Station storm water increase prediction method, equipment, medium and product

Abstract

The invention discloses a station storm water increase prediction method and device, a medium and a product. The method comprises the steps of obtaining n feature vectors and sending the n feature vectors to a TCN model trained in advance, so that the TCN model trained in advance outputs n storm surge preliminary predicted values after report starting time a, the n feature vectors are feature vectors corresponding to N forecast time periods of the target meteorological elements within a preset range with the target station as the center after the report starting time a; obtaining the report starting time a and the preorder storm water increasing values of the M time periods of the preorder of the report starting time a, and obtaining M + 1 preorder storm water increasing values; splicing the M + 1 preorder storm water increasing values and the n storm water increasing preliminary predicted values to obtain mixed features; and sending the mixed features to a pre-trained KAN model, so that the KAN model outputs n storm surge final predicted values after the report starting time a. According to the method, the station storm surge prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of storm surge prediction, in particular to a station storm surge prediction method, device, medium and product. BACKGROUND

[0002] China is one of the countries in the world that suffers the most serious storm surge disasters, and accurate prediction of storm surge is of great significance to improve the ability of disaster prevention and reduction and reduce personnel and property losses. In view of the rapid development of artificial intelligence technology and its wide application in various fields, intelligent prediction of storm surge is gradually becoming another important means to cope with storm surge prediction after numerical prediction, with its unique advantages. Station storm surge prediction is an indispensable part of the storm surge prediction system, which provides high-precision and refined prediction information, plays a key role in promoting the improvement of numerical prediction models, the development of intelligent prediction technology, and the support of emergency response and disaster management. At present, the intelligent prediction of station storm surge has developed from single variable time series prediction to multi-variable time series prediction, and further to multi-variable spatio-temporal sequence prediction. In this process, the prediction ability and accuracy of intelligent models are constantly improved. In addition, the input of meteorological numerical prediction data more effectively expresses the long-term change characteristics of storm surge. However, the change of storm surge is affected by nonlinear factors and multi-scale interaction, showing significant spatio-temporal complexity characteristics. The existing storm surge prediction methods based on BPNN (Backpropagation Neural Network), CNN (Convolutional Neural Networks), LSTM ((Long Short-Term Memory) and other neural networks and their combination models face the following challenges in the specific implementation process: first, the gradient stability is insufficient, it is difficult to effectively capture the long-distance time sequence dependence relationship, which is easy to cause gradient disappearance or gradient explosion problem, and then affect the convergence of model training; second, lack of strong function approximation ability, which may cause model underfitting or similar mean regression phenomenon. The above problems jointly restrict the further improvement of the prediction accuracy of storm surge, making it face a bottleneck.

[0003] In view of this, the application proposes a station storm surge intelligent prediction combination model based on a temporal convolutional network (TCN) and a Kolmogorov-Arnold network (KAN), and a corresponding instance is constructed. The model inputs multivariate meteorological background field spatiotemporal sequence numerical prediction data, first performs dimension reduction processing on the original input data by using a principal component analysis method (PCA) to reduce redundant information, second constructs a TCN model to capture long-distance dependence in the time sequence and perform meteorological feature extraction, and finally uses the powerful nonlinear function approximation capability of the KAN model to improve the accuracy of storm surge prediction. SUMMARY

[0004] The purpose of the application is to provide a station storm surge prediction method, device, medium and product to solve the problem of low accuracy of station storm surge prediction described in the background art.

[0005] To achieve the above purpose, the application provides the following solutions.

[0006] In a first aspect, the application provides a station storm surge prediction method, comprising:

[0007] n feature vectors are obtained and sent to a previously trained temporal convolutional network (TCN) model, so that the previously trained TCN model outputs n storm surge preliminary prediction values after a reporting time a, the n feature vectors being feature vectors corresponding to target meteorological elements in a predetermined range centered on a target station for N prediction times after the reporting time a;

[0008] M time sequence storm surge values before the reporting time a are obtained, to obtain M+1 time sequence storm surge values;

[0009] The M+1 time sequence storm surge values and the n storm surge preliminary prediction values are spliced to obtain a mixed feature;

[0010] The mixed feature is sent to a previously trained Kolmogorov-Arnold network (KAN) model, so that the KAN model outputs n storm surge final prediction values after the reporting time a.

