Typhoon wave forecasting method based on integrated machine learning

Abstract

The application discloses a typhoon wave forecasting method based on integrated machine learning. First, the application integrates historical typhoon wave data, meteorological data and marine environment data; the data is preprocessed, including integration, cleaning, filling and standardization; K nearest neighbor algorithm and spline interpolation method are used for multi-source data completion; secondly, key feature parameters are screened through Pearson correlation coefficient, and mutual information method is combined to strengthen nonlinear correlation representation; then, an integrated prediction model containing LSTM, XGBoost and Transformer is constructed; finally, a dynamic time sequence division strategy is used to divide the training set and the verification set, and the model hyperparameters are optimized, and the training and testing of the typhoon wave height forecasting model are completed. The application solves the problem of data sparsity through multi-source data fusion and feature selection optimization, improves the generalization ability through the integrated model architecture, and significantly shortens the training period, improves the prediction accuracy and timeliness compared with the traditional single model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of typhoon wave prediction, and particularly relates to a typhoon wave prediction method based on integrated machine learning. BACKGROUND

[0002] Meteorological disasters have caused great losses to human society, among which typhoon is a common and serious natural disaster. Typhoon, as a unique form of tropical cyclone in the northwest Pacific region, has the characteristics of low central pressure and strong wind, and can cause extreme weather phenomena, including gale, heavy rain and storm surge, which can cause serious damage to human life and natural environment.

[0003] Typhoon wave is an important marine disaster induced by typhoon, and it is one of the main factors affecting coastal areas. Typhoon wave can cause serious harm to marine transportation, offshore engineering construction and fishing in the nearshore area, and can cause damage to buildings such as dikes and piers in the coastal area. At the same time, sea waves are the main dynamic elements that cause sediment movement, and the wave upwelling in the nearshore area can affect the sediment carrying capacity of the flow, easily causing beach erosion, severe erosion and deposition of the bottom bed, and causing siltation of the channel and harbor.

[0004] The variability and complexity of marine environment bring challenges to traditional prediction methods, which mainly rely on numerical weather prediction models and empirical formulas. Although these models and formulas can provide information on typhoon waves to some extent, due to the complexity and uncertainty of ocean dynamics, their prediction results often have large errors.

[0005] There is usually more than 30% prediction deviation of wave height in the practical application of traditional SWAN model. SWAN model requires high accuracy of wind field, and the prediction result may be limited by the climate model, which amplifies the error of the climate model prediction. Currently, some scholars have proposed a classification framework based on Transformer encoder and empirical mode decomposition (EMD) for predicting significant wave height (SWH), or combined time series decomposition with machine learning to optimize prediction performance, but the optimization effect is general. Some scholars combined SWAN model and Self-Attention ConvLSTM network for time series prediction, but there are problems such as not focusing on typhoon events, not verifying the performance under extreme typhoon conditions, or not optimizing the parameters of SWAN model and deep integration with machine learning. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application proposes a typhoon wave forecasting method based on integrated machine learning. The method uses cutting-edge machine learning technology to optimize the parameter scheme of typhoon wave numerical forecasting, adopts the extreme gradient boosting algorithm with the best overall forecasting effect, uses multi-base model combination and explicit weighting strategy to establish a wave forecasting model coupled with SWAN and the extreme gradient boosting algorithm, and comprehensively analyzes historical and real-time meteorological and marine environmental data to construct an efficient and accurate typhoon wave prediction model.

