Physical-enhanced xgboost-based real-time prediction method for storm surge in lianyungang coastal area

Abstract

The application discloses a Lianyungang coastal storm surge instant prediction method based on physical enhanced XGBoost, including obtaining measured tidal level data, astronomical tidal level data and typhoon path intensity data of a Lianyungang tide observation station, obtaining a storm surge target variable through data preprocessing of the measured tidal level data, the astronomical tidal level data and the typhoon path intensity data, constructing input features and training a machine learning regression model to output a prediction result. The application constructs a 41-dimensional multi-source physical enhanced feature system, fuses site surge memory and typhoon physical forcing, improves prediction accuracy and physical interpretability; adopts an event grouping cross-validation strategy to avoid sample information leakage from the root, and to strengthen the model generalization ability and stability under independent typhoon events; based on the extreme gradient boosting model, combined with hyperparameter optimization and regular constraints, the prediction efficiency and performance are considered; and the application is developed for the Lianyungang sea area, and can provide reliable support for regional storm surge early warning and disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to the fields of marine disaster monitoring and forecasting, storm surge prediction and machine learning-based water level forecasting, specifically a real-time storm surge prediction method for the Lianyungang coastal area based on Physical Enhancement XGBoost. Background Technology

[0002] Storm surge, an abnormal rise and fall of nearshore water levels caused by severe weather systems such as tropical cyclones and extratropical cyclones, is a major marine disaster threatening the safety of my country's coastal areas. Lianyungang City in Jiangsu Province is located in Haizhou Bay, on the western edge of the South Yellow Sea. This bay is a semi-enclosed arc-shaped bay, connected to the open waters of the South Yellow Sea on its outer side and adjacent to a low-lying coastal plain and broad tidal flats on its inner side. This topographical pattern of land-sea connection, open bay entrance, and low-lying nearshore area makes nearshore water levels highly sensitive to wind, pressure, and tidal changes. Existing research shows that when a typhoon makes landfall directly on the Jiangsu coast or its path aligns with the coastline topography, significant storm surges are likely to occur along the Jiangsu coast, with extreme storm surges reaching approximately 2.5 meters in the Haizhou Bay area. Against the backdrop of continuous sea-level rise and the superposition of astronomical high tides, the water level risk facing the Lianyungang coast is further exacerbated. Therefore, conducting high-precision real-time storm surge forecasts is of significant practical importance for regional disaster prevention and mitigation.

[0003] Current storm surge prediction technologies mainly fall into two categories: numerical model forecasting and data-driven forecasting. Numerical models simulate water level evolution by solving ocean hydrodynamic control equations, revealing clear physical mechanisms. However, they consume significant computational resources and have long operating cycles, making them unsuitable for high-frequency, real-time forecasting. Data-driven methods, with machine learning at their core, have become an important technological direction for short-term storm surge prediction due to their fast computation speed and strong nonlinear fitting capabilities.

[0004] Existing machine learning-based storm surge prediction techniques include models such as persistence, linear regression, support vector regression, random forest, and extreme gradient boosting, but they still have significant shortcomings in practical applications.

[0005] First, the physical mechanisms of the input features are not sufficiently integrated. Most existing models only use historical flooding sequences and basic typhoon parameters as inputs, without systematically integrating physical information such as the short-term evolution trend of typhoon physical forcing and the directional coupling relationship between wind pressure forcing and the relative typhoon position. The feature dimensions are single and the physical constraints are weak, resulting in insufficient prediction accuracy for extreme flooding processes and poor interpretability.

[0006] Secondly, the model validation strategy has inherent flaws. Existing solutions generally use random sampling to divide the training and test sets, which can easily lead to adjacent samples from the same typhoon process being distributed in both the training and test sets simultaneously, causing information leakage. This results in the model evaluation results being biased towards interpolation performance, failing to truly reflect the model's generalization ability on independent typhoon events, and making it difficult to guarantee prediction stability in practical applications.

