.0. If action a3 (yellow warning) is taken, it is a basic match but slightly lower, and the cost is set at 3.0. If action a4 (red warning) is taken, it is a correct warning (typhoon red warning), and the cost is set at 0.5 (there is a slight possibility of false alarm).

[0166] Based on the fact that the actual intensity is higher than the predicted intensity, a regression penalty term is constructed. Taking 08:00 on July 15, 2024 as an example, the actual disaster at that time was a typhoon. The actual maximum wind speed of the typhoon was 42.5 meters per second. When action a1 (normal) is taken, the predicted wind speed output by the model is 35.0 meters per second. When action a2 (blue warning) is taken, the predicted wind speed is 38.2 meters per second. When action a3 (yellow warning) is taken, the predicted wind speed is 40.1 meters per second. When action a4 (red warning) is taken, the predicted wind speed is 42.5 meters per second.

[0167] The weighted anomaly index penalty term is calculated to be 0.3267 based on the weighted anomaly index corresponding to the state and the preset weighted anomaly index warning threshold.

[0168] Calculate the asymmetric cost for each action: the asymmetric cost for action a1 is 6.223, the asymmetric cost for action a2 is 3.4678, the asymmetric cost for action a3 is 1.7132, and the asymmetric cost for action a4 is 0.348. Compare the costs of each action. The cost of taking action a4 (red alert) in state s is minimized, which is consistent with the direction of solving the optimal policy.

[0169] The continuous value range of each disaster-causing factor is discretized into multiple level intervals according to a preset energy level. In this embodiment, each disaster-causing factor is divided into 5 level intervals, with the first disaster-causing factor... The value range is from 0.1 to 1.2. Level 1 is the lowest level, with a value greater than or equal to 0.1 and less than 0.3; Level 2 is the lowest level, with a value greater than or equal to 0.3 and less than 0.5; Level 3 is the middle level, with a value greater than or equal to 0.5 and less than 0.7; Level 4 is the highest level, with a value greater than or equal to 0.7 and less than 0.9; and Level 5 is the highest level, with a value greater than or equal to 0.9 and less than or equal to 1.2.

[0170] Second disaster-causing factor The value range is from 0.5 to 6.0, and it is also divided into 5 level intervals: Level 1 is the very low level, with a value greater than or equal to 0.5 and less than 1.2; Level 2 is the low level, with a value greater than or equal to 1.2 and less than 2.2; Level 3 is the medium level, with a value greater than or equal to 2.2 and less than 3.4; Level 4 is the high level, with a value greater than or equal to 3.4 and less than 4.6; and Level 5 is the very high level, with a value greater than or equal to 4.6 and less than or equal to 6.0.

[0171] The third disaster-causing factor The value range is from 0.5 to 25.0, divided into 5 level intervals: Level 1 is the very low level, with a value greater than or equal to 0.5 and less than 2.0; Level 2 is the low level, with a value greater than or equal to 2.0 and less than 5.0; Level 3 is the medium level, with a value greater than or equal to 5.0 and less than 10.0; Level 4 is the high level, with a value greater than or equal to 10.0 and less than 17.0; and Level 5 is the very high level, with a value greater than or equal to 17.0 and less than or equal to 25.0.

[0172] Each disaster-causing factor is discretized into 5 level intervals. The levels of the three disaster-causing factors are combined to construct a high-dimensional joint state space with a total of 125 states.

[0173] The preset set of warning level actions is used as the action space. In this embodiment, four warning level actions are set: action a1 is normal (no warning), action a2 is blue warning (attention level), action a3 is yellow warning (alert level), and action a4 is red warning (alarm level).

[0174] The negative value of the asymmetric cost function is used as the reward function. The larger the value of the reward function, the more likely it is to be used in this state. This asymmetric cost function is used to quantify the cost of taking action a in state s. This cost function comprehensively considers three aspects: classification error cost, regression underestimation penalty, and outlier exponent penalty.

[0175] In specific state For example, this state indicates that the first disaster-causing factor is at a medium level, the second disaster-causing factor is at a high level, and the third disaster-causing factor is at a low level. The actual disaster intensity corresponding to this state is 1.8 (representing the height of storm surge, in meters). The disaster intensity prediction value output by the XGBoost prediction submodule is 1.2 meters, the weighted anomaly index is 0.65, and the weighted anomaly index warning threshold is 0.50.