[0007] The present application comprises the following steps:

[0008] Data collection: Obtain multi-source heterogeneous data, including historical typhoon wave data, meteorological observation data, marine environmental data and human activity interference data;

[0009] Data preprocessing: The collected data is processed as follows:

[0010] Temporal and spatial alignment: The time resolution of data from different sources is unified to a preset interval;

[0011] Outlier processing: Based on the physical constraint condition, the observation values exceeding the reasonable range are identified and removed;

[0012] Missing value filling: The continuous missing data is filled by using the interpolation algorithm, and the discrete missing data is filled by using the interpolation algorithm;

[0013] Data standardization: Map each feature parameter to a unified numerical range;

[0014] Feature extraction: Through correlation analysis, the key feature parameters with high correlation with wave height prediction are screened, and the input feature set is determined by combining the nonlinear feature selection method;

[0015] Model construction: Construct an integrated machine learning model, adopt a coupling method of numerical model and machine learning algorithm, use multiple base model combination and explicit weighting strategy or error correction mechanism to establish a wave forecasting model;

[0016] Model training and optimization: Train the constructed integrated machine learning model using historical data, adjust the model parameters and structure to optimize the model performance, and evaluate the model by using cross-validation method to ensure the generalization ability and stability of the model;

[0017] Real-time forecasting: Input the real-time collected meteorological and marine data into the trained integrated machine learning model, and output the typhoon wave height prediction value and confidence interval in the future N hours through the multi-channel output of the coupled integrated machine learning model;

[0018] Result output and verification: generate a time-space distribution thermodynamic map, a prediction-observation comparison curve, and an error statistics panel, calculate the root mean square error, and iteratively optimize the model.

[0019] Preferably, in the model construction step, the integrated machine learning model comprises:

[0020] SWAN-ML-SSA weight ratio model: by integrating the SWAN numerical model and the machine learning algorithm, and dynamically optimizing the weight ratio using the sparrow search algorithm, the prediction accuracy and adaptability of the model are improved.

[0021] Preferably, the calculation process of the SWAN-ML-SSA weight ratio model comprises:

[0022] Obtain the effective wave height prediction sequence of the SWAN model and the machine learning model respectively;

[0023] Take 70% of each data set as the training set and 30% as the test set;

[0024] Take the wind speed and wave height of the first 1-3 hours as input items, and fuse the outputs of the SWAN model and the machine learning model through weight ratio;

[0025] Use the sparrow search algorithm to perform global optimization in the weight space to determine the optimal mixing ratio of the outputs of the SWAN model and the machine learning model;

[0026] Use Python to build a dual-model coupling framework, use the NetCDF4 library to real-time parse the grid wave field data output by the SWAN model, and use the machine learning model to load the pre-trained wave height prediction model.

[0027] Preferably, in the prediction phase of the SWAN-ML-SSA weight ratio model:

[0028] Build a prediction framework through a recurrent neural network, and form an input tensor by combining the wind field data output by SWAN and the historical wave height sequence;

[0029] The output end uses a weighted fusion layer to dynamically combine the SWAN physical prediction and the machine learning prediction;

[0030] The sparrow search algorithm optimizes the weight parameters every 6 hours online to ensure that the model adapts to the typhoon path mutation scenario.

[0031] Preferably, in the model construction step, the integrated machine learning model further comprises:

[0032] SWAN-ML error loss model: by constructing an error prediction-compensation mechanism, the systematic error of the SWAN numerical model is corrected, and the prediction accuracy of the model is improved.

[0033] Preferably, the calculation process of the SWAN-ML error loss model comprises:

[0034] Based on historical data, a SWAN simulation error database is established using Pandas, and the deviation between the effective wave height and wave period output by SWAN and the measured value at the corresponding time is taken as the training target;

[0035] An error prediction model is constructed using a machine learning algorithm, and the input features include wave elements, wind field parameters and terrain feature parameters output by SWAN in the previous 3 hours;

[0036] 70% of each data set is used as a training set, and 30% is used as a test set, and the dominant factors are selected through feature importance analysis;

[0037] The error compensation model is trained to learn the simulation deviation rule of SWAN in complex terrain areas, and the real-time output of SWAN is imported into the error compensation model in the prediction stage, and the corrected effective wave height is output.

[0038] Preferably, the SWAN-ML error loss model in the prediction stage:

[0039] A series structure of SWAN physical model and error compensation model is formed, and a GPU accelerated training module of a machine learning model is used;

[0040] The loss function is defined as Huber loss to enhance the robustness to extreme wave height and improve the prediction accuracy of typhoon eye passage.