[0007] Furthermore, existing general prediction models have not been adapted and optimized for the regional hydrodynamic characteristics of Haizhou Bay in Lianyungang, making it difficult to accurately depict the storm surge response patterns of this semi-enclosed bay, thus limiting their regional applicability. Summary of the Invention

[0008] The purpose of this invention is to provide a real-time storm surge prediction method for the Lianyungang coastal area based on physically enhanced XGBoost. It constructs a 41-dimensional multi-source physical enhancement feature system, integrating site storm surge memory and typhoon physical forcing to improve prediction accuracy and physical interpretability. An event-grouped cross-validation strategy is employed to fundamentally avoid sample information leakage, enhancing the model's generalization ability and stability under independent typhoon events. Based on an extreme gradient boosting model, it combines hyperparameter optimization and regularization constraints to balance prediction efficiency and performance. Customized for the Lianyungang sea area, it can provide reliable support for regional storm surge early warning and disaster prevention and mitigation, thus addressing the problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a real-time storm surge prediction method based on Physically Enhanced XGBoost for the Lianyungang coast, comprising acquiring measured tide level data, astronomical tide level data, and typhoon path and intensity data from the Lianyungang tide gauge station; obtaining storm surge target variables through data preprocessing of the measured tide level data, astronomical tide level data, and typhoon path and intensity data; constructing input features and training a machine learning regression model to output prediction results; specifically including the following steps:

[0010] S1. Data preprocessing and target variable construction: Unify the measured tide level data and the astronomical tide level data to the same elevation datum, calculate the difference between the measured tide level and the astronomical tide level at the same time, and obtain the storm surge target variable;

[0011] S2. Construction of Multi-Source Physical Enhancement Feature System: The input feature set is constructed from four dimensions: historical water increase inertia features, current physical forcing features, physical evolution trend features, and directional coupling features, and integrates site water increase memory and typhoon physical forcing information;

[0012] S3. Physically Enhanced Prediction Model Construction: Based on the extreme gradient boosting regressor as the basic framework, the Optuna hyperparameter optimization algorithm is combined to complete parameter optimization and construct a physically enhanced extreme gradient boosting prediction model.

[0013] S4. Event Group Cross-Validation Training: The training set and test set are divided into complete typhoon events as indivisible units. Event group cross-validation is performed to avoid information leakage caused by the cross-set distribution of samples of the same typhoon event.

[0014] S5. Real-time Storm Gain Prediction: Input the input feature set corresponding to the time to be predicted into the trained model, and output the real-time storm gain prediction value for the corresponding time along the Lianyungang coast.

[0015] Optionally, in step S1, the elevation datum is the abandoned Yellow River datum; the typhoon path intensity data covers 28 typhoon events that significantly affected the flood level of Lianyungang between 2002 and 2024, with a total sample size of 2168 groups, and the storm surge values ​​range from -0.82m to 1.47m.

[0016] Optionally, in step S2, the input feature set contains a total of 41 features, including 13 historical water increase inertia features, 9 current physical forcing features, 15 physical evolution trend features, and 4 directional coupling features.

[0017] Optionally, the historical water increase inertia characteristics include hourly lagging water increase, maximum water increase over a set time period, water increase difference over multiple time periods, and water increase change slope, which are used to characterize the recent water increase level, short-term persistence, and increase / decrease trend of the station.

[0018] Optionally, the current physical forcing characteristics include the typhoon's maximum wind speed, central pressure, pressure deficit, spherical distance from the typhoon center to the tide gauge station, relative azimuth cosine, relative displacement in the east-west and north-south directions, and astronomical tide level, which are used to characterize the typhoon's forcing intensity, spatial location, and tide level background at the current moment.

[0019] Optionally, the physical evolution trend characteristics include the changes and rates of change of typhoon wind speed, air pressure, and relative distance on 6-hour and 12-hour scales, which are used to characterize the short-term evolution of typhoon intensity and relative position.

