A physical perception rare sample enhancement and regional extreme value post-calibration method for extreme wave height prediction
Patent Information
- Application Number
- CN202611044587.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-10-13
AI Technical Summary
[0011]本发明的目的在于提供一种面向极端波高预测的物理感知稀有样本增强与区域极值后校准方法，以解决现有深度学习有效波高预测方法在极端波高样本不足、极端波高低估、风浪物理关系不一致以及不同海区尾部分布差异刻画不足等方面的问题
[0057]1.提高极端有效波高样本学习能力。本发明通过物理感知SMOGN方法增加极端波高稀有样本数量，缓解极端样本不足和样本分布不平衡问题。
Smart Images
Figure CN122903893A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent marine environmental forecasting, post-processing of numerical wave forecasts, deep learning spatiotemporal prediction, and early warning of extreme sea state risks, and particularly to a method for enhancing rare physical perception samples and post-calibrating regional extreme value theory for extreme wave height prediction. Background Technology
[0002] Significant wave height is a crucial parameter for describing sea state strength and marine disaster risk, with significant applications in ship navigation safety, offshore wind power operation and maintenance, marine engineering design, maritime search and rescue, fisheries production, disaster prevention and mitigation, and port scheduling. Ocean wave fields exhibit significant nonlinearity, non-stationarity, and regional variability. Particularly during typhoons, cold waves, strong winds, and severe convective weather, extreme significant wave heights can develop rapidly, posing a significant threat to coastal safety and offshore operations.
[0003] Traditional wave forecasting primarily relies on physical numerical models. While these models offer good physical interpretability, they suffer from high computational costs, error propagation, and local system biases in high-resolution regional forecasting, rapid updates, and complex nearshore conditions. In recent years, deep learning methods have been widely applied to significant wave height prediction, including recurrent neural networks, long short-term memory networks, convolutional long short-term memory networks, convolutional neural networks, graph neural networks, Transformers, and fully convolutional spatiotemporal prediction networks. These methods can automatically learn spatiotemporal evolution patterns from historical wave fields and meteorological fields, exhibiting high computational efficiency and strong data fitting capabilities.
[0004] However, existing deep learning-based effective wave height prediction methods still have the following shortcomings in extreme wave height prediction:
[0005] First, the training sample distribution is severely unbalanced. The number of samples with normal sea states far exceeds the number of samples with extreme wave heights. During model training, the normal sea states tend to dominate, resulting in a lower overall mean square error but a lower prediction of extreme wave heights.
[0006] Second, conventional data augmentation methods primarily generate samples from the perspective of statistical distance or Euclidean distance, without fully considering the physical consistency between wind speed, wind direction, and significant wave height. For wave prediction tasks, selecting nearest neighbors solely based on the numerical similarity of input features may generate synthetic samples with inconsistent wind-wave relationships, thereby weakening the model's ability to learn from extreme sea conditions.
[0007] Third, commonly used loss functions such as mean squared error or mean absolute error do not pay enough attention to the tail samples of extreme wave heights, especially lacking targeted penalties when extreme wave heights are underestimated. This can lead to systematic underestimation of wave heights by the model under high-risk conditions such as typhoon waves and strong winds.
[0008] Fourth, the statistical characteristics of extreme wave heights differ across different sea areas, exhibiting significant variations in water depth, wind field, frequency of typhoon impact, open sea area extent, and climatological wave height distribution. Using a uniform post-processing model for the entire sea area would make it difficult to simultaneously characterize the tail distribution features of different sea areas.
[0009] Fifth, existing deep learning models typically output deterministic predictions directly, lacking post-calibration mechanisms for extreme tail distributions. For risk warning, tail probability, tail amplitude, and prediction stability near extreme thresholds are particularly important; relying solely on the raw output of the neural network is insufficient to obtain adequately calibrated extreme wave height predictions.