[0011] Optionally, the target meteorological element includes any one of a sea level pressure, a first predetermined height zonal wind speed, and a second predetermined height meridional wind speed.

[0012] Optionally, the predetermined range includes a predetermined number of latitude and longitude grid regions centered on the target site.

[0013] Optionally, the n feature vectors are generated by the following method:

[0014] For the target element, meteorological forecast data for N forecast horizons after the reporting time a is obtained.

[0015] For the meteorological forecast data for the N forecast horizons, a sliding window with a window size W and a sliding step L is performed, and a feature vector is generated for each sliding, to obtain the n feature vectors.

[0016] Optionally, for the target element, before the meteorological forecast data for N forecast horizons after the reporting time a is obtained, the method further includes the following method:

[0017] The N forecast horizons of meteorological forecast data corresponding to the first predetermined number of latitude and longitude grid regions are sent to a principal component analysis (English full name: Principal Component Analysis, English abbreviation: PCA) model, so that the principal component analysis model performs dimensionality reduction processing on the N forecast horizons of meteorological forecast data, to obtain the N forecast horizons of meteorological forecast data of the second predetermined number of latitude and longitude grid regions after dimensionality reduction, wherein the second predetermined number is less than the first predetermined number.

[0018] Optionally, the first predetermined height is 10 meters, and the second predetermined height is 10 meters.

[0019] In a second aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method of any one of the first aspect.

[0020] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method of any one of the first aspect.

[0021] In a fourth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the method of any one of the first aspect.

[0022] According to the specific embodiments provided in the application, the application discloses the following technical effects:

[0023] The station storm surge prediction method provided in the embodiment of the application performs preliminary prediction of storm surge through the TCN model; then, the preliminary prediction of storm surge is spliced with the previous storm surge observation value to form a hybrid feature; then, the hybrid feature is sent to the KAN model, so that the KAN model is used to perform final prediction of storm surge, that is, the application constructs a TCN-KAN hybrid model, and the accuracy of storm surge prediction is improved. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0025] Figure 1 An application environment diagram of a station storm surge prediction method in an embodiment of the application;

[0026] Figure 2 A flowchart of a station storm surge prediction method provided in an embodiment of the application;

[0027] Figure 3 A flowchart of an n-feature vector generation method provided in an embodiment of the application;

[0028] Figure 4 A method diagram of n-feature vector generation provided in another embodiment of the application;

[0029] Figure 5 A TCN model construction and training diagram provided in an embodiment of the application;

[0030] Figure 6 A hybrid feature generation diagram provided in another embodiment of the application;

[0031] Figure 7 A KAN model construction and training diagram provided in an embodiment of the application;

[0032] Figure 8 A station storm surge prediction method flowchart provided in an embodiment of the application;

[0033] Figure 9 A comparison diagram between a storm surge prediction result of the application and a prediction result of another prediction method provided in an embodiment of the application;

[0034] Figure 10 A comparison diagram between the storm surge prediction result of the application and the prediction result of other prediction methods is provided for an embodiment of the application;

[0035] Figure 11 A structural diagram of a computer device is provided for an embodiment of the application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the application will be described clearly and completely in the following with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the application.

[0037] The above purposes, features and advantages of the application can be more obvious and easy to understand. The application will be further described in detail in the following with reference to the drawings and specific embodiments.