[0176] The asymmetric classification error cost term is constructed based on the prior principle that the false negative rate is much greater than the false positive rate. When action a1 (normal, no warning) is taken and an actual disaster is about to occur, the false negative cost is set to 10.0. When action a4 (red warning) is taken but there is no significant disaster, the false positive cost is set to 2.0. For intermediate actions, the classification error cost is linearly interpolated between the two to calculate the state. The classification error cost of taking action a1 is 8.5;

[0177] The regression penalty term is used to penalize situations where the disaster intensity is underestimated. The actual intensity is 1.8 meters, and the predicted intensity is 1.2 meters. The actual intensity is higher than the predicted intensity, so an underestimation penalty needs to be calculated. The regression penalty term is the square of the difference between the actual intensity and the predicted intensity, which equals 0.36. If the predicted intensity is higher than the actual intensity (overestimation), then the penalty term is 0.

[0178] The weighted anomaly index penalty term is used to penalize situations where the weighted anomaly index exceeds the warning threshold. The weighted anomaly index is 0.65, the warning threshold is 0.50, the excess is 0.15, and the penalty term is the square of the excess, which is 0.0225.

[0179] The complete calculation of the asymmetric cost function is a weighted sum of three terms, which equals 4.3625.

[0180] If action a4 (red alert) is taken, the cost of classification error (false alarm cost) is 1.5, the regression penalty (assuming the predicted intensity is 1.5 meters, which is still lower than the actual intensity of 1.8 meters) is 0.09, and the weighted anomaly index penalty is 0.0225. The value is 0.7815. In this state, the cost of taking a red alert action is much lower than not issuing an alert. Therefore, this state tends to trigger a higher level of alert. The lower the cost of the action, the better the strategy.

[0181] The high-dimensional joint state space (125 states in total) is used as the state space, and the pre-set set of warning level actions (4 actions) is used as the action space. The negative value of the asymmetric cost function is used as the reward function. The Bellman optimal equation is iteratively executed for all 125 states to solve for the optimal policy, and the optimal warning action in each state is obtained, thus forming the optimal policy set.

[0182] The optimal strategy set is mapped inversely to the value space of the disaster-causing factors. Taking action a4 (red alert) in the optimal strategy as an example, all states with the optimal strategy of a4 are counted, and the boundary values ​​of the level intervals of each disaster-causing factor in the state are extracted to determine the physical quantity threshold conditions corresponding to the red alert. All states with the optimal strategy of a3 (yellow alert) are counted, and the physical quantity threshold conditions corresponding to the yellow alert are extracted. All states with the optimal strategy of a2 (blue alert) are counted, and the physical quantity threshold conditions corresponding to the blue alert are extracted. All states with the optimal strategy of a1 (normal) are counted, and the physical quantity threshold conditions corresponding to the normal state are extracted.

[0183] Organize the above conditions into a hierarchical threshold set and construct Table 1, the hierarchical threshold set table.

[0184] Table 1. Hierarchical Threshold Set Table

[0185]

[0186] The conditions for the disaster-causing factors in a blue alert are: Condition B1: H1 ≥ 0.5, H2 ≥ 1.2, and H3 ≥ 2.0; Condition B2: H1 ≥ 0.3, H2 ≥ 2.2, and H3 ≥ 2.0. The conditions for the disaster-causing factors in a yellow alert are: Condition Y1: H1 ≥ 0.7, H2 ≥ 2.2, and H3 ≥ 2.0; Condition Y2: H1 ≥ 0.5, H2 ≥ 3.4, and H3 ≥ 2.0; Condition Y3: H1 ≥ 0.5, H2 ≥ 3.4, and H3 ≥ 2.0. The conditions for the red alert disaster-causing factor combination are: Condition R1: H1≥0.7, H2≥3.4, and H3≥2.0; Condition R2: H1≥0.9, H2≥3.4, and H3≥2.0; Condition R3: H1≥0.9, H2≥4.6, and H3≥0.5; Condition R4: H1≥0.5, H2≥4.6, and H3≥5.0.

[0187] The constructed hierarchical threshold set is configured in the real-time early warning system. For newly arriving real-time marine monitoring data, after calculating the three types of disaster-causing factors at the current moment using the aforementioned method, the data is then judged level by level according to the condition order in the hierarchical threshold set:

[0188] The first step is to determine whether the conditions for a red alert are met. If they are met, a red alert signal is output.

[0189] The second step is to determine whether the conditions for a red alert are met if they are not met, and if they are met, to output a yellow alert signal.

[0190] The third step is to determine whether the conditions for a yellow warning are met if they are not met, and if they are, to output a blue warning signal.

[0191] Fourth step: If none of the above warning conditions are met, then a normal signal is output.