[0010] To address the aforementioned issues, this invention proposes an integrated technical approach: augmentation of rare samples using physical perception, training with tail-perception loss, and post-calibration using regional extreme value theory. This method improves rare sample augmentation through wind and wave physical consistency constraints during the training sample construction phase; strengthens extreme wave height learning through tail-perception loss during model training; and performs tail calibration of extreme prediction results using regional extreme value theory models for different sea areas during the prediction post-processing phase. This improves the prediction accuracy, recall, and physical plausibility of extreme effective wave heights in complex sea areas. Summary of the Invention
[0011] The purpose of this invention is to provide a method for enhancing rare physical perception samples and post-calibrating regional extrema for extreme wave height prediction, addressing the problems of insufficient extreme wave height samples, underestimation of extreme wave height, inconsistent wind-wave physical relationships, and inadequate characterization of tail distribution differences in different sea areas in existing deep learning effective wave height prediction methods. This invention can improve the prediction accuracy of extreme effective wave height samples, reduce the underestimation bias of extreme wave height, increase the recall rate of extreme wave height identification, and enhance the spatial adaptability and physical rationality of prediction results in different sea areas while maintaining the overall prediction accuracy of conventional sea states.
[0012] To achieve the above objectives, the present invention adopts the following technical solution:
[0013] A method for enhancing rare physical-sensory samples and post-calibrating regional extrema for extreme wave height prediction includes the following steps:
[0014] S1. Obtain historical wave field and meteorological field data of the study sea area, and construct multivariate spatiotemporal sequence samples. The multivariate includes at least significant wave height, 10 m meridional wind speed, and 10 m zonal wind speed, and may further include one or more of the following: mean wave period, sine and cosine components of mean wave direction, land-sea mask, latitude and longitude, and seabed topography.
[0015] S2. Construct a sample-level rarity score based on the target effective wave height field, and filter extreme wave height rare samples according to the preset extreme threshold or quantile threshold.
[0016] S3. Perform Synthetic Minority Over-sampling (SMOGN) on the rare extreme wave height samples. When selecting the nearest neighbors of the rare samples, both input feature similarity and wind / wave physical consistency are considered to construct a comprehensive distance.
[0017] D = α·D feature + β·D physics ,
[0018] Among them, D feature D represents the normalized distance between sample input features or their latent space representations. physics This represents the physical distance constructed based on the relationship between wind speed and significant wave height, where α and β are weighting coefficients.
[0019] S4. Train the effective wave height spatiotemporal prediction model based on multivariate spatiotemporal sequence samples and enhanced samples generated by physical sensing SMOGN, and use the tail sensing loss function during the training process to give the extreme wave height region and the extreme wave height underestimation error a higher penalty weight than the conventional wave height region.
[0020] S5. Using the effective wave height prediction value and corresponding observation value or reanalysis true value output by the spatiotemporal prediction model on the validation set, establish a regional extreme value theory calibration model according to different sea areas or different spatial clustering regions of the study sea area. Among them, the generalized Pareto distribution of the prediction tail and the true tail are fitted to each region respectively.
[0021] S6. In the prediction stage, the spatiotemporal prediction model first outputs the basic predicted significant wave height field, and then the model is calibrated by calling the corresponding regional extreme value theory according to the region to which each ocean grid belongs. Probability mapping and smoothing fusion are performed on the predicted values that exceed the regional tail threshold to obtain the final predicted field after extreme significant wave height calibration.
[0022] S7, Output the final effective wave height prediction field Y after calibration final Furthermore, it can provide an evaluation of the model's prediction results.
[0023] The sample-level rarity score in step S2 is determined by the maximum value of the target effective wave height field, the average value of spatial grid points exceeding a preset proportion, the average value of spatial grid points exceeding a preset threshold, or a combination thereof.
[0024] The physical consistency of wind and waves mentioned in step S3 is determined by the dimensionless or quasi-dimensional relationship between the significant wave height and the square of the wind speed. The expression for this ratio includes:
[0025] R = H s / (U10² + ε), or R = g·H s / (U10² + ε),
[0026] Where H s The significant wave height is represented by U10² = u10² + v10², where u10 and v10 represent the meridional and zonal wind speeds at 10 m, respectively. g represents the gravitational acceleration, and ε is a stabilizing term to prevent the denominator from being too small.
[0027] In step S3, the comprehensive distance D feature and D physics After normalization, the results are weighted and summed, with β = 1 - α, where α ranges from 0.5 to 0.9.
[0028] In step S3, when generating enhanced samples, for the current rare extreme wave height sample i, the k physically sensed nearest neighbors with the smallest comprehensive distance are selected from the set of rare extreme wave height samples. Based on the comprehensive distance D(i, j) between the current rare sample i and the selected nearest neighbor sample j and the preset safe distance threshold safe_thr, the synthesis method is adaptively selected.