[0041] Preferably, in the feature extraction step, the Pearson correlation coefficient is used to analyze and select key feature parameters with a correlation greater than 0.7 with wave height, and the mutual information method is used to strengthen nonlinear feature selection.

[0042] Preferably, the model training stage adopts a dynamic time sequence division strategy, and the data set is divided into a training set, a validation set and a test set according to time continuity, and the adaptive cross-validation framework is used to optimize the hyperparameter combination.

[0043] Preferably, in the real-time prediction step, the integrated model can automatically switch modes according to different prediction requirements, such as automatically calling a more stable SWAN-ML error loss model when simulating the maximum effective wave height of a typhoon wave, and calling a SWAN-ML-SSA weight matching model if higher overall accuracy is required.

[0044] The beneficial effects of the present application mainly include the following aspects:

[0045] 1. Improved prediction accuracy: through machine learning technology, the model can automatically learn and extract the complex characteristics of typhoon wave formation and evolution, thereby improving the accuracy of prediction.

[0046] 2. Strong versatility: This method can adapt to different intensity and different path of typhoon wave forecasting needs, and can make more accurate prediction of typhoon wave under various marine environment and meteorological conditions.

[0047] 3. Cost savings: Traditional typhoon wave forecasting relies on expensive observation equipment and human resources, while the integrated machine learning-based forecasting method can significantly reduce the cost of forecasting through algorithm optimization and automation, and has higher timeliness and economy compared with traditional numerical models.

[0048] 4. Technical promotion potential: The prediction method can be applied to other types of marine and meteorological disasters, and has broad application prospects and market potential. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flowchart of the method of the present application;

[0050] Figure 2 is a schematic diagram of the data filling effect in the implementation of the present application;

[0051] Figure 3 is a schematic diagram of the calculation process of the SWAN-ML-SSA weight matching model;

[0052] Figure 4 is a schematic diagram of the calculation process of the SWAN-ML error loss model;

[0053] Figure 5 is a scatter plot of the prediction results of the single ML model and the actual observation values;

[0054] Figure 6 is a scatter plot of the prediction results of the SWAN-ML-SSA weight matching model and the actual observation values. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0056] The specific implementation of the present application will be described in detail below in combination with specific examples:

[0057] As shown in Figure 1 , one embodiment of the present application provides a typhoon wave forecasting method based on integrated machine learning, which constructs a SWAN-ML hybrid prediction architecture, adopts a coupling method of numerical model and machine learning, and uses a hierarchical architecture with parallel running of weight matching and error correction dual channels. The method specifically includes the following steps:

[0058] I. Data Collection:

[0059] Firstly, historical typhoon wave data, meteorological data (wind speed, wind direction, air pressure), and marine environmental data (tide, ocean current, etc.) are collected, as well as other influencing factors including tide, ocean current, seabed topography, and human activity interference data. Data sources include but are not limited to the following ways:

[0060] 1. Buoy monitoring: Buoy deployed in the ocean can monitor real-time meteorological and marine environmental data on the sea surface, including wind speed, wind direction, air pressure, wave height, wave direction, etc., providing important observation data for typhoon wave forecasting.

[0061] 2. Coastal weather stations / ground observation stations: Coastal weather stations can monitor and record meteorological data during the landing process of typhoon, including rainfall, air temperature, air pressure, etc., providing meteorological data for the coastal area for the prediction model.

[0062] 3. Numerical model: A numerical model of typhoon waves in the sea area is established based on SWAN, and the optimal parameter scheme is determined through sensitivity analysis, where the wind energy input is selected as Komen mode, the wave breaking parameter γ=0.79, the bottom friction is adopted as Jonswap mode, and the three-wave and four-wave nonlinear interaction is turned on. The wave field during the historical typhoon period is used to verify the adaptability of the parameter scheme to typhoons of different intensities (super typhoon, strong typhoon, typhoon) and paths (westward, northwestward, northward), so as to reduce the average relative error of significant wave height and supplement the data set to provide diversified and more accurate data sources for the prediction model.