[0020] Optionally, the directional coupling features include a combination of the square of wind speed and the cosine of azimuth, and a combination of pressure deficit and distance, used to characterize the differences in the effect of wind pressure forcing on water increase under different relative azimuths of typhoons.

[0021] Optionally, in step S3, the extreme gradient boosting regressor uses an additive regression tree to fit the nonlinear mapping relationship between the input features and storm surge. The training process is constrained by the loss function and the model complexity regularization term to control the risk of model overfitting.

[0022] Optionally, in step S4, all samples of a single typhoon event belong to only one set in the training set or the test set, and cannot be split across sets; the model evaluation metrics include mean absolute error, root mean square error, coefficient of determination, Pearson correlation coefficient, and systematic bias.

[0023] Optionally, in step S4, an inner validation set is set up under the event grouping framework for early model stopping, while the outer test set does not participate in model training and parameter tuning, and is used to evaluate the model's generalization ability on independent typhoon events.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. The real-time storm surge prediction method based on physical enhancement XGBoost for Lianyungang coast provided by this invention constructs a 41-dimensional multi-source physical enhancement feature system covering historical storm surge inertia, current physical forcing, physical evolution trends, and directional coupling features. This achieves deep integration of site storm surge memory and typhoon physical forcing, enabling multi-dimensional characterization of the storm surge evolution driving mechanism and effectively improving prediction accuracy and physical interpretability. It employs an event grouping cross-validation strategy with complete typhoon events as indivisible units, fundamentally avoiding information leakage caused by the cross-set distribution of adjacent samples of the same typhoon. Combined with inner and outer layer validation frameworks, this ensures that the evaluation results fit actual application scenarios, significantly improving the model's generalization ability and prediction stability on independent typhoon events. Simultaneously, it uses an extreme gradient boosting regressor as its basic architecture, combined with Optuna... Hyperparameter optimization and regularization constraint training reduce the risk of overfitting while ensuring the efficiency of real-time prediction computation, and the prediction performance is better than the traditional benchmark model. Moreover, the method is customized for the regional geomorphology and hydrodynamic characteristics of Haizhou Bay in Lianyungang, with strong adaptability, and can provide reliable data support for storm surge disaster early warning, port safety operation and maintenance, and disaster prevention and mitigation decision-making in Lianyungang.

[0026] 2. In this invention, a 41-dimensional input feature set is constructed from four dimensions: historical flood inertia characteristics, current physical forcing characteristics, physical evolution trend characteristics, and directional coupling characteristics. On the one hand, the short-term continuous pattern of station flood is captured through historical flood lag information, rolling statistics, and change slope. On the other hand, the system integrates physical information such as typhoon wind pressure forcing, spatial relative position, 6-12 hour scale physical evolution trend, and directional coupling effect, realizing the deep integration of historical flood memory and typhoon physical forcing. It can characterize the evolution driving mechanism of storm flood in multiple dimensions and effectively improve the model prediction accuracy and physical interpretability.

[0027] 3. In this invention, the training set and test set are divided into complete typhoon events as the smallest indivisible unit, ensuring that all samples of a single typhoon event belong to only the same set, thus avoiding the information leakage problem caused by the cross-set distribution of adjacent samples of the same typhoon from the root. At the same time, an inner and outer layer validation framework is constructed, with the inner layer validation set supporting the model's early stopping mechanism, and the outer layer test set not participating in the training and parameter tuning process, which is dedicated to evaluating the model's generalization performance on independent typhoon events, making the model evaluation results more in line with actual application scenarios and significantly improving the reliability and stability of the prediction results.

[0028] 4. In this invention, an extreme gradient boosting regressor is used as the basic architecture. An additive regression tree is used to fit the nonlinear mapping relationship between input features and storm surge. The Optuna hyperparameter optimization algorithm is combined to achieve automated parameter optimization. The training process is constrained by the loss function and the model complexity regularization term. While ensuring the efficiency of real-time prediction computation, the risk of overfitting is further reduced. The model also integrates the fitting advantages of data-driven modeling and the mechanism constraints of physical features, and has better prediction error performance than traditional benchmark models.