[0038] The site storm surge prediction method provided by the embodiments of the application can be applied in an application environment as shown in Figure 1 The application environment includes a terminal and a server. The terminal communicates with the server through a network. A data storage system can store data required to be processed by the server. The data storage system can be separately arranged, integrated on the server, placed on a cloud or other servers. The terminal can send the related data to be processed to the server. After receiving the data to be processed, the server can first store the data, and then acquire the data from the storage location when it needs to process the data, or can perform the processing task while storing the data. The server performs site storm surge prediction based on the related data. The server feeds back the predicted site storm surge prediction value to the terminal. In addition, in some embodiments, the site storm surge prediction method can also be realized by the server or the terminal alone, for example, the terminal can directly perform site storm surge prediction on the related data to be processed, or the server can acquire the related data to be processed from the data storage system and perform site storm surge prediction on the related data to be processed.

[0039] The terminal can be, but is not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server can be realized by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0040] In an exemplary embodiment, reference is made toFigure 2 As shown, a station storm surge prediction method is provided, which is executed by a computer device, specifically, can be executed by a terminal or a server, or both, in the embodiments of the present application, the method is applied to Figure 1 The server in the method is taken as an example for illustration, including the following steps 201 to 204:

[0041] Before describing the specific steps, it should be noted that the method uses the numerical prediction data of a target meteorological element on a certain day to predict storm surge, and for other target meteorological elements, the station storm surge prediction method provided by the present application can be used, which is similar to the present application, and the present application will not be described in detail.

[0042] Step 201, obtaining n feature vectors and sending them to a pre-trained TCN model, so that the pre-trained TCN model outputs n storm surge preliminary prediction values after the reporting time a, the n feature vectors are the feature vectors corresponding to the target meteorological elements in a predetermined range centered on the target station for N prediction times after the reporting time a.

[0043] Step 202, obtaining the storm surge value of the reporting time a and the M time before the reporting time a, obtaining M+1 storm surge values before the reporting time a.

[0044] It should be noted that for the convenience of description, the storm surge value of the reporting time a is also recorded as the storm surge value before the reporting time a.

[0045] Step 203, concatenating the M+1 storm surge values before the reporting time a and the n storm surge preliminary prediction values to obtain a mixed feature.

[0046] Step 204, sending the mixed feature to a pre-trained KAN model, so that the KAN model outputs n storm surge final prediction values after the reporting time a.

[0047] It should be noted that the present application uses the prediction data of the target meteorological element of the same station and the same time period to predict the storm surge prediction value of the same station and the same time period.

[0048] The station storm surge prediction method provided by the embodiments of the present application uses the TCN model to preliminarily predict the storm surge, concatenates the storm surge preliminary prediction with the storm surge observation value before the reporting time a to form a mixed feature, and then sends the mixed feature to the KAN model to use the KAN model to finally predict the storm surge, that is, the present application constructs a TCN-KAN hybrid model to improve the accuracy of storm surge prediction.

[0049] Optionally, in another example embodiment of the present application, the target meteorological element includes any one of a sea level pressure, a first predetermined height zonal wind speed, and a second predetermined height meridional wind speed.

[0050] The first predetermined height and the second predetermined height are predetermined heights from the sea level, which can be empirical values, for example, both are 10 meters.

[0051] Of course, other target meteorological elements can also be included in actual use, which are not limited in the present application.

[0052] Optionally, in another example embodiment of the present application, the predetermined range includes a predetermined number of latitude and longitude grid regions centered on the target site.

[0053] The present application takes the grid region as the meteorological feature background field information.

[0054] Optionally, referring to Figure 3 , in another example embodiment of the present application, the n feature vectors are generated by the following steps 401-402:

[0055] Step 401, for the target element, obtaining meteorological forecast data of N forecast horizons after the reporting time a.

[0056] Step 402, for the meteorological forecast data of the N forecast horizons, sliding according to the window size W and the sliding step L, generating a feature vector each time the window is slid, and obtaining the n feature vectors.

[0057] The n feature vectors are obtained by sliding the window according to the window size W and the sliding step L.

[0058] When the window is slid each time, the meteorological forecast data corresponding to the window is a feature vector, that is, the meteorological forecast data included in the feature vector is the same as the window size.

