A data-driven nearshore wind and wave prediction method and system

Abstract

The present invention discloses a data-driven nearshore wind and wave prediction method and system, the method comprising: obtaining multi-source data of long-time series data of various meteorological data variables in the study area and preprocessing, selecting target variables and characteristic variables, dividing the study area into regional grid points; taking the characteristic variables of the open sea grid points as the reference sequence, taking the target variables of the inland sea or nearshore grid points as the target sequence, analyzing the time lag effect and spatial propagation mode between the reference sequence and the target sequence; introducing the analysis results into a deep learning model as prior knowledge, and training the deep learning model to obtain a typhoon prediction model; obtaining real-time data of various meteorological data variables in the study area and preprocessing, inputting the characteristic variables therein into the typhoon prediction model, and obtaining the prediction results of the target variables. The present invention comprehensively considers the time lag effect and spatial relationship of various open sea variables on nearshore waters information, realizes the fusion and integration of multi-source data, and performs efficient and accurate real-time prediction.

Description

Technical Field

[0001] The present invention relates to meteorological forecasting technology, and in particular to a data-driven nearshore wind and wave forecasting method and system. Background Art

[0002] Typhoon is a typical tropical weather extreme event, which seriously threatens the safety of life and property of people in coastal areas and economic development. Typhoon prediction and research can not only reduce the direct impact of typhoon disasters, but also predict extreme weather such as heavy rain and strong winds in advance, which can provide certain help for the early deployment of disaster prevention measures and greatly reduce personal injury and property losses. At present, the prediction of typhoon wave height and other factors during typhoons mainly uses numerical model methods. These models require a lot of computing resources and take a long time to calculate. In addition, the rapid changes in the marine environment put forward higher requirements on the real-time and accuracy of the prediction system. In recent years, with the development of artificial intelligence technology, a new data-driven method has been provided for typhoon prediction, especially deep learning has been widely used in wave prediction. However, most prediction models predict single points, ignoring the cross-correlation between points in space, which inevitably affects the prediction accuracy. Summary of the invention

[0003] The technical problem to be solved by the present invention is as follows: In view of the above-mentioned problems of the prior art, a data-driven nearshore wind and wave prediction method and system are provided, which can comprehensively consider the time lag effect and spatial relationship of various offshore variables on nearshore waters information, realize the fusion and integration of multi-source data, and use a hybrid prediction model to perform efficient and accurate real-time prediction.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A data-driven nearshore wind and wave prediction method comprises the following steps:

[0006] Obtain multi-source data of long-term series data of various meteorological data variables in the study area and preprocess them, select the target variable and find the characteristic variables significantly correlated with the target variable, and divide the study area into regional grid points according to longitude and latitude;

[0007] The characteristic variables of the outer sea grid points are taken as the reference sequence, and the target variables of the inner sea or nearshore grid points are taken as the target sequence, and the time lag effect and spatial propagation pattern between the reference sequence and the target sequence are analyzed;

[0008] The analysis results of time lag effect and spatial propagation mode are introduced into the deep learning model as prior knowledge, and the deep learning model is trained using the long time series data of characteristic variables and target variables to obtain the typhoon prediction model.

[0009] The long time series data of characteristic variables and target variables are input into the typhoon prediction model to obtain the prediction results of the target variables. The real-time observation data of each meteorological data variable in the study area are obtained and preprocessed. The prediction results of the target variables are compared with the corresponding actual observation data. If the accuracy of the prediction results of the target variables meets the requirements, the prediction results of the target variables are output.

[0010] Furthermore, when obtaining multi-source data of long time series data of various variables in the study area and preprocessing them, the preprocessing steps include:

[0011] The multi-source data are complemented and verified to integrate to obtain the feature set, and then the missing values ​​in the feature set are filled, while the abnormal data points in the feature set are identified and processed, and finally the data in the feature set is normalized.

[0012] Furthermore, when selecting the target variable and finding the characteristic variables significantly correlated with the target variable, it includes:

[0013] Calculate the Pearson correlation coefficient between each meteorological data variable and the target variable respectively, and select the meteorological data variable whose absolute value of the Pearson correlation coefficient with the target variable is greater than a preset threshold as the candidate variable;

[0014] A corresponding univariate regression model is established for each candidate variable, and the degree of fit and variable significance of each univariate regression model are evaluated to screen highly correlated variables that have significant positive or negative effects on the target variable;

[0015] A corresponding multivariate regression model was established for each highly correlated variable, and the performance of each multivariate regression model was comprehensively evaluated to screen the characteristic variables with high correlation.

[0016] Furthermore, the analysis of the time lag effect and spatial propagation pattern between the reference sequence and the target sequence includes:

[0017] The first characteristic variable among the characteristic variables of the outer sea grid points is taken as the reference sequence, and the cross-correlation algorithm is used to analyze the time lag effect between the reference sequence and the target sequence;

[0018] The first and second characteristic variables of the characteristic variables of the outer sea grid points are taken as reference sequences, and the dynamic time warping algorithm is used to calculate the optimal matching path between the reference sequence and the target sequence to analyze the nonlinear time lag effect and spatial propagation pattern between the reference sequence and the target sequence.

[0019] Furthermore, when the cross-correlation algorithm is used to analyze the time lag effect between the reference sequence and the target sequence, the expression is as follows:

[0020]

[0021] Among them, R xy (τ) is the cross-correlation function between the reference sequence and the target sequence, τ represents the time lag, x(t) represents the reference sequence, and y(t+τ) represents the target sequence.

[0022] Furthermore, when the dynamic time warping algorithm is used to calculate the optimal matching path between the reference sequence and the target sequence, the expression is as follows:

[0023] D(i,j)=d(x i ,y j )+min(D(i-1,j),D(i,j-1),D(i-1,j-1))

[0024] Where D(i,j) represents the minimum cumulative distance between the first point to the i-th point of the reference sequence x and the first point to the j-th point of the target sequence y; d(x i ,y j ) is the Euclidean distance between the i-th element in the reference sequence x and the j-th element in the target sequence y; D(i-1,j) represents the minimum accumulated distance when the current point of the reference sequence x moves back one step while the current point of the target sequence y remains unchanged in the matching path; D(i,j-1) represents the minimum accumulated distance when the current point of the reference sequence x remains unchanged while the current point of the target sequence y moves back one step in the matching path; D(i-1,j-1) represents the minimum accumulated distance when the current points of both the reference sequence x and the target sequence y move back one step in the matching path.