[0063] The data covers typhoon events of different intensities and paths, as well as observation values under various marine environmental and meteorological conditions, ensuring the comprehensiveness and representativeness of model training. Through comprehensive analysis and integration of multi-source data, a comprehensive and accurate typhoon wave forecasting model is established to improve the accuracy and reliability of the prediction.

[0064] II. Data Preprocessing:

[0065] 1. Data integration: In order to facilitate subsequent analysis and modeling work, the original data is integrated uniformly to build a mixed data set containing SWAN simulation data and measured buoy data, and spatiotemporal matching technology is used to realize the fusion of grid data and station data. Specifically, it includes:

[0066] Time accuracy unification: The time accuracy of different stations and different test time periods in the original data may be inconsistent. In order to unify the time interval of the data, the data is unified to the same time interval. Based on the time resolution of the collected measured data, which is 10 min to 1 h, the data is generally unified to hourly data to unify the output interval of the model and improve the comparability and consistency of the data.

[0067] Data format standardization: unify and standardize the data formats of different sites and different data sources to ensure consistency, compatibility and operability of data for subsequent data processing and analysis.

[0068] 2. Data cleaning: there are some outliers in the original observation data, which obviously violate the physical law, such as abnormally high wind speed or abnormally low air pressure, etc. These unreasonable data need to be removed to ensure the accuracy and reliability of the data.

[0069] 3. Data filling: in the measured data, due to the need to maintain the observation equipment, there are many wind speed and wave height data with only one valid data, or 1-3 invalid data between two valid data. If these data with missing values are deleted in whole row, the data set will be too small, which will affect the prediction ability of integrated machine learning model.

[0070] When dealing with missing values of meteorological and oceanic elements with certain variation rules, the commonly used methods include K-Nearest Neighbor (KNN) algorithm and spline interpolation method.

[0071] K-Nearest Neighbor algorithm is used to fill in the missing data. This method is a supervised classification machine learning algorithm based on instances. Compared with the ordinary mean method, K-Nearest Neighbor interpolation method has obvious improvement, which can fully consider the samples near the missing samples, so as to supplement the better data. Among them, the heterogeneous Euclidean-overlap metric (HEOM) is generally selected as the similarity measurement standard between samples in K-Nearest Neighbor interpolation method.

[0072] When the wave height and wind speed have missing values longer than 3, the data column is truncated and saved, and after all the processing, a complete data set is formed.

[0073] Spline interpolation is a flexible interpolation method that uses a set of basis functions called splines to approximate data. Spline functions are a series of smooth and intersecting curves that can well adapt to the changes of data and maintain smoothness near the interpolation points. Spline interpolation method is particularly suitable for processing data with certain variation rules, because it can capture the local trends and shapes in the data.

[0074] The spline interpolation method is used to fill in the missing values of the wind speed and significant wave height data of the station. However, this method is also not suitable when there are large areas of missing wave height and wind speed. When the wave height and wind speed have missing values longer than 4, the data column is truncated and saved. After all the processing is completed, a complete data set is formed. As shown in Figure 2 , the missing data is repaired by K-Nearest Neighbor Imputation. Figure 2 In the middle (a) is the original time series with missing data, and (b) is the complete sequence after imputation. It can be seen that the wind speed and wave height curves after imputation maintain physical continuity and do not appear abnormal mutations.

[0075] 4. Standardization / Normalization: Due to the differences in the scale and size range of different data inputs of machine learning, the model is more sensitive to the size of the data. Therefore, this embodiment normalizes the input components, converts the dimensional expression to a dimensionless expression, and the standardized data set can better predict the parameters and obtain more reasonable prediction results. The normalized method divides the entire data set into [0, 1] set.

[0076] III. Model Prediction:

[0077] 1. Feature extraction (i.e., feature engineering stage): According to the demand of typhoon wave prediction, the key meteorological and marine environmental features related to typhoon wave height prediction are extracted from the preprocessed data. Selecting reasonable features from numerous marine data features as input features can effectively avoid overfitting, speed up the training model rate, and quickly obtain good results.