[0192] Secondly, a marine disaster risk prediction system based on machine reinforcement learning includes:

[0193] Data processing module: used to collect historical and real-time marine monitoring data, perform spatiotemporal alignment, calculate the forced field gradient, and divide the historical marine monitoring data into first data and second data according to a random ratio;

[0194] Disaster Factor Module: Used to correct energy flux using forced field gradient in the first data and to construct disaster factors by coupling sea surface state and seabed state;

[0195] Prediction Model Module: Used to build XGBoost prediction models, process the second data input to the XGBoost prediction model, and obtain the model output results;

[0196] The hierarchical threshold set module is used to calculate the weighted anomaly index based on the classification result probability obtained from the model output, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set.

[0197] Early warning output module: It is used to process real-time marine monitoring data, input it into the XGBoost prediction model, obtain the probability of real-time classification results, calculate the real-time weighted anomaly index, compare the stratified threshold set with the real-time weighted anomaly index, and obtain real-time early warning stratified signals.

[0198] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting marine disaster risk based on machine reinforcement learning, characterized in that, Includes the following steps: S1. Collect historical and real-time marine monitoring data and perform spatiotemporal alignment, calculate the forced field gradient, and divide the historical marine monitoring data into first data and second data according to a random ratio. S2. In the first data, the energy flux is corrected by a forced field gradient, and the disaster-causing factor is constructed by coupling the sea surface state and the seabed state. S3. Construct an XGBoost prediction model, process the second data and input it into the XGBoost prediction model to obtain the classification result probability and calculate the weighted anomaly index, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set. S4. Process the real-time marine monitoring data and input it into the XGBoost prediction model to obtain the real-time classification result probability, calculate the real-time weighted anomaly index, compare the stratified threshold set with the real-time weighted anomaly index, and obtain the real-time early warning stratified signal. The forced field gradient includes the wind energy gradient, the pressure gradient, and the tsunami wave ray path density gradient. The energy fluxes include wind energy flux, latent heat flux, and tsunami wave energy flux; The disaster-causing factors include a first disaster-causing factor, a second disaster-causing factor, and a third disaster-causing factor; The historical data includes wind field, pressure field, geological source information, sea surface condition, and seabed condition; the sea surface condition includes sea surface roughness, eddy current, divergence, and sea surface temperature; the seabed condition includes seabed roughness, turbulent dissipation rate, bottom boundary layer flow regime, and seabed temperature.

2. The marine disaster risk prediction method based on machine reinforcement learning according to claim 1, characterized in that, The method for constructing the disaster-causing factor includes: The wind speed gradient is used as the wind stress enhancement factor, and the wind energy flux is scaled and adjusted by a product. The wind speed gradient and pressure gradient are used together as thermodynamic adjustment factors, and the latent heat flux is scaled and adjusted by a two-factor product. The tsunami wave ray path density gradient is used as the diffusion adjustment factor, and the tsunami wave energy flux is scaled and adjusted by an exponential decay correction. The sea surface roughness and absolute value of eddy current are harmonic averaged to obtain the basic wind stress quantity. The divergence is used as a gated symbol variable, and the basic wind stress quantity is asymmetrically modulated by a sigmoid-type gated function to construct the first disaster-causing factor. The sea surface temperature and the absolute value of divergence are inversely proportionally saturated and coupled to obtain the dynamic convergence enhancement coefficient. A vorticity inverse suppression term is constructed from the vorticity square term. The dynamic convergence enhancement coefficient is then adjusted inversely to construct the second disaster factor. By saturating and mapping the bed surface roughness to the bottom boundary flow regime, a basic term for bottom boundary energy dissipation is obtained. A secondary acceleration dissipation term is constructed using the turbulent dissipation rate. The basic term for bottom boundary energy dissipation and the secondary acceleration dissipation term are multiplied to construct a triple product exponential decay kernel. The seabed temperature is used as a compensation term to adjust the logarithmic gain of the decay result and construct a third disaster factor.

3. The marine disaster risk prediction method based on machine reinforcement learning according to claim 1, characterized in that, The method for constructing the XGBoost prediction model includes: The disaster-causing factors were randomly divided into a training set and a test set in a 7:3 ratio to train and test the XGBoost prediction model; the XGBoost prediction model includes an input layer, a discriminant layer, a prediction layer and an output layer; The discriminant layer unfolds the disaster-causing factors in chronological order and extracts the energy fingerprint feature set, which is then input into the XGBoost multi-classifier to output the disaster type discrimination result. The prediction layer allocates disaster-causing factors to preset prediction sub-modules according to the disaster type discrimination results, and outputs disaster risk prediction results; the prediction sub-modules include typhoon prediction sub-module, tsunami prediction sub-module, storm surge prediction sub-module, typhoon-storm prediction sub-module and calm period prediction sub-module. The XGBoost prediction model improves prediction accuracy through mean squared error loss term and physical constraint loss term; the XGBoost prediction model uses an adaptive momentum estimation algorithm to optimize model hyperparameters; The physical constraint loss term uses the gradient wind balance equation to constrain the typhoon prediction submodule, the tidal theory equation to constrain the tsunami prediction submodule, the tidal-storm surge nonlinear equation to constrain the storm surge prediction submodule, and the wind stress curl coupled water level response equation to constrain the typhoon-storm surge prediction submodule.