[0025] Furthermore, the deep learning model adopts the TCN-CNN model, and the analysis results of the time lag effect and the spatial propagation mode are introduced into the deep learning model as prior knowledge, including:

[0026] The sliding window is used to expand the input data from the features of a single time point to a sequence containing multiple time points, and multiple related feature variables are provided to the model as multi-dimensional inputs, expanding the model analysis data dimension;

[0027] The number of TCN and CNN convolutional layers and the size of convolutional kernels are dynamically adjusted according to the time lag effect, and a fully connected layer and an attention layer are added after the TCN and CNN layers to fuse the temporal features and the spatial features, so that the TCN module of the TCN-CNN model effectively captures the long-term dependency and nonlinear time lag effect in the time series through one-dimensional dilated convolution, and the CNN module of the TCN-CNN model realizes spatial feature extraction by using the information of the spatial propagation mode;

[0028] During the training process, the results of the cross-correlation algorithm and the dynamic time warping algorithm are used for pre-training to find the optimal solution more quickly. At the same time, a regularization term is constructed to punish the part of the model prediction results that does not conform to the law of the time lag effect. Specifically, the result of the dynamic time warping algorithm is converted into a part of the regularization term. By measuring the path similarity between the predicted sequence and the actual sequence, the learning process of the model is constrained, so that it pays more attention to the dynamic similarity between sequences.

[0029] Furthermore, after training the deep learning model using long time series data of feature variables and target variables, it also includes: calculating error evaluation indicators to evaluate the prediction performance of the typhoon prediction model, and the error evaluation indicators include one or more of mean square error, root mean square error, evaluation absolute error, determination coefficient, minimum distance error, and maximum speed error.

[0030] Furthermore, after comparing the predicted results of the target variable with the corresponding actual observed data, it also includes:

[0031] If the accuracy of the prediction result of the target variable does not meet the requirements, adjust the parameters of the typhoon prediction model and use the long time series data of the characteristic variables and the target variable to train the typhoon prediction model again, input the long time series data of the characteristic variables and the target variable into the retrained typhoon prediction model, obtain a new prediction result of the target variable, compare the prediction result of the target variable with the corresponding actual observation data, and obtain the corresponding accuracy;

[0032] The changing trend of the accuracy of the prediction results of the target variable is analyzed. If the changing trend does not meet the requirements, the typhoon prediction model is trained again using the long time series data of the characteristic variables and the target variable containing the real-time observation data.

[0033] The present invention also proposes a data-driven nearshore wind and wave prediction system, comprising a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute any one of the data-driven nearshore wind and wave prediction methods.

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] The present invention introduces the analysis results of time lag effects and spatial propagation patterns as prior knowledge into the deep learning model, which can not only learn the surface features in the data, but also deeply understand the inherent mechanisms and interaction relationships between variables. Especially when processing grid data, the model can go beyond the traditional point-to-point analysis perspective and analyze the cross-influence of variables in space and the propagation patterns between regions from the perspective of spatial integrity, thereby achieving comprehensive understanding and accurate prediction of the complex marine environment.

[0036] The present invention uses the dynamic time warping algorithm (DTW algorithm) and the cross-correlation algorithm (CCF algorithm) to deeply analyze the complex time lag effects and dynamic spatial propagation patterns of variables between offshore and nearshore grids. The DTW algorithm effectively captures the nonlinear relationship between different time series caused by the time lag effect by calculating the optimal matching path between sequences, providing an accurate time alignment basis for model training; while the CCF algorithm reveals the intensity and directionality of variable propagation in the spatial dimension by quantifying the cross-correlation between variables, providing an important basis for model construction of a spatial feature extraction module. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the overall concept of an embodiment of the present invention.

[0038] Figure 2 The figure is a flow chart of a method according to an embodiment of the present invention.

[0039] Figure 3 The figure is a flow chart of multi-source data fusion and data preprocessing in an embodiment of the present invention.

[0040] Figure 4 Schematic diagram of a typhoon prediction model framework according to an embodiment of the present invention.

[0041] Figure 5 This is the loss curve for training the nearshore significant wave height prediction model.

[0042] Figure 6 Schematic diagram of the fitting effect of the nearshore grid significant wave height prediction model.

[0043] Figure 7 This is the prediction result diagram of the nearshore grid significant wave height prediction model test set.

[0044] Figure 8 This is the prediction result of effective wave height in the next 24 hours for nearshore grid points.

[0045] Fig. 9 This is a comparison chart between the predicted and actual values ​​of the significant wave height at the nearshore grid point. DETAILED DESCRIPTION

[0046] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0047] Embodiment 1

[0048] In order to improve the precision and accuracy of typhoon prediction, this embodiment proposes a data-driven nearshore wind and wave prediction method. Figure 1 As shown, including:

[0049] Multi-source data collection and preprocessing: collect satellite remote sensing data, buoy observation data, reanalysis data and numerical simulation data for data preprocessing;

[0050] Feature engineering and extraction: Use preprocessed multi-source data to build a comprehensive feature set of variables, and extract feature variables from the comprehensive feature set;

[0051] Model selection and training: Identify the time lag effect and spatial propagation pattern between offshore characteristic variables and nearshore information, then integrate their algorithms into the deep learning model to further analyze the long-term dependencies and spatial relationships of the time series, and output the inland or nearshore information at future moments;

[0052] Real-time prediction and feedback mechanism: Build a real-time data processing channel and a loop feedback mechanism to pre-process, extract features, and make real-time predictions on data from newly arrived sites;

[0053] Dynamic model optimization and retraining: Dynamically adjust model parameters according to typhoon forecast results to improve forecast accuracy and real-time performance;

[0054] Real-time prediction results after optimization: Outputs the prediction results of the optimized and retrained model.