[0078] Correlation analysis is to analyze all features and the prediction target, i.e., the wave height of the sea, to measure the correlation between the two factors, which can effectively extract parameters with strong correlation with the prediction target. Pearson correlation coefficient is used for feature analysis to explore the correlation between wind direction, wind speed, air temperature, air pressure, wave period, average wave direction, and the prediction target (significant wave height). The Pearson correlation coefficient calculation formula is as follows:

[0079]

[0080] where cov(x, y) is the covariance of x and y, σ x and σ y are the standard deviations of x and y, n is the sample size, x i and y i are sample data points, and are the sample means of x and y, respectively, and the Pearson correlation coefficient The value between -1 and 1, the value is positive is a positive correlation, for negative is a negative correlation; Its absolute value is closer to 1, the stronger the correlation, closer to 0, the weaker the correlation. In order to prevent the model from over-learning the feature parameters, leading to the occurrence of over-fitting, only select the feature parameters with correlation greater than 0.7 as input factors.

[0081] 2. Model construction: Based on the extracted features, an integrated machine learning model is constructed to predict the wave height of typhoon waves. According to the characteristics of the data and the complexity of the prediction, select the appropriate model to improve the accuracy and stability of the prediction. Select the extreme gradient boosting (XGBoost) machine learning algorithm.

[0082] 3. Model training and optimization: Use the collected historical data to train the constructed integrated machine learning model, and adjust the model parameters and structure to optimize the model performance. Among them, the sparrow search algorithm (SSA) is used, which is a new type of swarm intelligence optimization algorithm. This algorithm has the characteristics of fast convergence speed, strong optimization ability, and strong robustness.

[0083] At the same time, cross-validation and other methods are used to evaluate the model to ensure its generalization ability and stability. Through continuous iteration and optimization of the model, its prediction effect under different conditions is improved. Specifically, time series cross-validation can be used to preserve the continuity of the time series, and the validation set is always located in the time period after the training set.

[0084] Four, forecast result generation and output:

[0085] 1. Forecast: Input the collected meteorological and marine environmental data into the trained model to generate the wave height prediction results of the typhoon wave.

[0086] 2. Establish SWAN-ML-SSA weight matching model: This model integrates the advantages of SWAN numerical model and XGBoost machine learning algorithm, and uses sparrow search algorithm to dynamically optimize the weight matching.

[0087] First, get the effective wave height prediction sequence of the SWAN model and the XGBoost model respectively, take 70% of each data set as the training set and 30% as the test set, take the wind speed of the first 1~3 hours and the wave height of the first 1~3 hours as the input item, and the output wave height will be matched with the corresponding SWAN numerical model result. After matching, the wave height is used as the input of the next round, that is, the prediction value of this round is recalculated and used as the input of the next round. On the basis of recursive prediction, the SWAN model result is matched to construct a prediction framework based on recursive neural network.

[0088] The objective function is established by the sparrow search algorithm to minimize the root mean square error, and the optimal mixing ratio of SWAN and XGBoost output is determined in the 0-1 weight space. A double-model coupling framework is built using Python, and the grid wave field data (including Hs, Tp, etc.) output by SWAN is analyzed in real time using the NetCDF4 library. At the same time, the pre-trained site wave height prediction model is loaded using the Booster API of XGBoost. The sparrow search algorithm is implemented based on the optimization module of Scipy, and the objective function is defined as the minimization of the validation set RMSE, with the weight parameter space constraint ω∈[0,1]. The core code segment uses matrix operations to accelerate weight iteration calculation.

[0089] The calculation process of the SWAN-ML-SSA weight ratio model is shown in Figure 3 In the prediction phase, an LSTM time series prediction module is built using PyTorch, and the wind field data (U10, WDIR) output by SWAN and the historical wave height sequence are combined into a three-dimensional tensor input. The output end uses a weighted fusion layer to dynamically combine SWAN physical prediction and XGBoost data-driven prediction. The SSA algorithm optimizes the weight parameter every 6 hours online to ensure that the model adapts to the sudden change of typhoon path.