[0029] 5. In this invention, a model is custom-developed for the geomorphological and hydrodynamic characteristics of Haizhou Bay in Lianyungang. Based on 28 typical typhoon events affecting the region between 2002 and 2024, the model can accurately adapt to the storm surge response pattern of Lianyungang station. The prediction method has high computational efficiency and can realize the real-time output of storm surge data. The prediction results can provide direct data support for storm surge disaster early warning, port safety operation and maintenance, and disaster prevention and mitigation decision-making along the Lianyungang coast. Attached Figure Description

[0030] Figure 1 This is a map showing the research area of ​​the real-time storm surge prediction method for the Lianyungang coastal area based on Physically Enhanced XGBoost, as described in this invention.

[0031] Figure 2 This is a comparison of the RMSE and R² of the physical enhancement XGBoost of this invention with five benchmark models on the outer test set;

[0032] Figure 3 This is a comparison chart of observed and predicted storm surge processes in highly correlated typhoon events according to the present invention;

[0033] Figure 4 This is a comparison chart of observed and predicted storm surge processes in a typhoon event with low error stability, as presented in this invention.

[0034] Figure 5 This is a comparison chart of observed and predicted storm surge processes in typhoon events where the present invention underestimates and has local biases.

[0035] Figure 6This is a system block diagram of the real-time storm surge prediction method for the Lianyungang coastal area based on Physically Enhanced XGBoost of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please refer to Figures 1 to 6. This embodiment provides a real-time prediction method for storm surge along the Lianyungang coast based on a physically enhanced extreme gradient boosting model. The method includes acquiring measured tide level data, astronomical tide level data, and typhoon path and intensity data from the Lianyungang tide gauge station. The measured tide level data, astronomical tide level data, and typhoon path and intensity data are preprocessed to obtain storm surge target variables. Input features are constructed, and a machine learning regression model is trained to output prediction results. The specific steps include:

[0038] S1. Data preprocessing and target variable construction: Unify the measured tide level data and the astronomical tide level data to the same elevation datum, calculate the difference between the measured tide level and the astronomical tide level at the same time, and obtain the storm surge target variable;

[0039] S2. Construction of Multi-Source Physical Enhancement Feature System: The input feature set is constructed from four dimensions: historical water increase inertia features, current physical forcing features, physical evolution trend features, and directional coupling features, and integrates site water increase memory and typhoon physical forcing information;

[0040] S3. Physically Enhanced Prediction Model Construction: Based on the extreme gradient boosting regressor as the basic framework, the Optuna hyperparameter optimization algorithm is combined to complete parameter optimization and construct a physically enhanced extreme gradient boosting prediction model.

[0041] S4. Event Group Cross-Validation Training: The training set and test set are divided into complete typhoon events as indivisible units. Event group cross-validation is performed to avoid information leakage caused by the cross-set distribution of samples of the same typhoon event.

[0042] S5. Real-time Storm Gain Prediction: Input the input feature set corresponding to the time to be predicted into the trained model, and output the real-time storm gain prediction value for the corresponding time along the Lianyungang coast.

[0043] To eliminate systematic biases caused by different elevation benchmarks and accurately extract storm surge signals driven by typhoons, providing a unified and reliable target variable for model training, in step S1 of this embodiment, data preprocessing and target variable construction, the measured tide level data and astronomical tide level data are unified to the abandoned Yellow River elevation benchmark. The difference between the measured tide level and the astronomical tide level at the same moment is calculated to obtain the storm surge target variable, as shown in equation (1). This embodiment uses 28 typhoon events that significantly affected the storm surge in Lianyungang between 2002 and 2024 to construct a sample set. The total number of modeling samples is 2168 groups, and the target storm surge value ranges from -0.82m to 1.47m. The detailed parameters of the dataset are shown in Table 1. By calculating the residuals after unifying the benchmark, the interference of astronomical tide cycle fluctuations can be effectively eliminated, the storm surge component can be accurately separated, and the consistency and accuracy of the data basis for subsequent model training can be guaranteed.