[0055] Based on the above ideas, the specific implementation steps of the method in this embodiment are as follows: Figure 2 As shown, the following steps are included:

[0056] S101) obtaining multi-source data of long-term series data of various meteorological data variables in the study area and preprocessing them, selecting the target variable and finding the characteristic variable significantly correlated with the target variable, and dividing the study area into regional grid points according to longitude and latitude;

[0057] S102) taking the characteristic variables of the outer sea grid points as the reference sequence, taking the target variables of the inner sea or nearshore grid points as the target sequence, and analyzing the time lag effect and spatial propagation pattern between the reference sequence and the target sequence;

[0058] S103) introducing the analysis results of the time lag effect and the spatial propagation mode into the deep learning model as prior knowledge, and using the long time series data of the characteristic variables and the target variables to train the deep learning model to obtain a typhoon prediction model;

[0059] S104) Input the long time series data of the characteristic variables and the target variables into the typhoon prediction model to obtain the prediction result of the target variable, obtain the real-time observation data of each meteorological data variable in the study area and preprocess it, compare the prediction result of the target variable with the corresponding actual observation data, and output the prediction result of the target variable if the accuracy of the prediction result of the target variable meets the requirements.

[0060] Each step is described in detail below.

[0061] In step S101 of this embodiment, the meteorological data variables in the study area include effective wave height, average wave period, average wavelength, average sea level pressure, 10-meter U wind component, 10-meter V wind component, sea surface temperature, etc. The multi-source data include satellite remote sensing data (such as SAR, MODIS, etc.), buoy observation data, meteorological reanalysis data (such as ECMWF, ERA5, etc.) and high-precision numerical simulation data of these meteorological data variables. By collecting satellite remote sensing data, buoy observation data, reanalysis data and numerical simulation data and integrating them, richer data dimensions and variables are provided, the depth and breadth of data mining are deepened, the selection of temporal resolution and spatial resolution is optimized, and the cross-validation and complementarity of multi-source data reduce the errors caused by a single data source, thereby improving the comprehensiveness and accuracy of the data set.

[0062] In step S101, when obtaining multi-source data of long time series data of each variable in the study area and preprocessing, such as Figure 3 As shown: The preprocessing steps include:

[0063] S201) complementing and verifying the multi-source data to obtain a comprehensive feature set, specifically including:

[0064] Use multi-source data for spatial complementarity, use the wide coverage of remote sensing data to fill in the data gaps in space, and combine the high-resolution characteristics of numerical simulation data to perform spatial interpolation and refinement of data;

[0065] Use multi-source data for temporal complementation. Buoy observation data usually have high temporal resolution but limited coverage. Satellite remote sensing data have wide coverage but low temporal resolution. By combining the two, more comprehensive time series data can be obtained. Meteorological reanalysis data provides continuous observations of long time series, which can be used to fill the gaps in short-term observation data.

[0066] Utilize multi-source data to complement variables. Different data sources contain different variable information. For example, satellite remote sensing data provides sea surface temperature, wave parameters, etc., while buoy observation data contains more detailed ocean current field information. By integrating these variables, a more comprehensive set of marine environmental characteristics can be constructed.

[0067] Use multi-source data for cross-validation, check data consistency, check whether the values ​​of variables at the same time and location are consistent between different data sources, correct errors and deviations in the data, verify data accuracy, and improve the quality of the data set;

[0068] S202) The missing values ​​in the feature set are filled by interpolation method, and the Z-score is used to identify and process abnormal data points in the feature set. Finally, in order to avoid the difference in the dimension between the feature variable and the target variable affecting the prediction accuracy and improve the model training efficiency and stability, the deviation standardization is used to normalize the data in the feature set. The expression is as follows:

[0069]

[0070] Among them, x i is the original value of the variable, x j is the normalized value of the variable, x max and x min are the maximum and minimum values ​​of the variables respectively;

[0071] S203) performing feature analysis on the feature set, including time series features such as trend features, period features and statistical features, and physical features such as spatial distribution features, dynamics and interaction features.

[0072] In step S101, when selecting a target variable and searching for characteristic variables significantly correlated with the target variable, feature selection and optimization are performed to reduce redundant features, improve model efficiency and prediction accuracy, and through correlation coefficient analysis, quantitative indicators analyze the strength and direction of the linear relationship between variables, identify characteristic variables significantly associated with the target variable, and perform regression analysis to verify the correlation coefficient analysis results, and further screen high-correlation characteristic variables, including the following steps:

[0073] S301) respectively calculating the Pearson correlation coefficient between each meteorological data variable and the target variable, and selecting the meteorological data variable whose absolute value of the Pearson correlation coefficient with the target variable is greater than a preset threshold as a candidate variable;

[0074] In this embodiment, the effective wave height is used as the target variable, and the Pearson correlation coefficient is used to accurately quantify the strength and direction of the linear relationship between these characteristic variables and the effective wave height in step S203. Its value range is between -1 and 1, which is convenient for interpreting the overall picture from complete negative correlation to complete positive correlation between variables. A positive value indicates a positive correlation, that is, an increase in one variable is accompanied by an increase in another variable; a negative value indicates a negative correlation, that is, an increase in one variable is accompanied by a decrease in another variable; and 0 means that there is no linear relationship between the two:

[0075]

[0076] In the above formula: cov is the covariance, σ is the standard deviation, X represents the feature variable, and Y represents the target variable.

[0077] Based on the calculation results of the Pearson correlation coefficient, the characteristic variables (independent variables X) that are significantly correlated with the effective wave height (assuming the target variable Y) are identified. 1 ,X 2 ,…,X n ), the significance judgment is based on ρ X,Y and the size of the correlation coefficient (close to 1 or -1 indicates a strong correlation);

[0078] Next, univariate and multivariate regression analysis methods are used to screen highly correlated characteristic variables. Univariate regression analysis focuses on the impact of a single independent variable on the target variable, while multivariate regression analysis can comprehensively consider multiple independent variables and reveal their joint effects on the target variable, including:

[0079] S302) establishing a corresponding univariate regression model for each candidate variable, and evaluating the degree of fit and variable significance of each univariate regression model to screen highly correlated variables that have a significant positive or negative impact on the target variable;

[0080] In this embodiment, a univariate regression analysis is performed for each characteristic variable significantly correlated with the significant wave height to evaluate the influence of the single variable on the significant wave height;