[0059] Optionally, in another example embodiment of the present application, before the step 401, the method further includes the following step 501:

[0060] Step 501, sending the N forecast horizons of meteorological forecast data corresponding to the first predetermined number of latitude and longitude grid regions to the principal component analysis model, so that the principal component analysis model performs dimensionality reduction processing on the N forecast horizons of meteorological forecast data, and obtains the N forecast horizons of meteorological forecast data of the second predetermined number of latitude and longitude grid regions after dimensionality reduction, wherein the second predetermined number is less than the first predetermined number.

[0061] That is, the dimensionality reduction is to reduce the number of region grids.

[0062] Further, the process of dimension reduction using principal component analysis is as follows:

[0063] Step one, flatten the two-dimensional grid data into a one-dimensional array. For each meteorological element Q, flatten the numerical prediction two-dimensional grid data of all the reporting times and prediction lead times into a one-dimensional array one by one, and stack them into the original matrix X Q .

[0064] Step two, diagonalize the matrix X Q . Let the covariance matrix of X Q be C Q , and P Q be a matrix composed of a set of bases by rows, let Y Q = P Q X Q , then Y Q is the data after the basis transformation of X Q to P Q . Let the covariance matrix of Y Q be D Q , then the relationship between D Q and C Q is expressed as the following formula (1):

[0065]

[0066] Where m is the number of samples. Thus, the optimization goal is converted to finding a matrix P Q that satisfies is a diagonal matrix, and the diagonal elements are arranged in descending order.

[0067] Step three, determine the feature vectors after dimension reduction. Calculate the contribution ratio of the principal component variance of each meteorological element Q, and when the sum of the contribution ratios of the first K principal component variances of all meteorological elements is greater than 85%, the first K diagonal elements are determined as the feature vectors of the meteorological element Q grid data after dimension reduction.

[0068] Exemplarily, for sea level pressure (MSLP), 10-meter zonal wind speed (U10) and 10-meter meridional wind speed (V10) elements, the regional background data extracted from each prediction lead time in each prediction file is flattened into a one-dimensional matrix with a length of 225, and then stacked into a two-dimensional matrix (m*120, 225) (where m is the total number of numerical prediction files collected), through the PCA method, 4 feature values are extracted from the 225 dimensions, and the principal component variance contribution ratio of each feature value is calculated, as shown in the following Table 1.

[0069] Table 1 Principal component variance contribution ratio of each feature after PCA dimension reduction

[0070] Meteorological element Feature 1 Feature 2 Feature 3 Feature 4 MSLP 0.9774 0.0133 0.0060 0.0014 U10 0.6627 0.1051 0.0453 0.0380 V10 0.7198 0.0970 0.0427 0.0390

[0071] According to the requirement that the proportion of the principal component contribution of the total characteristics is more than 85%, the first four characteristics are selected to form the feature vector after dimension reduction. Thus, the original sample with a dimension of (m*120, 225) is reduced to (m*120, 4), which greatly reduces the data amount on the basis of retaining most of the characteristics.

[0072] The site observation data and site neighborhood meteorological numerical prediction products are collected. The site observation data includes historical tidal level observation data and astronomical tide data, and the meteorological numerical prediction products include numerical prediction data of three elements, i.e., sea level pressure, 10-meter zonal wind speed and 10-meter meridional wind speed, within a certain range of the site neighborhood.

[0073] For example, the hourly tidal level and astronomical tide data of Shandong Dongying Port Station (latitude 38.14 °N, longitude 119.04 °E) from January 1, 2020 to December 31, 2024 are collected, and the 00 (universal time) meteorological numerical prediction products of the Copernicus Atmosphere Monitoring Service (CAMS) of the same period are downloaded from the European Centre for Medium-Range Weather Forecasts (ECMWF), including three elements, i.e., sea level pressure (MSLP), 10-meter zonal wind speed (U10) and 10-meter meridional wind speed (V10). The spatial resolution of the meteorological numerical prediction product is 0.4°, the prediction time is 120 hours, and the prediction time interval is 1 hour.