[0044] The target variable for storm surge is defined as the difference between the measured tide level and the astronomical tide level at the same moment. To ensure comparability between the two types of water level data, the measured tide level and the astronomical tide level are first unified to the abandoned Yellow River datum before the residuals are calculated. The calculation formula is as follows:

[0045]

[0046] in, For a moment Storm surge To unify the measured tide levels after datum alignment, These are the astronomical tide levels under the same datum.

[0047] Table 1. Dataset Sample Information

[0048]

[0049] To characterize the evolutionary driving mechanism of storm surge from multiple dimensions, and to integrate short-term continuous memory of storm surge and typhoon physical forcing information, thereby enhancing the physical interpretability and prediction accuracy of the model, in step S2 of this embodiment, the construction of a multi-source physical enhancement feature system is carried out. A total of 41-dimensional input feature set is constructed from four dimensions: historical storm surge inertia features, current physical forcing features, physical evolution trend features, and directional coupling features. The number and physical meaning of each type of feature are shown in Table 2.

[0050] Table 2 Feature Dataset

[0051]

[0052] To capture the short-term continuity and dynamic trends of water level increases at monitoring stations and preserve the temporal inertia information of historical water level increases, this embodiment sets 13 historical water level increase inertia features, including hourly lagging water level increases, maximum water level increases over a set time period, water level difference across multiple time periods, and water level change slope. These features characterize the recent water level increase, short-term continuity, and the changing patterns of water level rises or falls at the monitoring stations. These features allow the model to fully utilize the temporal memory of water level evolution, improving the fitting effect on continuous water level increase processes.

[0053] To comprehensively characterize the current typhoon's forcing intensity, spatial relative position, and tidal background conditions, this embodiment sets nine current physical forcing features, including the typhoon's maximum wind speed, central pressure, pressure deficit, spherical distance from the typhoon center to the tide gauge station, relative azimuth cosine, east-west and north-south relative displacement, and astronomical tide level. These features directly reflect the real-time intensity and spatial relationship of the typhoon's external force on the station, providing clear physical forcing inputs for the model and enhancing the physical rationality of the prediction results.

[0054] To characterize the short-term evolution of typhoon intensity and relative position, and to reflect the continuous impact of typhoon dynamics on flooding, this embodiment sets 15 physical evolution trend features, covering the changes and rates of change of typhoon wind speed, air pressure, and relative distance at 6-hour and 12-hour scales. By introducing evolution information at multiple time scales, the model can perceive the development trend of the typhoon, improve its ability to predict the dynamic changes in flooding, and adapt to the lag characteristics of flooding response during typhoon movement.

[0055] To characterize the differences in the effect of wind pressure forcing on water level rise under different relative typhoon orientations and to reflect the regulatory role of directional coupling on nearshore water level rise in Haizhou Bay, this embodiment sets up four directional coupling features, including a combination of the square of wind speed and the cosine of azimuth, and a combination of pressure deficit and distance. These features further strengthen the constraints of the physical mechanism, can reproduce the differentiated impact of typhoons from different directions on regional water level rise, and improve the model's adaptability to complex typhoon paths and the prediction accuracy of extreme scenarios.