[0081] For each feature variable X i , establish a univariate regression model Y = β 0 +β 1 X i +ε, where Y is the effective wave height, X i is the characteristic variable, β 0 is the intercept term, β 1 is the slope coefficient, ε is the error term, and β is estimated using the least squares method. 0 , β 1 The value of , the determination coefficient and residual analysis are used to evaluate the model fit and variable significance, and X with significant positive or negative impact on the target variable is selected. i As a highly correlated variable;

[0082] S303) establishing a corresponding multiple regression model for each highly correlated variable, and comprehensively evaluating the performance of each multiple regression model to screen highly correlated characteristic variables;

[0083] In this embodiment, after determining the highly correlated variables in the univariate regression analysis, a multivariate regression analysis is further performed, and all the selected highly correlated variables are used as independent variables to establish a multivariate regression model Y = β 0 +β 1 X 1 +β 2 X 2 +…+β n X n+ε, the values ​​of all β coefficients are estimated using the least squares method, and the model performance is comprehensively evaluated through the determination coefficient, residual analysis and VIF (variance inflation factor), to check whether there are multicollinearity problems between the independent variables and to determine the highly correlated characteristic variables.

[0084] Through the above steps, multiple characteristic variables are analyzed, including u10 (the latitudinal component of the ten-meter wind speed), v10 (the longitudinal component of the ten-meter wind speed), mean wave direction, mean wave period, mean sea level pressure and sea surface temperature, etc. These variables together constitute a multidimensional framework for analyzing ocean dynamic processes and their impact on significant wave height.

[0085] In step S101 of this embodiment, the study area is divided into regional grid points according to longitude and latitude. Specifically, the entire study area is divided into a regular grid consisting of Q spatial grid points according to longitude and latitude. Each node contains the initial time Q i0 To the current time Q it By orthogonally dividing the entire study area into countless small grids, a refined spatial expression of the study area is achieved, so as to capture more subtle spatial change characteristics; the division of regular grids enables geographic data from different sources and resolutions to be unified into the same spatial framework, and also facilitates data integration and comparative analysis, so that data can be standardized in space, and data between different grids are comparable, providing a basis for cross-regional comparative studies.

[0086] In step S102 of this embodiment, the characteristic variables of the open sea grid points and the target variables of the inland sea or nearshore grid points are the variable data matrices corresponding to the longitude and latitude of each grid. For example, the matrix H represents the effective wave height, the matrix T represents the average wave period, the matrix U represents the 10-meter U wind component, and the matrix V represents the 10-meter V wind component V, which includes Q spatial grid points, each of which contains n hours of data:

[0087]

[0088] Step S102 of this embodiment extracts and analyzes time, space and physical characteristics based on multi-source and multi-dimensional variable data, including wave significant wave height, average wave period, 10-meter U wind component, 10-meter V wind component, sea surface temperature and other variable data. While paying attention to the immediate changes in the nearshore waters, it deeply identifies the lag effect and spatial propagation relationship of the offshore characteristic variables on the nearshore waters, realizes cross-regional time lag effect analysis, analyzes the interaction mechanism between different marine meteorological elements, and analyzes the time lag effect and spatial propagation mode between the reference sequence and the target sequence. The offshore grid point Q is selected. i The long-term characteristic variable data is used as the reference series x(t), the inland or nearshore Q jThe target variable is taken as the target sequence y(t), and the cross-correlation function is used to preliminarily capture the time lag effect between time series. Through spatial autocorrelation analysis and spatial regression analysis, the influence of spatial position on the variable relationship is quantified, and the spatial propagation pattern is analyzed, including:

[0089] The outer sea grid point S i The first characteristic variable among the characteristic variables (in this embodiment, U is selected i 、V i Variable) as the reference sequence x(t), the effective wave height H at the nearshore grid point j As the target sequence y(t), the cross-correlation algorithm (CCF) is used to analyze the time lag effect between the reference sequence and the target sequence. The expression is as follows:

[0090]

[0091] Among them, R xy (τ) is the cross-correlation function between the reference sequence and the target sequence, τ represents the time lag, x(t) represents the reference sequence, and y(t+τ) represents the target sequence;

[0092] In the cross-correlation algorithm (CCF) analysis of offshore variables (U i 、V i ) and nearshore significant wave height (H j ) to further refine the analysis and consider more offshore factors that may affect the nearshore wave height, an additional offshore variable, wave height (H i ) and period (T i ), the addition of new variables can provide a more comprehensive description of the marine environment, which may reveal more complex nonlinear time-lag effects and spatial propagation patterns;

[0093] Next, the first characteristic variable (U is selected in this embodiment) among the characteristic variables of the outer sea grid point is i 、Vi i variable) and the second characteristic variable (in this embodiment, H i , T i Variable) as the reference sequence x(t), the effective wave height H at the nearshore grid point j As the target sequence y(t), the dynamic time warping (DTW) algorithm is used to calculate the optimal matching path between the reference sequence and the target sequence to analyze the nonlinear time lag effect and spatial propagation pattern between the reference sequence and the target sequence.

[0094] When constructing the dynamic time warping (DTW) distance matrix, we are no longer limited to a single or a few variables, but instead consider multiple variables as a multidimensional vector. For each time point t, the state of the open sea can be represented by a matrix containing H i (t),T i(t),U i (t),V i The vector representation of (t) is denoted as x(t), which is the nearshore significant wave height H j (t) is considered as a vector containing only one element, or as a single time series corresponding to a multidimensional open sea vector;

[0095] For two multidimensional time series x(t) and y(t), the DTW distance is defined as a cost function that minimizes the cumulative distance by warping the time axis. The cumulative distance is usually defined as:

[0096] D(x,y)=min φ (∑ (i,j)∈φ d(x i ,y j ))

[0097] Where φ is a set of time index pairs (i, j), representing a possible matching path between the reference sequence x and the target sequence y, and d(x i ,y j ) is the i-th element x in the reference sequence i and the jth element y in the target sequence j The distance measure between

[0098] For multidimensional vectors, use the Euclidean distance as d(x i ,y j ) definition:

[0099]

[0100] Where n is the dimension of the vector, and They are x i and j The kth component of ;