[0029] When D(i, j) < safe_thr, nearest neighbor interpolation is used to generate the synthesis input and synthesis target:
[0030] X new = X i + λ·(X j - X i ),
[0031] Y new = Y i + λ·(Y j - Y i ),
[0032] When D(i, j) ≥ safe_thr, a Gaussian perturbation method is used to generate synthetic input and synthetic target near the current rare samples:
[0033] X new = X i + η X ,
[0034] Y new = Y i + η Y ,
[0035] Among them, X i Y iThese represent the current rare sample and its target effective wave height field, X. j Y j These represent the effective wave height fields of its physical sensing nearest neighbor samples and its target, respectively, where λ is a random interpolation coefficient between 0 and 1, and η is the target wave height field. X and η Y The Gaussian perturbation term is subject to constraints such as land-sea masking, non-negativity of effective wave height, and physical consistency of wind and waves, ensuring that the generated samples maintain spatial continuity and reasonable wind-wave relationships.
[0036] The safe distance threshold safe_thr is determined by the statistical distribution of the comprehensive distance between samples in the rare sample set of extreme wave heights, or by one or more of the preset ratio parameter and the comprehensive distance mean, comprehensive distance median, and comprehensive distance quantile; the comprehensive distance D(i, j) is the weighted sum of the input feature distance and the wind and wave physical consistency distance after normalization.
[0037] The tail-perceiving loss function described in step S4 includes a basic error term and a tail-weighted term, expressed as:
[0038] L = Σ M·w(Y)·ρ(Y hat -Y) + λ u Σ M·I(Y≥u)·max(YY hat ,0)^q,
[0039] Where M is the ocean mask, Y is the true significant wave height, and Ys is the true significant wave height. hat To predict the significant wave height, ρ represents the squared error, absolute error, or Huber error, w(Y) is the weighting function that increases with Y, u is the tail threshold, I is the indicator function, and λ... u To underestimate the penalty weight, q is an exponent greater than or equal to 1.
[0040] The weighting function w(Y) can be a step function, a power function, or a sigmoid smoothing function, where the sigmoid smoothing function is expressed as:
[0041] w(Y) = 1 + λ t ·sigmoid((Yu) / s),
[0042] Where, λ t s is the tail enhancement weight, and s is the smoothing scale.
[0043] The post-calibration model based on the regional extremum theory described in step S5 includes the following process:
[0044] For each region r, extract the threshold u from the true significant wave height of the validation set and the model-predicted significant wave height, respectively. r The tail samples are used to obtain the true transcendence Z. obs= Y - u r And the predicted excess Z pred =Y hat -u r ;
[0045] For Z respectively obs and Z pred By fitting a generalized Pareto distribution, the true tail distribution parameters and the predicted tail distribution parameters are obtained.
[0046] When the number of tail samples in a certain region is less than the preset minimum number of tail samples, adjacent regions are merged, parameters for the entire sea area are rolled back, or empirical quantile mapping is rolled back.
[0047] The probability mapping and smoothing fusion described in step S6 includes:
[0048] When the basic predicted value Y hat Greater than the regional prediction threshold u pred, r When the prediction tail is distributed, calculate the cumulative probability p.
[0049] Substituting the cumulative probability p into the inverse function of the true tail distribution, we obtain the regional extreme value theoretical calibration value Y. evt ;
[0050] Using a gating function to apply Y hat and Y evt To merge:
[0051] Y final = Y hat + a·G(Y hat -u pred,r )·(Y evt -Y hat ),
[0052] Where a is the fusion coefficient and G is the Sigmoid gate function.
[0053] The spatiotemporal prediction model is a convolutional neural network, ConvLSTM, U-Net, SimVP, SimVPv2, Transformer, MLP-Mixer, fully convolutional prediction network or a combination thereof, used to predict the effective wave height field for one or more forecast lead times in the future based on multivariate marine meteorological fields from several past time periods.
[0054] During training, validation, and prediction, land areas are shielded using a land-sea mask, so that the loss function, sample sparseness calculation, extreme value distribution fitting, and prediction result evaluation are all performed only on the ocean grid.