[0056] To ensure efficient real-time prediction while enhancing the model's nonlinear fitting ability and overfitting resistance, in step S3 of this embodiment, the physical augmentation prediction model is built using an extreme gradient boosting regressor as the basic framework. An additive regression tree is employed to fit the nonlinear mapping relationship between input features and storm surge. During training, the loss function and model complexity regularization term are used to jointly constrain the model and control the risk of overfitting. Simultaneously, the Optuna hyperparameter optimization algorithm is used to automate parameter optimization, resulting in a physical augmentation extreme gradient boosting prediction model. Detailed parameters of the dataset are shown in Table 3. This model architecture retains the prediction advantages of the extreme gradient boosting algorithm in small-sample, nonlinear scenarios, while achieving a deep integration of data-driven and physical mechanisms through multi-source physical feature inputs. Under the same validation conditions, the prediction error is lower than that of traditional machine learning benchmark models.

[0057] Table 3. Model hyperparameters after Optuna optimization

[0058]

[0059] To fundamentally avoid information leakage caused by the cross-set distribution of adjacent samples from the same typhoon and ensure that the model evaluation results truly reflect the generalization ability on independent typhoon events, in step S4 of this embodiment, event grouping cross-validation training, the training set and test set are divided using complete typhoon events as indivisible units. This ensures that all samples of a single typhoon event belong to only one set in either the training or test set, and are not split across sets. This method employs strict event grouping cross-validation, with the specific division shown in Table 4 below. The outer layer validation uses 5-fold event grouping cross-validation and is stratified according to the peak water level of the typhoon event, so that complete typhoon events are treated as indivisible units entering the training or test fold. Outer layer test events do not participate in model training, parameter tuning, or early termination processes; they are only used for final test evaluation. The prediction results of each outer layer test fold are summarized and used as the main evaluation basis for the model's generalization performance.

[0060] Table 4 Overview of Outer Event Grouped Cross-Validation

[0061]

[0062] To further enhance the reliability of model training and the objectivity of evaluation results, this embodiment sets an inner validation set for early model stopping within the event grouping framework. The outer test set is not involved in model training or parameter tuning throughout the process and is dedicated solely to evaluating the model's generalization ability on independent typhoon events. Model evaluation metrics include mean absolute error, root mean square error, coefficient of determination, Pearson correlation coefficient, and systematic bias. This validation strategy avoids inflated model evaluation results, ensuring the trained model exhibits more stable and reliable predictive performance in actual typhoon events.

[0063] To evaluate the relative performance of the physics-enhanced XGBoost model, this paper sets up five benchmarks: Persistence, LinearRegression, SVR, Random Forest, and XGBoost. All benchmarks use the same outer event grouping evaluation logic as the main model.

[0064] To examine the impact of different information sources on the model's predictive performance, this method conducts ablation experiments on representative feature sets within a strict event-grouped cross-validation framework. Specific experimental settings are shown in Table 5 below. The ablation experiments use a complete 41-feature model as a baseline. The outer test event partitioning, inner validation rules, early stopping strategy, and XGBoost parameter settings are all consistent with the main model; only the input feature set is changed. Representative ablation settings include: removing all historical hysteresis features, removing historical hysteresis trend features, removing distance-spatial relationship features, removing astronomical tide background features, and removing current physical forcing features.

[0065] Table 5. Ablation experimental setup for representative characteristic groups

[0066]

[0067] Under the final evaluation on the strict event grouping outer test set, the Physically Enhanced XGBoost model achieved an MAE of 0.0819 m, an RMSE of 0.1282 m, an R² of 0.7941, a Corr of 0.8928, and a Bias of -0.0024 m on 2168 samples. These results indicate that, under the evaluation condition where complete typhoon events are independently excluded, the model can reconstruct the overall change process of storm surge at Lianyungang station relatively well, with an overall average bias close to zero.

[0068] The results of the five outer test layers show that the model performance fluctuates somewhat between different event combinations, but the overall model remains relatively stable. The overall performance and the performance of each layer are shown in Table 6. The RMSE of each layer ranges from 0.1054 to 0.1404 m, the R² ranges from 0.7672 to 0.8042, and the Corr ranges from 0.8761 to 0.9104. The biases of each layer are -0.0227, 0.0059, -0.0143, -0.0023, and 0.0152 m, respectively. This indicates that there are slight positive and negative deviations under different typhoon event combinations, but no obvious systematic overall shift was observed.