[0101] The DTW algorithm finds the optimal matching path by filling a two-dimensional matrix D(i,j), where D(i,j) is the minimum cumulative distance from x(1) to x(i) and from y(1) to y(j). The initial conditions are usually D(0,j) = ∞ (for all j>0) and D(i,0) = ∞ (for all i>0), and D[0,0] = 0. This matrix can be calculated by the following recursive relationship:

[0102] D(i,j)=d(x i ,y j )+min(D(i-1,j),D(i,j-1),D(i-1,j-1))

[0103] Where D(i,j) represents the minimum cumulative distance between the first point to the i-th point of the reference sequence x and the first point to the j-th point of the target sequence y; d(x i ,y j ) is the Euclidean distance between the i-th element in the reference sequence x and the j-th element in the target sequence y; D(i-1,j) represents the minimum accumulated distance when the current point of the reference sequence x moves back one step while the current point of the target sequence y remains unchanged in the matching path; D(i,j-1) represents the minimum accumulated distance when the current point of the reference sequence x remains unchanged while the current point of the target sequence y moves back one step in the matching path; D(i-1,j-1) represents the minimum accumulated distance when the current points of both the reference sequence x and the target sequence y move back one step in the matching path.

[0104] After the DTW distance matrix is ​​filled, the optimal matching path from D[1,1] to D[m,n] (where m and n are the lengths of x and y, respectively) can be found by backtracking. This path reveals the nonlinear time correspondence between x and y, so that the time lag effect and spatial propagation pattern can be analyzed.

[0105] The DTW algorithm effectively captures the nonlinear relationship between different time series caused by time lag effects by calculating the optimal matching path between sequences, providing an accurate basis for time alignment for model training; while the CCF algorithm quantifies the cross-correlation between variables, revealing the intensity and directionality of variable propagation in the spatial dimension, providing an important basis for model construction of spatial feature extraction modules.

[0106] Step S103 of this embodiment integrates the time lag effect and spatial propagation relationship analysis algorithm with the deep learning model, and deeply mines the complex patterns and laws of the data through big data driving, uses the long time series data set of each feature variable and the target variable as input for model training, further extracts the time domain features and spatial features, mines the long-term dependency and spatial propagation pattern, and obtains the predicted value of the target variable at the future moment, including but not limited to the effective wave height, wind speed, etc.

[0107] The deep learning model of this embodiment adopts the TCN-CNN model. The TCN-CNN model uses a time series model (TCN) to capture long-term dependencies. The core advantages of TCN are its parallel computing capability, the advantages of long-term dependency modeling, and efficient feature extraction capabilities.

[0108] Using causal convolution makes each output time step depend only on the current and previous time step input data, but not on future time steps, ensuring that the model maintains causality when processing time series data and avoiding deviations caused by future information leakage;

[0109] Dilated convolution applies convolution kernels at intervals so that it can increase the receptive field without losing temporal resolution, allowing it to capture information at a longer distance;

[0110] Adding residual connections alleviates the gradient vanishing problem of deep networks and promotes cross-layer information transmission. Each residual block contains two convolutional layers and one residual connection.

[0111] Weight normalization is added to improve the training efficiency and stability of the model. By scaling the data to a reasonable range, the neural network weight update process is smoother, reducing the sensitivity to the learning rate and preventing overfitting.

[0112] Combined with the spatial model (CNN) to process spatial feature extraction, the convolution layer extracts local features of the input data by means of a sliding window. The number of input channels corresponds to the number of feature variables. The number of convolution layers is selected according to the complexity of the data to extract features at different levels. After the convolution layer, the pooling layer is averaged and down-sampled to reduce the data dimension and the amount of calculation, and retain important features.

[0113] Combined with the time series features extracted by the TCN part, it can effectively handle feature extraction and long-term dependency capture to simultaneously capture temporal dependency and spatial correlation, achieving efficient and accurate real-time prediction.

[0114] like Figure 4 As shown in the figure, based on the analysis of time lag effect and spatial propagation mode, by deeply integrating the advantages of time series analysis (TCN) and spatial feature extraction (CNN), and incorporating prior knowledge of nonlinear time lag effect and spatial propagation mode, high-precision prediction of nearshore significant wave height is achieved.

[0115] The TCN module in the TCN-CNN model can accurately capture the dependencies of time series, and effectively capture the long-term dependencies and nonlinear time lag effects in the time series through one-dimensional dilated convolution. The dilated convolution enables the model to expand the receptive field without sacrificing temporal resolution, thereby capturing the mutual influence between more distant time points. At the same time, the application of causal convolution ensures that each time step of the model output depends only on the current and previous time step inputs, strictly maintaining the causality of the time series data and avoiding the risk of future information leakage.

[0116] The CNN module in the TCN-CNN model deepens the extraction of spatial features. In order to make full use of the spatial dimension information in the input data, the model introduces the CNN module, which flexibly adjusts the number of convolution layers according to the complexity of the data to extract spatial structural features at different levels. After the convolution layer, downsampling is performed through the average pooling layer, which effectively reduces the data dimension and the amount of calculation while retaining the key features. These spatial features are combined with the temporal features extracted by the TCN module to provide a more comprehensive information basis for the model.

[0117] The TCN-CNN fusion model achieves deep fusion of temporal features and spatial features through fusion strategies. In the preprocessing stage, the DTW and CCF algorithms are used to analyze the time lag effect and spatial propagation mode. This integrated method may not have been widely used in the field of ocean wave height prediction, providing the model with more accurate input data alignment and feature extraction;

[0118] In the model design stage, the analysis results of nonlinear time lag effects and spatial propagation patterns are introduced as prior knowledge. The application of the DTW algorithm in the preprocessing stage provides the model with information about the optimal matching path between sequences. This information is implicitly used in the model training process, which helps the model better understand the dynamic similarities between sequence data during the training stage. Using this information as part of the input and as a regularization term in the model training process can optimize the model's processing of time series data.

[0119] In order to introduce the time lag effect, the input data is expanded from the features of a single time point to a sequence containing multiple time points through a sliding window, and multiple related feature variables are provided to the model as multi-dimensional input. The expanded model analyzes the data dimension, which helps the model analyze the data more comprehensively and improves the modeling ability of complex dynamic relationships.