[0055] The evaluation metrics for the model prediction results in step S7 mainly include: root mean square error (RMSE) of the whole sample; mean absolute error (MAE) of the whole sample; bias of the whole sample; RMSE, MAE, and bias of extreme samples; precision of extreme wave height recognition; recall of extreme wave height recognition; and F1 score of extreme wave height recognition.
[0056] Compared with existing technologies, the beneficial effects achieved by the technical solution of this invention are:
[0057] 1. Improve the learning ability of extreme effective wave height samples. This invention increases the number of rare extreme wave height samples through the physical sensing SMOGN method, alleviating the problems of insufficient extreme samples and imbalanced sample distribution.
[0058] 2. Enhance the physical plausibility of the enhanced samples. This invention introduces wind and wave physical consistency distance into the nearest neighbor screening process, avoiding the generation of samples with inconsistent wind and wave relationships based solely on statistical distance.
[0059] 3. Reduce the underestimation bias of extreme wave height. This invention enhances the weight of grid points for extreme wave height through a tail-perception loss function and applies additional penalties for underestimating extreme wave height, making the model more inclined to learn tail features under high-risk sea conditions.
[0060] 4. Improve adaptability to different sea areas. This invention employs a regional extreme value theory post-calibration mechanism to characterize the tail distribution differences in different spatial regions, thereby enhancing the model's generalization ability in complex sea areas.
[0061] 5. Improved extreme prediction calibration capability. This invention enables statistical calibration of the original neural network output within the extreme tail range by using a probability mapping from the predicted tail distribution to the true tail distribution.
[0062] 6. Balancing overall accuracy and extreme performance. This invention introduces a Sigmoid-gated smoothing fusion mechanism, initiating extreme value calibration correction only for the extreme tail prediction region, without interfering with conventional wave height forecast results. While significantly improving the accuracy of extreme sea state forecasts, it fully preserves the model's forecasting advantages for conventional sea states, achieving optimal balance between global accuracy and extreme performance. Attached Figure Description
[0063] Figure 1 This is a flowchart illustrating the overall process of the method for predicting extreme wave heights according to the present invention.
[0064] Figure 2 This is a flowchart of the rare sample enhancement process for Physical Sensing SMOGN in this invention.
[0065] Figure 3 This is a schematic diagram of the tail-sensing loss function in this invention.
[0066] Figure 4This is a flowchart of the post-calibration process based on the regional extremum theory in this invention. Detailed Implementation
[0067] To make the technical problems, technical solutions, and beneficial effects of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, this invention is not limited thereto. Equivalent modifications and adaptations made by those skilled in the art based on the core concept of this invention are all within the scope of protection of this invention. Details not described in detail can be implemented using conventional techniques in the art.
[0068] Figure 1 The following is a flowchart of the method for extreme wave height prediction of this invention. The overall process is as follows: input historical ocean wave / meteorological data, perform S1 data preprocessing and sample construction, S2 rare sample identification, S3 physical sensing SMOGN enhancement, S4 tail sensing loss training model, S5 validation set prediction and regional extreme value theory parameter fitting, S6 prediction stage basic prediction and regional extreme value theory post-calibration, and finally output the calibrated effective wave height prediction result.
[0069] Example: Prediction of 48-hour significant wave height over the Chinese coast based on ERA5 data
[0070] 1. Data Preparation
[0071] ERA5 wave and meteorological reanalysis data from the China Sea region were selected, with a temporal resolution of 1 hour and a spatial resolution of 0.5°×0.5°. The study area ranges from 104.5°E to 134°E and from 2.5°N to 42°N, with a spatial grid of 80×60.
[0072] The input variables include the following historical 48-hour data: significant wave height Hs; 10 m meridional wind speed u10; 10 m zonal wind speed v10; mean wave period Tm; mean wave direction sine component Ds; and mean wave direction cosine component Dc.
[0073] The output variable is the effective wave height field for the next 48 hours.
[0074] Ocean and land grid points are identified using a land-sea mask M. During training and evaluation, loss and metrics are calculated only for ocean grid points.
[0075] 2. Sample Construction
[0076] For each reporting time t, the multivariate spatial field of the past 48 hours is taken as input:
[0077] X t = [F{t-47}, F{t-46}, ... , F{t}].
[0078] Take the effective wave height field of the next 48 hours as the output:
[0079] Y t = [H s {t+1}, H s {t+2}, ... , Hs{t+48}].