[0090] In actual process, XGBoost model and sparrow search algorithm SSA are combined to speed up model convergence, and the results obtained by different weight ratios are analyzed for error, and the error is selected as the minimum. The result is that when the weight of SWAN model is about 0.3, the weight ratio model has the best effect.

[0091] 3. Establish a SWAN-ML error loss model: This model focuses on correcting the systematic error of the SWAN numerical model and builds an error prediction-compensation mechanism.

[0092] First, based on historical data, a SWAN simulation error database is established using Pandas, and the effective wave height, wave period, etc. output by SWAN and the measured value deviation at the corresponding time are used as the training target. The XGBoost algorithm is used to build an error prediction model, and the input features include the wave elements, wind field parameters and terrain feature parameters output by SWAN in the previous 3 hours. 70% of each data set is used as the training set and 30% as the test set. The dominant factors are selected through feature importance analysis, and the error compensation model is trained to learn the simulation deviation rules of SWAN in complex terrain areas.

[0093] In the prediction stage, the real-time output of SWAN is introduced into the error compensation model to output the corrected significant wave height, forming a series structure of SWAN physical model and error compensation model, using the GPU accelerated training module of XGBoost, and customizing the loss function as Huber loss to enhance the robustness to extreme wave height. The prediction accuracy of sudden processes such as typhoon eye passage is improved. The SWAN-ML error loss model calculation process is shown in Figure 4 .

[0094] 4. Result output: the final model can automatically switch modes according to requirements. When the maximum significant wave height of the typhoon wave needs to be simulated, the more stable SWAN-ML error loss model is automatically called, and if higher overall accuracy is required, the SWAN-ML-SSA weight matching model is called. The prediction results are presented in graphical and numerical forms, including the predicted typhoon wave height curve and spatial distribution map over time. The prediction results are presented in an intuitive visual way to help users better understand and utilize the prediction information.

[0095] Figure 5 The scatter plot of the prediction results of the single ML (XGBoost) model and the actual observation values is shown in Figure 6 , and the scatter plot of the SWAN-ML-SSA weight matching model is shown in , the R² of the former is 0.86, while the R² of the latter reaches 0.94, and the prediction points of the integrated model are concentrated near the y=x reference line, indicating that the multi-model fusion significantly improves the prediction accuracy.

[0096] 5. Result verification: compare the prediction results with the actual observation values to verify the accuracy of the prediction, and according to the verification results, the model is iteratively optimized, and a typhoon type feature library is established to realize automatic recognition of the typhoon path type and switching of the prediction mode.

[0097] In summary, the present application integrates multi-source meteorological and ocean data, uses K-nearest neighbor algorithm and spline interpolation method for data completion, effectively solves the problem of data missing in typhoon wave prediction; selects key feature parameters through Pearson correlation coefficient, constructs an integrated prediction model, which has stronger generalization ability than single model; uses cross-validation and dynamic time sequence division strategy to significantly improve the training efficiency and prediction accuracy of the model.