[0120] Further, the number of TCN and CNN convolutional layers and the size of convolutional kernels are dynamically adjusted according to the time lag effect to make the model more suitable for the data characteristics, and a fully connected layer and an attention layer are added after the TCN and CNN layers to fuse the temporal and spatial features. In the training process, the results of the DTW and CCF algorithms are used for pre-training to find the optimal solution faster and maintain high performance during training, so as to accelerate the training process and improve the model performance, improve the training efficiency and enhance the generalization ability of the model.

[0121] By adjusting the dilation rate of TCN and the convolution kernel configuration of CNN, the model can better adapt to these specific data characteristics; when the analysis results of the time lag effect and the spatial propagation pattern are introduced into the deep learning model as prior knowledge, the TCN module of the TCN-CNN model effectively captures the long-term dependency and nonlinear time lag effect in the time series through one-dimensional dilated convolution, and the CNN module of the TCN-CNN model uses the information of the spatial propagation pattern to realize spatial feature extraction;

[0122] At the same time, regularization is constructed to reflect prior knowledge, and regularization terms are constructed to punish the part of the model prediction results that does not conform to the law of time lag effect. For example, if prior knowledge indicates that the time lag effect in a time series data should be expressed in a certain specific pattern (such as exponential decay, linear growth, etc.), then the constraint of this pattern can be introduced in the regularization term so that the model can meet this constraint during the training process; the result of the DTW algorithm is converted into a part of the regularization term, and the learning process of the model is constrained by measuring the path similarity between the predicted sequence and the actual sequence, so that it pays more attention to the dynamic similarity between the sequences;

[0123] The improved model will focus more on identifying nonlinear time-lag effects and spatial propagation patterns through prior knowledge, which enables the model to use these key information more efficiently during training and prediction, and is more targeted and adaptive. Through adaptive improvements, the model can not only better fit the training data, but also better handle new data with similar nonlinear time-lag effects and spatial propagation patterns, enhancing its generalization ability on new data. Because the model structure is more targeted, it reduces unnecessary calculations, improves the efficiency of training and prediction, and improves the model's adaptability and problem-solving ability for specific problems. Through these structural and strategic adjustments, the model can more effectively handle data with nonlinear time-lag effects and spatial propagation characteristics, thereby providing more accurate prediction results in specific application scenarios.

[0124] The residual connection mechanism is incorporated into the model design of this embodiment, which directly connects the output of the shallow CNN to the output of the deep CNN. This not only deepens the network depth, but also significantly improves the performance stability of the deep network, effectively avoiding the problems of gradient vanishing and performance degradation during training.

[0125] To further enhance the generalization ability of the model and prevent overfitting, the TCN-CNN model embeds dropout layers at key positions, especially before the linear layer after the CNN module. The dropout rate is set to randomly discard the outputs of some neurons, forcing the network to learn more robust feature representations.

[0126] During the model training phase, the model receives variable long time series data (U i 、Vi , H i , T i The data are organized into a time series matrix and contain information about the spatial dimension. The analysis results of nonlinear time lag effects and spatial propagation patterns are introduced into the deep learning model as prior knowledge. The stochastic gradient descent optimizer (SGD) is used, supplemented by a dynamic learning rate adjustment strategy, such as Figure 5 As shown in the figure, when the loss on the validation set has not improved for five consecutive rounds, the learning rate is automatically halved to promote the model to converge to a better solution, and the weight decay parameter of the optimizer is set to effectively control the size of the model weight, avoid the excessive influence of individual parameters on the overall prediction results, and further improve the prediction accuracy and stability of the model. By adjusting the time window and the prediction step size for future moments, the model outputs the prediction results of the nearshore significant wave height for the next 24 moments.

[0127] In this embodiment, after the deep learning model is trained using the long time series data of the characteristic variables and the target variables, a systematic error analysis framework is constructed to comprehensively reflect the different performances of the model. Specifically, according to the specific needs and goals of typhoon prediction, representative and complementary error evaluation indicators are selected, and the error evaluation indicators are calculated to evaluate the prediction performance of the typhoon prediction model. The error evaluation indicators include but are not limited to: mean square error (MSE), root mean square error (RMSE), absolute error (MAE), determination coefficient (R 2 ) etc.; according to the characteristics of typhoon prediction, specific evaluation indicators such as minimum distance error (MDE) and maximum velocity error (MSE_v) are introduced to comprehensively evaluate the prediction performance of the model, compare and quantify the performance of the model in different aspects. The relevant calculation formulas are as follows:

[0128]

[0129] Where: n is the total number of samples in the test set, o i is the predicted value, u i It is the actual observed value. MAE is insensitive to outliers and reflects the average deviation of the prediction.

[0130]

[0131] Where: n is the total number of samples in the test set, o i is the predicted value, u i is the actual observed value. MSE is sensitive to outliers and is used to evaluate the overall error level.

[0132]

[0133] Where: n is the total number of samples in the test set, o i is the predicted value, u iIt is the actual observed value. RMSE has the same dimension as the true value, which makes it easier to intuitively understand the error size.

[0134]

[0135] Where: n is the total number of samples in the test set, o i is the predicted value, u i is the actual observed value, is the average of the observed values, and the coefficient of determination R 2 Measures the degree of fit between the model prediction value and the actual observation value. The closer the value is to 1, the better the model fit is.

[0136] like Figure 4 As shown, this embodiment establishes a real-time feedback mechanism and a real-time data processing channel, pre-processes and extracts features of the newly arrived measuring point data, inputs the extracted feature data into the prediction model in real time, compares it with the prediction result, and uses the comparison result as feedback to continuously input it into the model optimization and verification process, forming a closed-loop optimization mechanism, using the mean square error, mean absolute error and other indicators to evaluate the prediction accuracy, dynamically adjusts the parameters of the prediction model through an automated algorithm according to the feedback results, so as to quickly find the optimal parameter combination, retrain the model and optimize the parameter adjustment, optimize the real-time prediction result, and output the optimized effective wave height real-time prediction result; and regularly evaluates the model performance, pays attention to the changing trend of the prediction accuracy, introduces an online learning mechanism, continuously learns new observation data, automatically updates the knowledge base, and improves the prediction ability of future typhoon events. The real-time feedback mechanism is established, and by establishing an efficient data acquisition system, the typhoon wind speed, wind direction, air pressure and other related data of each measuring point are collected in real time, and the collected data are further pre-processed and feature extracted and input into the prediction model, and the online learning mechanism is introduced to enable it to dynamically adjust the parameters in real time according to the newly arrived data, so as to improve the model's adaptability to new data and the prediction accuracy. Correspondingly, in step S104, after comparing the prediction result of the target variable with the corresponding actual observation data, the following steps are also included:

[0137] If the accuracy of the prediction result of the target variable does not meet the requirements, adjust the parameters of the typhoon prediction model and use the long time series data of the characteristic variables and the target variables to train the typhoon prediction model again, input the long time series data of the characteristic variables and the target variables into the retrained typhoon prediction model to obtain a new prediction result of the target variable, compare the prediction result of the target variable with the corresponding actual observation data to obtain the corresponding accuracy, and judge whether the progress meets the requirements again. If the requirements are met, the new prediction result of the target variable is output. If the requirements are not met, this step is iteratively executed;

[0138] The accuracy of the prediction results of the target variable each time in the specified time interval is obtained, and the changing trend of the accuracy of the prediction results of the target variable is analyzed. If the changing trend does not meet the requirements (such as the accuracy continues to decrease), the long time series data of the characteristic variables and the target variables containing real-time observation data are used to train the typhoon prediction model again.

[0139] The comparison between the prediction results output by the typhoon prediction model that is continuously optimized according to the above steps and the actual observation results is shown in the figure below. Figures 6 to 9 As shown, it can be seen Figure 6 and Figure 7 The model fitting effect and the prediction effect on the test set are presented respectively, both showing that the model is highly fitted to the historical data. Figure 6 In , the model’s accurate fit to historical data verifies the rationality of its internal logic and shows its strong potential in processing similar data. In addition, over time, the predicted values ​​and true values ​​have similar changing trends, which shows the model’s ability to capture data changing trends. Trend consistency is particularly important for scenarios that require predicting future data changes. Figure 7 It can be seen from the prediction results of the test set that the difference between the predicted value and the true value is small, and the model prediction value remains relatively stable over the entire time range, with trend consistency, indicating that the model has good robustness and generalization ability, ensuring that the model can still achieve stable and accurate predictions in complex environments; Figure 8 This is the forecast result of the effective wave height in the nearshore grid for the next 24 hours. The forecast value changes smoothly over time and maintains a similar trend to the true value. Fig. 9 By comparing the predicted value of the effective wave height in the next 24 hours with the true value, the prediction accuracy of the model is quantitatively demonstrated. The results show that the predicted value shows a similar trend to the true value over time, without large fluctuations and abnormal values. The prediction is relatively stable. The evaluation indicators mean square error, root mean square error, evaluation absolute error and determination coefficient are 0.01174, 0.10833, 0.0818 and 0.8973 respectively, and the maximum error percentage is 12%, indicating that the model prediction has high accuracy. The model shows excellent performance, verifies the rationality of its internal logic by highly fitting historical data, and demonstrates its strong potential and accuracy in processing similar marine environment data; the model not only has good trend capture ability and can accurately track data changes, but also shows good robustness and generalization ability, ensuring stable and accurate prediction in complex environments. The predicted value changes smoothly over time, maintains a similar trend to the true value, and has a low error, which fully proves the high accuracy of the model in wave height prediction and provides reliable technical support for marine engineering, meteorological forecasting, shipping safety and other fields.

[0140] Embodiment 2

[0141] This embodiment proposes a data-driven nearshore wind and wave prediction system, including a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute the data-driven nearshore wind and wave prediction method described in the first embodiment.

[0142] In summary, the present invention proposes a data-driven nearshore wind and wave prediction method and system, which integrates multi-source variable data sets of different time resolutions and spatial resolutions as comprehensive feature data sets, and performs high-resolution spatial division of the research area through refined grid division, and then deeply analyzes the complex time lag effect and dynamic spatial propagation mode of variables between offshore and nearshore grids. On this basis, a multi-level and multi-dimensional variable relationship analysis framework is constructed to analyze the interaction relationship between grid points, as well as the variable coupling mechanism between grid points and their respective regions, and between different regions, which are introduced into the deep learning prediction model as prior knowledge, and the self-feedback loop and online learning mechanism are integrated to realize the instant response and adaptive adjustment of the prediction model to changes in external meteorological data. When the offshore monitoring station captures the latest meteorological data, efficient data analysis and prediction are carried out, thereby realizing real-time and accurate prediction of typhoons and related marine environmental parameters in the nearshore waters.

[0143] Different from the traditional single analysis perspective of point-to-point variable propagation, this invention analyzes the cross-influence of variables in space and the propagation mode between regions from the perspective of spatial integrity. Through the integration of innovative data analysis methods and deep learning, it realizes high-precision modeling and prediction of complex time lag effects and spatial propagation modes in the marine environment. It has the following advantages:

[0144] Improve prediction accuracy: Through spatiotemporal correlation analysis and multi-source data fusion, multiple factors affecting the target variable are comprehensively considered to improve the accuracy and reliability of the prediction;

[0145] Realize real-time prediction: Build a real-time data processing and prediction system that can make real-time predictions of inland and coastal sea information under extreme weather conditions such as typhoons, and provide timely and accurate decision-making support for relevant departments and the public;

[0146] Flexibility and scalability: The design of the hybrid prediction model enables the system to be flexibly adjusted and optimized according to different prediction needs and data sources. It has good scalability. The model introduces an online learning mechanism to continuously learn new observation data, automatically update the knowledge base, and improve the ability to predict future typhoon events. It builds an efficient and accurate real-time data processing channel and prediction model to provide strong support for disaster prevention and mitigation.