4. The marine disaster risk prediction method based on machine reinforcement learning according to claim 3, characterized in that, The method for constructing the energy fingerprint feature set includes: The energy fingerprint feature set includes frequency domain fingerprint, evolution rate fingerprint, wave morphology fingerprint, and coupling correlation fingerprint; The divergence gating term is used to calculate the physical gated energy concentration ratio, construct the physical gated frequency scale, and calculate the physical gated multi-scale entropy rate. The minimum point of the physical gated multi-scale entropy rate and the physical gated energy concentration ratio together constitute the frequency domain fingerprint vector. The recursive matrix is ​​obtained by reconstructing the phase space of the time series. The physical damping coefficient is calculated by using the vortex inverse suppression term, and the physical damping evolution weighted index is constructed. The physical damping weighted determinism is calculated based on the recursive matrix. The evolution rate fingerprint vector is formed by the physical damping weighted evolution index and the physical damping weighted determinism. Hilbert transforms are performed on the time series of the first and second disaster-causing factors respectively, and the physical coupling index and non-uniform physical phase coupling strength of the disaster-causing factors are calculated simultaneously to determine the coupling correlation fingerprint vector. Homology analysis is performed on the time series to obtain the persistence of connected components. The physical decay weighted persistence energy entropy is calculated. The reciprocal of the triple product exponential decay kernel is used as the physical decay time scale. The time position corresponding to the maximum persistence is divided by the negative value of the physical decay time scale as the decay weight of the exponent. The exponent is then multiplied by the maximum persistence and the total energy of the time-domain waveform to determine the physical decay topology-energy coupling index and obtain the waveform morphology fingerprint vector.

5. The marine disaster risk prediction method based on machine reinforcement learning according to claim 1, characterized in that, The method for constructing a hierarchical threshold set includes: The range of the disaster-causing factor is used as the basic dimension of the state space. The continuous range of the disaster-causing factor is discretized into multiple level intervals and combined to construct a high-dimensional joint state space and an asymmetric cost function. Using a high-dimensional joint state space as the state space and a preset set of warning level actions as the action space, the optimal strategy is solved. The optimal strategy is back-mapped to the value space of the disaster-causing factors, the state corresponding to the optimal action is marked, the level interval boundary values ​​of each disaster-causing factor in the state are extracted, the physical quantity thresholds corresponding to each disaster-causing factor under different warning levels are determined, and a hierarchical threshold set is constructed.

6. The marine disaster risk prediction method based on machine reinforcement learning according to claim 5, characterized in that, The method for constructing the asymmetric cost function includes: Obtain the disaster discrimination probability output by the XGBoost prediction model, and calculate the weighted anomaly index based on the disaster-causing factors and the disaster discrimination probability; The disaster category in the high-dimensional joint state space is taken as the state, and the warning level in the preset warning level action set is taken as the candidate warning action to construct an asymmetric classification error cost term. The historical disaster values ​​corresponding to the state are used as the true intensity, and the predicted output values ​​corresponding to the candidate warning actions are used as the predicted intensity to construct a regression penalty term. Based on the weighted anomaly index and the preset weighted anomaly index, a weighted anomaly index penalty term is constructed; An asymmetric cost function is constructed based on the asymmetric classification error cost matrix, regression penalty term, and weighted anomaly index penalty term.

7. A marine disaster risk prediction system based on machine reinforcement learning, used to perform the method according to any one of claims 1-6, characterized in that, include: Data processing module: used to collect historical and real-time marine monitoring data, perform spatiotemporal alignment, calculate the forced field gradient, and divide the historical marine monitoring data into first data and second data according to a random ratio; Disaster Factor Module: Used to correct energy flux using forced field gradient in the first data and to construct disaster factors by coupling sea surface state and seabed state; Prediction Model Module: Used to build XGBoost prediction models, process the second data input to the XGBoost prediction model, and obtain the model output results; The hierarchical threshold set module is used to calculate the weighted anomaly index based on the classification result probability obtained from the model output, perform Markov decision to obtain the optimal early warning strategy, and construct a hierarchical threshold set. Early warning output module: It is used to process real-time marine monitoring data, input it into the XGBoost prediction model, obtain the probability of real-time classification results, calculate the real-time weighted anomaly index, compare the stratified threshold set with the real-time weighted anomaly index, and obtain real-time early warning stratified signals.

Images