[0074] For site observation data, the threshold method is used for quality control of historical tidal level data, and abnormal data are deleted, and the linear interpolation method is used for interpolation processing of short-period missing tidal level data; the tidal level observation data at the corresponding time is subtracted by the astronomical tide data to obtain the storm surge time series data. For meteorological numerical prediction products, a certain size of neighborhood range is defined, and the grid data of each element and each prediction time in the specified neighborhood range is extracted as the meteorological characteristic background field information.

[0075] For example, for tidal level observation data, the threshold method is used to remove data records with tidal level greater than 30 meters or less than -30 meters, and the linear interpolation method is used for interpolation processing of data missing for no more than 3 hours, and missing for more than 3 hours is regarded as data interruption. After processing, the corresponding tidal level height is subtracted by the astronomical tide height to obtain the storm surge height. For meteorological numerical prediction products, 7 grids are taken in each of the four directions with the grid where Dongying Port Station is located as the center, and the area range of the meteorological characteristic background field of the station is defined, i.e., a square area of 15 rows and 15 columns with Dongying Port Station as the center.

[0076] Alternatively, referring to Figure 8 In another exemplary embodiment of the present application, the process of TCN model construction, training and prediction is as follows:

[0077] Step one, organize the feature dataset of target meteorological element forecast value. For meteorological element Q, the feature vector of target meteorological element forecast value at the bth time after the starting time a after dimension reduction is denoted as (by the dimension reduction process, is a one-dimensional array with a length of K). Let the sliding time window length be W, the sliding step length be L, the number of meteorological elements be C, and the maximum forecast validity of numerical forecast product be N. The sample organization process is shown in Figure 4 . When the end of the sliding time window is taken at time b<w, the feature vector of w-b-1 time is supplemented by the previous day numerical forecast product. Therefore, for each meteorological numerical forecast product, the feature vector with a dimension of is organized and generated under the condition that storm surge data is not missing.

[0078] After the feature data is organized, is the n th feature record of meteorological element Q at the starting time a in the forecast product file, and the storm surge value at the end of the forecast time in each feature record is extracted as the label data, denoted as As shown in Figure 4 , SL a,1 corresponds to the feature record time a+6. The label matrix extracted from each forecast product file is denoted as SL a , and the dimension is

[0079] Step two, data normalization and dataset division. The feature data and label data organized are normalized by the extreme value normalization method, which is represented by the following formulas (2) and (3):

[0080]

[0081] In the formula, Z i is the i th feature in the feature data, i=1,2,3,…,K*J; max(Z i ) is the maximum value in the sample space of Z i , min(Z i ) is the minimum value in the sample space of Z i ; X i is the normalized feature data; L is the label data, max(L) is the maximum value in the label space, min(L) is the minimum value in the label space; Y is the standardized label data. Then, a data subset of a continuous period of time at the end is extracted from the normalized dataset as a test set, and the previous data subset is randomly divided into a training set and a validation set according to a certain proportion (such as 8:2), forming the final sample dataset.

[0082] Step three, TCN model construction. The network structure and parameters of the TCN model are preliminarily drafted, the training set data is input into the TCN model for iterative training, and the training results are verified by the validation set data. The network structure of the TCN model is set as shown in Figure 5

[0083] Figure 5 In the above formula, the input data includes feature data and label data, and the dimensions are (None, W, CxK) and (None, 1) respectively; the TCN receives the input data, designs three hidden layers, and the channel numbers of the three hidden layers are H1, H2 and H3 respectively; the residual connection is used between the hidden layers; the convolution mode of the hidden layer is dilated convolution, that is, the more holes in the convolution window, the larger the dilated convolution coefficient d. The output data dimension of the hidden layer H3 is (None, W, H3), which is flattened to (None, WxH3) and input into the fully connected (FCN) layer; after the operation of the FCN layer, the model result, that is, the prediction data Y ' .