[0147] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A data-driven nearshore wind and wave prediction method, characterized in that: The following steps are involved: Obtain multi-source data of long-term series data of various meteorological data variables in the study area and preprocess them, select the target variable from the meteorological data variables and find the characteristic variables significantly correlated with the target variable from other meteorological data variables other than the target variable, divide the study area into regional grid points according to longitude and latitude, and the meteorological data variables in the study area include effective wave height, average wave period, average wavelength, average sea level pressure, 10-meter U wind component, 10-meter V wind component, and sea surface temperature; The characteristic variables of the offshore grid points are used as the reference sequence, and the target variables of the inland or nearshore grid points are used as the target sequence. The time lag effect and spatial propagation pattern between the reference sequence and the target sequence are analyzed, including: The first characteristic variable among the characteristic variables of the outer sea grid points is used as the reference sequence. The first characteristic variable includes the 10-meter U wind component and the 10-meter V wind component. The cross-correlation algorithm is used to analyze the time lag effect between the reference sequence and the target sequence. The first and second characteristic variables of the characteristic variables of the outer sea grid points are taken as the reference sequence. The second characteristic variable includes the significant wave height and the average wave period. The dynamic time warping algorithm is used to calculate the optimal matching path between the reference sequence and the target sequence to analyze the nonlinear time lag effect and spatial propagation mode between the reference sequence and the target sequence. The analysis results of time lag effect and spatial propagation mode are introduced into the deep learning model as prior knowledge, and the deep learning model is trained using the long time series data of characteristic variables and target variables to obtain the typhoon prediction model. The long time series data of characteristic variables and target variables are input into the typhoon prediction model to obtain the prediction results of the target variables. The real-time observation data of each meteorological data variable in the study area are obtained and preprocessed. The prediction results of the target variables are compared with the corresponding actual observation data. If the accuracy of the prediction results of the target variables meets the requirements, the prediction results of the target variables are output.

2. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: When obtaining multi-source data of long time series data of various variables in the study area and preprocessing them, the preprocessing steps include: The multi-source data are complemented and verified to integrate to obtain the feature set, and then the missing values ​​in the feature set are filled, while the abnormal data points in the feature set are identified and processed, and finally the data in the feature set is normalized.

3. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: When selecting a target variable and finding characteristic variables that are significantly correlated with the target variable, include: Calculate the Pearson correlation coefficient between each meteorological data variable and the target variable respectively, and select the meteorological data variable whose absolute value of the Pearson correlation coefficient with the target variable is greater than a preset threshold as the candidate variable; A corresponding univariate regression model is established for each candidate variable, and the degree of fit and variable significance of each univariate regression model are evaluated to screen highly correlated variables that have significant positive or negative effects on the target variable; A corresponding multivariate regression model was established for each highly correlated variable, and the performance of each multivariate regression model was comprehensively evaluated to screen the characteristic variables with high correlation.

4. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: When the cross-correlation algorithm is used to analyze the time lag effect between the reference sequence and the target sequence, the expression is as follows: Among them, R xy (τ) is the cross-correlation function between the reference sequence and the target sequence, τ represents the time lag, x(t) represents the reference sequence, and y(t+τ) represents the target sequence.

5. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: When the dynamic time warping algorithm is used to calculate the optimal matching path between the reference sequence and the target sequence, the expression is as follows: D(i,j)=d(x i ,y j )+min(D(i-1,j),D(i,j-1),D(i-1,j-1)) Where D(i,j) represents the minimum cumulative distance between the first point to the i-th point of the reference sequence x and the first point to the j-th point of the target sequence y; d(x i ,y j ) is the Euclidean distance between the i-th element in the reference sequence x and the j-th element in the target sequence y; D(i-1,j) represents the minimum accumulated distance when the current point of the reference sequence x moves back one step while the current point of the target sequence y remains unchanged in the matching path; D(i,j-1) represents the minimum accumulated distance when the current point of the reference sequence x remains unchanged while the current point of the target sequence y moves back one step in the matching path; D(i-1,j-1) represents the minimum accumulated distance when the current points of both the reference sequence x and the target sequence y move back one step in the matching path.

6. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: The deep learning model adopts the TCN-CNN model, and the analysis results of the time lag effect and the spatial propagation mode are introduced into the deep learning model as prior knowledge, including: The sliding window is used to expand the input data from the features of a single time point to a sequence containing multiple time points, and multiple related feature variables are provided to the model as multi-dimensional inputs, expanding the model analysis data dimension; The number of TCN and CNN convolutional layers and the size of convolutional kernels are dynamically adjusted according to the time lag effect, and a fully connected layer and an attention layer are added after the TCN and CNN layers to fuse the temporal features and the spatial features, so that the TCN module of the TCN-CNN model effectively captures the long-term dependency and nonlinear time lag effect in the time series through one-dimensional dilated convolution, and the CNN module of the TCN-CNN model realizes spatial feature extraction by using the information of the spatial propagation mode; During the training process, the results of the cross-correlation algorithm and the dynamic time warping algorithm are used for pre-training to find the optimal solution more quickly. At the same time, a regularization term is constructed to punish the part of the model prediction results that does not conform to the law of the time lag effect. Specifically, the result of the dynamic time warping algorithm is converted into a part of the regularization term. By measuring the path similarity between the predicted sequence and the actual sequence, the learning process of the model is constrained, so that it pays more attention to the dynamic similarity between sequences.

7. The data-driven nearshore wind and wave prediction method according to claim 1, characterized in that: After training the deep learning model using long time series data of feature variables and target variables, it also includes: calculating error evaluation indicators to evaluate the prediction performance of the typhoon prediction model, and the error evaluation indicators include one or more of mean square error, root mean square error, evaluation absolute error, determination coefficient, minimum distance error, and maximum speed error.

8. The data-driven nearshore wind and wave prediction method according to claim 7, characterized in that: After comparing the predicted results of the target variable with the corresponding actual observed data, it also includes: If the accuracy of the prediction result of the target variable does not meet the requirements, adjust the parameters of the typhoon prediction model and use the long time series data of the characteristic variables and the target variable to train the typhoon prediction model again, input the long time series data of the characteristic variables and the target variable into the retrained typhoon prediction model, obtain a new prediction result of the target variable, compare the prediction result of the target variable with the corresponding actual observation data, and obtain the corresponding accuracy; The changing trend of the accuracy of the prediction results of the target variable is analyzed. If the changing trend does not meet the requirements, the typhoon prediction model is trained again using the long time series data of the characteristic variables and the target variable containing the real-time observation data.

9. A data-driven nearshore wind and wave prediction system, characterized in that: The invention comprises a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute the data-driven nearshore wind and wave prediction method according to any one of claims 1 to 8.

Images