[0080] The training set, validation set, and test set are divided into three parts according to time sequence to ensure that the validation set and test set do not participate in parameter updates during model training.
[0081] 3. Rare Sample Identification
[0082] For each training sample, calculate the maximum value in the target's effective wave height field:
[0083] score t = max(Y t ·M).
[0084] In this embodiment, the extreme wave height threshold is set to 4.0m, i.e., score t If the value is ≥4.0 m, the training sample is marked as a rare sample with extreme wave height. Alternatively, the 95th percentile of the distribution of the maximum effective wave height of the target in the training set can be used as an adaptive rarity threshold.
[0085] 4. Enhanced Physics-Aware SMOGN (refer to...) Figure 2 )
[0086] For each rare sample i, calculate the input feature distance D between it and other rare samples j. feature In this embodiment, the distance is calculated on the standardized input tensor, or it can be calculated after extracting low-dimensional features through an encoder.
[0087] Then calculate the wind-wave physical consistency distance between sample i and sample j. First, calculate the square of the wind speed:
[0088] U10² = u10² + v10².
[0089] Next, calculate the wind-wave ratio:
[0090] R = g·H s / (U10² + ε).
[0091] The difference in R is calculated at the ocean grid points at the target extreme wave height region or at the end of the input sequence to obtain the wind and wave physical distance D. physics .
[0092] D feature and D physics Normalize each value separately, and calculate the composite distance by setting α=0.7 and β=0.3:
[0093] D = 0.7·D feature_norm + 0.3·D physics_norm .
[0094] Among them, D feature D represents the normalized distance between sample input features or their latent space representations. physics This represents the physical distance constructed based on the relationship between wind speed and significant wave height.
[0095] For the current rare sample i with extreme wave height, select the 5 physically sensed nearest neighbors with the smallest comprehensive distance from the set of rare samples with extreme wave height. Based on the comprehensive distance D(i, j) between the current rare sample i and the selected nearest neighbor sample j and the preset safety distance threshold of 0.5, adaptively select the synthesis method:
[0096] When D(i, j) < 0.5, nearest neighbor interpolation is used to generate the synthesis input and synthesis target:
[0097] X new = X i + λ·(X j - X i ),
[0098] Y new = Y i + λ·(Y j - Y i ),
[0099] When D(i, j) ≥ 0.5, a Gaussian perturbation method is used to generate synthetic input and synthetic target near the current rare sample:
[0100] X new = X i + η X ,
[0101] Y new = Y i + η Y ,
[0102] Non-negative constraints are applied to the synthesized effective wave height, and land grid points are kept invalid using a land-sea mask.
[0103] Figure 2 This is a flowchart of the rare sample enhancement process for Physical Sensing SMOGN in this invention, using a high-frequency, high-rarity sample set {(X i ,Y i Based on )}, calculate the input feature distance D respectively. feature (i,j) and the physical distance D from the wind and waves physics(i,j) are normalized, and the comprehensive distance D = α·D is constructed. feature + β·D physics Select k physical sense neighbors, and generate synthetic samples by means of nearest neighbor interpolation or Gaussian perturbation based on the relationship between the comprehensive distance D(i, j) between the current rare sample i and the selected nearest neighbor sample j and the safety threshold safe_thr. Finally, output the enhanced training samples.
[0104] 5. Model Training (refer to...) Figure 3 )
[0105] The original training samples and augmented samples are mixed and then input into the spatiotemporal prediction model. This embodiment uses a fully convolutional spatiotemporal prediction network, but SimVP, ConvLSTM, U-Net, or Transformer can also be used.
[0106] The loss function uses tail-aware MSE:
[0107] L = ΣM·w(Y)·(Y hat -Y)² + λ u ΣM·I(Y≥4.0)·max(YY hat ,0)².
[0108] The weighting function is:
[0109] w(Y) = 1 + λ t ·sigmoid((Y-4.0) / s).
[0110] The loss function increases the training weights for extreme wave height grid points near and above the extreme wave height threshold of 4.0 m, and imposes additional penalties for cases where the true value is higher than 4.0 m but the predicted value is lower.