[0069] Table 6 Overall and analytical performance of Physically Enhanced XGBoost on the outer test set

[0070]

[0071] In order to achieve efficient and real-time output of storm surge data and support the real-time decision-making needs for coastal storm surge disaster early warning and disaster prevention and mitigation, in step S5 of this embodiment, the multi-source input feature set corresponding to the time to be predicted is input into the trained physical augmentation extreme gradient boosting model, and after rapid calculation, the real-time storm surge prediction value of the corresponding time along the coast of Lianyungang is output.

[0072] Figure 2 shows the comparison of RMSE and R² between Physics-Enhanced XGBoost and the five benchmark models. Table 7 shows the complete evaluation metrics for each model. The Persistence model achieved a MAE of 0.0829 m, an RMSE of 0.1340 m, an R² of 0.7719, a Corr of 0.8869, and a Bias of -0.0005 m. Physics-Enhanced XGBoost reduced the RMSE by approximately 4.31% and increased the R² by approximately 0.0222 compared to Persistence.

[0073] The RMSE for Random Forest is 0.1459 m and R² is 0.7334; the RMSE for the base XGBoost is 0.1488 m and R² is 0.7227; the RMSE for Linear Regression is 0.1549 m and R² is 0.6996; and the RMSE for SVR is 0.1773 m and R² is 0.6064. Compared to these benchmarks, the physics-enhanced XGBoost shows a reduction in RMSE of approximately 12.12%, 13.83%, 17.22%, and 27.68%, respectively, and an improvement in R² of 0.0607, 0.0714, 0.0945, and 0.1878, respectively.

[0074] Table 7 Performance comparison of Physically Enhanced XGBoost and the baseline model on the outer test set

[0075]

[0076] Working principle:

[0077] This method, based on a data-driven and physical mechanism-integrated approach, conducts real-time storm surge prediction for the semi-enclosed, low-lying nearshore area of ​​Haizhou Bay in Lianyungang. First, it preprocesses three types of multi-source baseline data—measured tide level, astronomical tide level, and typhoon track and intensity—unifying them to the abandoned Yellow River elevation datum. Then, it separates the storm surge target variable by the difference between the measured and astronomical tide levels, eliminating interference from astronomical tidal fluctuations and datum differences, and constructing standardized modeling samples. Second, it constructs 41 models based on four dimensions: historical storm surge temporal inertia, real-time typhoon physical forcing, short-term typhoon evolution trend, and azimuth coupling effect. The system employs a comprehensive physical enhancement feature framework, fully covering the intrinsic continuity and external driving factors of storm surge, providing the model with clear physical meaning inputs, strengthening the model's physical constraints and interpretability, and compensating for the lack of physical features in conventional models. Subsequently, an extreme gradient boosting regressor is used as the core architecture, combined with hyperparameter optimization algorithms and regularization constraints to build a prediction model, balancing the model's nonlinear fitting ability and anti-overfitting performance to ensure efficient prediction computation. The model training process adopts a strict event-group cross-validation strategy, using complete typhoon events as the smallest sample unit, and employing inner and outer validation frameworks to achieve complete isolation between training, validation, and testing, thoroughly avoiding sample information leakage and ensuring that the model maintains stable accuracy even on unseen independent typhoon events, truly reflecting the model's actual generalization ability. Finally, by inputting the corresponding features of the time to be predicted into the trained model, the storm surge prediction results can be quickly output, achieving high-precision and high-reliability real-time storm surge prediction along the Lianyungang coast, providing reliable data support for regional storm surge disaster early warning, port safety operation and maintenance, and disaster prevention and mitigation decision-making.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time storm surge prediction method based on Physically Enhanced XGBoost (PEX) for the Lianyungang coastal area, comprising acquiring measured tide level data, astronomical tide level data, and typhoon path and intensity data from the Lianyungang tide gauge station, characterized in that: The measured tide data, astronomical tide data, and typhoon track and intensity data are preprocessed to obtain the storm surge target variable. Input features are constructed, and a machine learning regression model is trained to output prediction results. The specific steps include: S1. Data preprocessing and target variable construction: Unify the measured tide level data and the astronomical tide level data to the same elevation datum, calculate the difference between the measured tide level and the astronomical tide level at the same time, and obtain the storm surge target variable; S2. Construction of Multi-Source Physical Enhancement Feature System: The input feature set is constructed from four dimensions: historical water increase inertia features, current physical forcing features, physical evolution trend features, and directional coupling features, and integrates site water increase memory and typhoon physical forcing information; S3. Physically Enhanced Prediction Model Construction: Based on the extreme gradient boosting regressor as the basic framework, the Optuna hyperparameter optimization algorithm is combined to complete parameter optimization and construct a physically enhanced extreme gradient boosting prediction model. S4. Event Group Cross-Validation Training: The training set and test set are divided into complete typhoon events as indivisible units. Event group cross-validation is performed to avoid information leakage caused by the cross-set distribution of samples of the same typhoon event. S5. Real-time Storm Gain Prediction: Input the input feature set corresponding to the time to be predicted into the trained model, and output the real-time storm gain prediction value for the corresponding time along the Lianyungang coast.

2. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 1, characterized in that: In step S1, the elevation datum is the abandoned Yellow River datum; the typhoon path intensity data covers 28 typhoon events that significantly affected the water level increase in Lianyungang between 2002 and 2024, with a total sample size of 2168 groups and storm surge values ​​ranging from -0.82m to 1.47m.

3. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 1, characterized in that: In step S2, the input feature set contains a total of 41 features, including 13 historical water increase inertia features, 9 current physical forcing features, 15 physical evolution trend features, and 4 directional coupling features.

4. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 3, characterized in that: The historical water increase inertia characteristics include hourly lagging water increase, maximum water increase over a set time period, water increase difference over multiple time periods, and water increase change slope, which are used to characterize the recent water increase level, short-term sustainability, and increase / decrease trend of the station.

5. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 3, characterized in that: The current physical forcing characteristics include the typhoon's maximum wind speed, central pressure, pressure deficit, spherical distance from the typhoon center to the tide gauge station, relative azimuth cosine, relative displacement in the east-west and north-south directions, and astronomical tide level, which are used to characterize the typhoon's forcing intensity, spatial location, and tide level background at the current moment.

6. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 3, characterized in that: The physical evolution trend characteristics include the changes and rates of change of typhoon wind speed, air pressure, and relative distance on 6-hour and 12-hour scales, which are used to characterize the short-term evolution of typhoon intensity and relative position.

7. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 3, characterized in that: The directional coupling features include a combination of the square of wind speed and the cosine of azimuth, and a combination of pressure deficit and distance, which are used to characterize the differences in the effect of wind pressure forcing on water increase under different relative azimuths of typhoons.

8. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 1, characterized in that: In step S3, the extreme gradient boosting regressor uses an additive regression tree to fit the nonlinear mapping relationship between the input features and storm surge. The training process is constrained by the loss function and the model complexity regularization term to control the risk of model overfitting.

9. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 1, characterized in that: In step S4, all samples of a single typhoon event belong to only one set in the training or test set and cannot be split across sets; the model evaluation metrics include mean absolute error, root mean square error, coefficient of determination, Pearson correlation coefficient, and systematic bias.

10. The real-time storm surge prediction method based on Physically Enhanced XGBoost for Lianyungang coastal areas according to claim 9, characterized in that: In step S4, an inner validation set is set up within the event grouping framework for early model stopping, while the outer test set does not participate in model training and parameter tuning. This set is used to evaluate the model's generalization ability on independent typhoon events.

Images