[0098] The above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can modify the technical solutions described in the foregoing examples, or make equivalent substitutions for part of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A typhoon wave forecasting method based on integrated machine learning, characterized by, Comprise the following steps: Data collection: acquire multi-source heterogeneous data, including historical typhoon wave data, meteorological observation data, marine environmental data and human activity interference data; Data preprocessing: the collected data are processed as follows: Temporal and spatial alignment: unify the time resolution of data from different sources to a preset interval; Outlier processing: identify and remove observation values outside the reasonable range based on physical constraints; Missing value filling: continuous missing data is filled by interpolation algorithm, and discrete missing data is filled by interpolation algorithm; Data standardization: map each feature parameter to a unified numerical range; Feature extraction: select key feature parameters with high correlation with wave height prediction through correlation analysis, and determine the input feature set by combining nonlinear feature selection methods; Model construction: construct an integrated machine learning model, use the coupling method of numerical model and machine learning algorithm, use multiple base models and clear weighting strategy or error correction mechanism to establish a wave prediction model; Model training and optimization: train the constructed integrated machine learning model using historical data, optimize the model performance by adjusting the model parameters and structure, and evaluate the model by cross-validation method to ensure the generalization ability and stability of the model; Real-time prediction: input real-time collected meteorological and ocean data into the trained integrated machine learning model, and output the typhoon wave height prediction value and confidence interval of future N hours through the multi-channel output of the coupled integrated machine learning model; Result output and verification: generate a space-time distribution heat map, a prediction-observation comparison curve and an error statistics panel, calculate the root mean square error and iteratively optimize the model; In the model construction step, the integrated machine learning model comprises: SWAN-ML-SSA weight ratio model: by integrating SWAN numerical model and machine learning algorithm, using sparrow search algorithm to dynamically optimize weight ratio, to improve the prediction accuracy and adaptability of the model; The integrated machine learning model further comprises: SWAN-ML error loss model: by constructing error prediction-compensation mechanism, correcting the systematic error of SWAN numerical model, improving the prediction accuracy of the model; The calculation process of the SWAN-ML-SSA weight ratio model comprises: Obtain the effective wave height prediction sequence of SWAN model and machine learning model respectively; Take 70% of each data set as training set and 30% as test set; Take the wind speed and wave height of the first 1-3 hours as input, and fuse the outputs of SWAN model and machine learning model through weight ratio; Use sparrow search algorithm to perform global optimization in weight space to determine the optimal mixing ratio of SWAN model and machine learning model output; Use Python to build a double model coupling framework, use NetCDF4 library to real-time parse the grid wave field data output by SWAN model, and use machine learning model to load the pre-trained wave height prediction model; In the prediction stage of the SWAN-ML-SSA weight ratio model: Construct a prediction framework by using recurrent neural network, and input the wind field data output by SWAN and the historical wave height sequence to form an input tensor; The output end dynamically combines the SWAN physical prediction and the machine learning prediction by using a weighted fusion layer; The sparrow search algorithm optimizes the weight parameters every 6 hours online to ensure that the model adapts to the typhoon path mutation scenario; The calculation process of the SWAN-ML error loss model includes: Based on historical data, a SWAN simulation error database is established by using Pandas, and the deviation between the effective wave height and the wave period output by SWAN and the measured value at the corresponding time is taken as the training target; A machine learning algorithm is used to construct an error prediction model, and the input features include wave elements, wind field parameters and terrain feature parameters output by SWAN in the previous 3 hours; 70% of each data set is used as the training set, and 30% is used as the test set, and the dominant factors are selected through feature importance analysis; The error compensation model learns the simulation deviation rules of SWAN in complex terrain areas, and in the prediction stage, the real-time output of SWAN is imported into the error compensation model to output the corrected effective wave height; In the prediction stage, the SWAN-ML error loss model: The series structure of the SWAN physical model and the error compensation model is formed, and the GPU acceleration training module of the machine learning model is used; The loss function is defined as Huber loss to enhance the robustness to extreme wave height and improve the prediction accuracy of the typhoon eye passage; In the real-time prediction step, the integrated machine learning model can automatically switch modes according to different prediction requirements, such as automatically calling the more stable SWAN-ML error loss model when simulating the maximum effective wave height of the typhoon wave, or calling the SWAN-ML-SSA weight ratio model if higher overall accuracy is required. 2.The integrated machine learning based typhoon wave forecasting method according to claim 1, wherein, In the feature extraction step, the Pearson correlation coefficient is used to analyze and select key feature parameters with a correlation greater than 0.7 with wave height, and the mutual information method is used to strengthen nonlinear feature selection. 3.The integrated machine learning based typhoon wave forecasting method according to claim 1, wherein, In the model training stage, a dynamic time sequence division strategy is used to divide the data set into training set, validation set and test set according to time continuity, and the adaptive cross-validation framework is used to optimize the hyperparameter combination.

Images