[0111] Figure 3 This is a schematic diagram of the tail-sensing loss function structure in this invention, using the true effective wave height Y and the model-predicted effective wave height Yi as examples. hat With the land-sea mask M as input, the basic error term ρ(Y) is first calculated. hat -Y) and introduce a tail weight w(Y). For samples that exceed the tail threshold u and are undervalued, an additional underestimation penalty term max(YY) is added. hat ,0)^q and the corresponding underestimated penalty weight λ u The weighted summation yields the overall loss function L, which is then used to update the model parameters via backpropagation.
[0112] 6. Fitting the theory of regional extrema
[0113] After training, run the model on the validation set to obtain the validation set prediction value Y. hat_valand the true value Y of the validation set val .
[0114] In this embodiment, it can be further subdivided into more regions based on latitude, longitude, and the distribution of extreme wave heights in climatic conditions.
[0115] For each region r, samples with actual and predicted significant wave heights exceeding 4.0 m are extracted. A generalized Pareto distribution (GPD) is fitted to both the actual and predicted exceedance values to obtain the actual and predicted tail distribution parameters for each region.
[0116] When the number of samples exceeding 4.0 m in a certain area is less than 200, a strategy of merging adjacent sea areas or parameter rollback for the entire sea area is adopted.
[0117] 7. Post-prediction calibration (refer to...) Figure 4 )
[0118] During the testing or business forecasting phase, the model first outputs the basic predicted effective wave height field Y. hat For each ocean grid point, its region r is determined based on its latitude and longitude.
[0119] If Y hat If the value is ≤4.0m, it is considered a normal sea state and no extreme value calibration is required. In this case, the predicted value remains unchanged or a weak calibration is performed.
[0120] If Y hat If the depth is greater than 4.0m, then calculate:
[0121] p = G pred,r (Y hat - 4.0).
[0122] Then mapped to the actual tail distribution:
[0123] Y evt = 4.0 + G obs,r ^{-1}(p).
[0124] Finally, perform a smooth blending process:
[0125] Y final = Y hat + a·sigmoid(k(Y hat -4.0))·(Y evt -Y hat ).
[0126] In this embodiment, a is 0.7 and k is 5.0.
[0127] Figure 4This is a flowchart of the post-calibration process for the regional extremum theory in this invention, divided into two parts: a verification stage and a prediction stage. Verification stage: using the true value Y of the verification set... val Compared with the predicted value Y hat_val The regions are divided according to sea area / latitude and longitude / cluster, and the true tail excess Z of each region is extracted. obs =Y−u obs,r And the predicted tail excess Z predr =Y hat -u pred,r ; respectively fit the true tail GPD and the predicted tail GPD distribution parameters, and establish and save the regional extreme value theoretical parameter library {θ}. obs,r ,θ pred,r ,u r Prediction phase: Determine the region r for each grid point and determine the base predicted value Y. hat Has the region tail calibration trigger threshold been exceeded? pred,r If the value exceeds the limit, the predicted tail cumulative probability p is calculated and mapped to the true tail distribution to obtain the calibration value Y. evt Then, the final predicted field Y is obtained by smoothing and fusing the data through a Sigmoid gate function. final For areas with normal sea states, the baseline prediction results are directly retained; if the sea state is not exceeded, the baseline prediction or weak calibration is maintained. Here, Y represents the true significant wave height, and u... obs,r The extreme value fitting threshold for the real sample in the r-th sub-region of the sea area is denoted by u, where r is the sea area sub-index. pred,r The tail calibration trigger threshold for the predicted sample of the r-th sub-region of the sea area; θ obs,r For true tail GPD, θ pred,r To predict tail GPD.
[0128] 8. Result Output
[0129] Output the hourly effective wave height prediction field for the next 48 hours and provide an evaluation of the model prediction results, as shown in Table 1.
[0130] Table 1 Evaluation of 48h Prediction Results
[0131]
[0132] As shown in Table 1, the method of this invention significantly improves the forecasting performance of extreme wave heights while maintaining good overall forecast accuracy. The overall error level of the entire sample is low, the bias of extreme samples is significantly reduced, the recall rate of extreme wave heights reaches 0.7688, and the F1 score reaches 0.7805. This effectively alleviates the problem of systematic underestimation of extreme wave heights, greatly enhances the ability to identify extreme sea states, and demonstrates good forecasting performance and regional adaptability for extreme wave heights. This invention improves the quality of extreme samples by adding wind and wave physical distances to construct an improved SMOGN with a comprehensive distance, specifically corrects the underestimation bias by addressing tail perception loss, and calibrates regional extreme values to adapt to the statistical characteristics of different sea areas. These three aspects work synergistically to improve extreme forecasting performance. This invention is applicable to marine extreme sea state risk forecasting scenarios such as marine disaster early warning, ship navigation safety assurance, offshore wind power operation and maintenance, marine engineering design, maritime search and rescue, port scheduling, and disaster prevention and mitigation in fisheries production.
[0133] The above embodiments are merely preferred embodiments of the present invention, and the parameter settings are all exemplary and not intended to limit the overall technical solution of the present invention. All equivalent modifications, adaptations, and variant implementations made based on the core method of the present invention fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for enhancing rare physical-sensory samples and post-calibrating regional extrema for extreme wave height prediction, characterized in that, Includes the following steps: S1. Obtain historical wave field and meteorological field data of the study sea area, and construct multivariate spatiotemporal sequence samples. The multivariate includes at least significant wave height, 10 m meridional wind speed, and 10 m zonal wind speed, and may further include one or more of the following: mean wave period, sine and cosine components of mean wave direction, land-sea mask, latitude and longitude, and seabed topography. S2. Construct a sample-level rarity score based on the target effective wave height field, and filter extreme wave height rare samples according to the preset extreme threshold or quantile threshold. S3. Perform Physically Perceived Minority Oversampling (SMOGN) on the rare extreme wave height samples. When selecting the nearest neighbors of the rare samples, both input feature similarity and wind / wave physical consistency are considered to construct a composite distance. D = α·D feature + β·D physics , Among them, D feature D represents the normalized distance between sample input features or their latent space representations. physics This represents the physical distance constructed based on the relationship between wind speed and significant wave height, where α and β are weighting coefficients; S4. Train the effective wave height spatiotemporal prediction model based on multivariate spatiotemporal sequence samples and enhanced samples generated by physical sensing SMOGN, and use the tail sensing loss function during the training process to give the extreme wave height region and the extreme wave height underestimation error a higher penalty weight than the conventional wave height region. S5. Using the effective wave height prediction value and corresponding observation value or reanalysis true value output by the spatiotemporal prediction model on the validation set, establish a regional extreme value theory calibration model according to different sea areas or different spatial clustering regions of the study sea area. Among them, the generalized Pareto distribution of the prediction tail and the true tail are fitted to each region respectively. S6. In the prediction stage, the spatiotemporal prediction model first outputs the basic predicted significant wave height field, and then the model is calibrated by calling the corresponding regional extreme value theory according to the region to which each ocean grid belongs. Probability mapping and smoothing fusion are performed on the predicted values that exceed the regional tail threshold to obtain the final predicted field after extreme significant wave height calibration. S7, Output the final effective wave height prediction field Y after calibration final Furthermore, it can provide an evaluation of the model's prediction results.
2. The method according to claim 1, characterized in that, The sample-level rarity score in step S2 is determined by the maximum value of the target effective wave height field, the average value of spatial grid points exceeding a preset proportion, the average value of spatial grid points exceeding a preset threshold, or a combination thereof.
3. The method according to claim 1, characterized in that, The physical consistency of wind and waves mentioned in step S3 is determined by the dimensionless or quasi-dimensional relationship between the significant wave height and the square of the wind speed. The expression for this ratio includes: R = H s / (U10² + ε),or R = g·H s / (U10² + ε), Where H s The significant wave height is represented by U10² = u10² + v10², where u10 and v10 represent the meridional and zonal wind speeds at 10 m, respectively. g represents the gravitational acceleration, and ε is a stabilizing term to prevent the denominator from being too small.
4. The method according to claim 1, characterized in that, In step S3, the comprehensive distance D feature and D physics After normalization, the results are weighted and summed, with β = 1 - α, where α ranges from 0.5 to 0.
9.
5. The method according to claim 1, characterized in that, In step S3, when generating enhanced samples, for the current rare extreme wave height sample i, the k physical sense nearest neighbors with the smallest comprehensive distance are selected from the set of rare extreme wave height samples. Based on the comprehensive distance D(i, j) between the current rare sample i and the selected nearest neighbor sample j and the preset safe distance threshold safe_thr, the synthesis method is adaptively selected. When D(i, j) < safe_thr, nearest neighbor interpolation is used to generate the synthesis input and synthesis target: X new = X i + λ·(X j - X i ), AND new = And i + λ·(Y j - AND i ), When D(i, j) ≥ safe_thr, a Gaussian perturbation method is used to generate synthetic input and synthetic target near the current rare samples: X new = X i + the X , AND new = And i + η Y , Among them, X i Y i These represent the current rare sample and its target effective wave height field, X. j Y j These represent the effective wave height fields of its physical sensing nearest neighbor samples and its target, respectively, where λ is a random interpolation coefficient between 0 and 1, and η is the target wave height field. X and η Y The Gaussian perturbation term is subject to constraints imposed by land-sea masking, non-negativity of effective wave height, and physical consistency of wind and waves, ensuring that the generated samples maintain spatial continuity and a reasonable relationship between wind and waves. The safe distance threshold safe_thr is determined by the statistical distribution of the comprehensive distance between samples in the rare sample set of extreme wave heights, or by one or more of the preset ratio parameter and the comprehensive distance mean, comprehensive distance median, and comprehensive distance quantile; the comprehensive distance D(i, j) is the weighted sum of the input feature distance and the wind and wave physical consistency distance after normalization.
6. The method according to claim 1, characterized in that, The tail-perceiving loss function described in step S4 includes a basic error term and a tail-weighted term, expressed as: L = Σ M·w(Y)·ρ(Y hat -Y) + λ u Σ M·I(Y≥u)·max(YY hat ,0)^q Where M is the ocean mask, Y is the true significant wave height, and Ys is the true significant wave height. hat To predict the significant wave height, ρ is the squared error, absolute error, or Huber error, w(Y) is the weighting function that increases with the true significant wave height Y, u is the tail threshold, I is the indicator function, and λ u To underestimate the penalty weight, q is the exponent of the underestimated error term, and q≥1; The weighting function w(Y) is a step function, a power function, or a sigmoid smoothing function, where the sigmoid smoothing function is expressed as: w(Y) = 1 + λ t ·sigmoid((Y-u) / s) Where, λ t s is the tail enhancement weight, and s is the smoothing scale.
7. The method according to claim 1, characterized in that, The post-calibration model based on the regional extremum theory described in step S5 includes the following process: For each region r, extract the threshold u from the true significant wave height of the validation set and the model-predicted significant wave height, respectively. r The tail samples are used to obtain the true transcendence Z. obs = Y - u r And the predicted excess Z pred =Y hat -u r ; respectively for the true transcendent quantity Z obs And the predicted excess Z pred By fitting a generalized Pareto distribution, the true tail distribution parameters and the predicted tail distribution parameters are obtained. When the number of tail samples in a certain region is less than the preset minimum number of tail samples, adjacent regions are merged, parameters are backed up across the entire sea area, or empirical quantile mapping is used for backing up.
8. The method according to claim 1, characterized in that, The probability mapping and smoothing fusion described in step S6 includes: When the basic predicted value Y hat Greater than the regional prediction threshold u pred, r When the prediction tail is distributed, calculate the cumulative probability p. Substituting the cumulative probability p into the inverse function of the true tail distribution, we obtain the regional extreme value theoretical calibration value Y. evt ; Using a gating function to apply Y hat and Y evt To merge: AND final = And hat + a·G(Y hat -or pred,r )·(AND evt -AND hat ), Where a is the fusion coefficient and G is the Sigmoid gate function; The spatiotemporal prediction model is a convolutional neural network, ConvLSTM, U-Net, SimVP, SimVPv2, Transformer, MLP-Mixer, fully convolutional prediction network or a combination thereof, used to predict the effective wave height field for one or more forecast lead times in the future based on multivariate marine meteorological fields from several past time periods.
9. The method according to claim 1, characterized in that, During training, validation, and prediction, land areas are shielded using a land-sea mask, so that the loss function, sample sparseness calculation, extreme value distribution fitting, and prediction result evaluation are all performed only on the ocean grid.
10. The method according to claim 1, characterized in that... The evaluation metrics for the model prediction results in step S7 include: root mean square error (RMSE) of the whole sample; mean absolute error (MAE) of the whole sample; bias of the whole sample; RMSE, MAE, and bias of extreme samples; precision of extreme wave height recognition; recall of extreme wave height recognition; and F1 score of extreme wave height recognition.