[0084] Step four, TCN model training. The model parameters (such as the number of nodes of the hidden layers H1 to H3, the dilated convolution coefficient, the convolution kernel size, etc.) are optimized and selected by the grid search method, and the combination with the smallest prediction error on the validation set is selected as the final parameters by enumerating multiple parameter combinations. The activation function of the model is ReLU, and the loss function is mean square error MSE. The number of training rounds of the model is determined, and after the model training is completed, the model structure, optimal parameters and weight data are saved to a file.

[0085] Step five, initial prediction of storm surge. The optimal model obtained by running the above training is input into the training set, the validation set and the test set, and the prediction value Y ' a,n is calculated for each feature record in the sample data set. ' a,n In the above formula, a is the start time of the forecast product, The same as the start time of the forecast product, a total of prediction values are calculated, and the time interval between adjacent prediction values is L.

[0086] Among them, the data from the storm surge time series data to the PCA dimension reduction module is the label data during training.

[0087] ​Exemplarily, TCN model construction, training and prediction. For each meteorological numerical prediction file, by the method of sliding window, set the sliding step to 1, the sliding window length to 24 (for time less than 24 hours, use the prediction data of the corresponding time in the previous day's numerical prediction file to supplement), and extract the dimension-reduced feature vector of each element every 24 hours in the 120-hour prediction time limit by sliding. In this way, each numerical prediction file extracts a matrix with dimensions (120, 24, 12). After extraction of each numerical prediction file, stack according to the first dimension to generate a feature data with dimensions (120*m, 24, 12). At the same time, extract the storm surge value at the end of each sliding window and organize it as label data. Then, independently standardize the extreme value according to the third dimension of the feature data and the label data in order to generate a normalized data set. The data from October 1, 2023 to December 31, 2024 is used as the test set, and the data between January 1, 2020 and September 30, 2023 is randomly divided into training set and validation set in the ratio of 8:2. Construct the TCN model, the input dimension is (None, 24, 12), determine the output channel numbers H1, H2 and H3 of the three hidden layers of TCN as 128, 32 and 8 respectively through grid search algorithm, then pass through the fully connected layer, and output the preliminary prediction data with dimension (None, 1). During the model training process, the maximum number of iterations is 1000, the minimum batch size is 400, the learning rate is 0.0001, the optimizer selects Adam algorithm, the activation function uses ReLU, the loss function uses MSE, and no early stopping mechanism is set. The processed training set is input into the model for training, and the validation set is used to verify the training result until the training is completed, and the optimal model structure, parameters and weights are saved. After training is completed, read and run the model file to predict all feature records in the training set, validation set and test set, and save the preliminary prediction results in the order of numerical prediction starting time and prediction time limit, with dimensions (m, 120).

[0088] Alternatively, in another exemplary embodiment of the present application, the process of KAN model construction, training and prediction is as follows:

[0089] Step one, organize mixed feature sample data set. For each meteorological numerical prediction file, the feature data organization includes two parts: one is the storm surge observation data S a-n (n = 0, 1, 2, 3,.., M) at the starting time a and the previous M times, and the other is the predicted value

[0090] p is the total number of meteorological numerical prediction products, and the organization method is shown in Figure 6 :

[0091] Therefore, the dimension of the feature data is , denoted as XL. The label data is the multi-step prediction value, denoted as SL a+t , where t is between 1 and N time (I is the forecast time interval), such as the 3-hour forecast Therefore, the dimension of the label data is

[0092] Here, I can be different from L, they can be the same or different, and they are independent of each other, that is, the time interval between the initial forecast and the final forecast can be different. For example, in this example, L takes the value of 1 and I takes the value of 3.

[0093] Step 2: Data normalization and dataset partitioning. The method is the same as in the TCN part to generate training set, validation set and test set.

[0094] Step 3: KAN model construction. Preliminary design of the network structure and parameters of the KAN model, input the above training set data into the KAN model for iterative training, and verify the training results with the validation set data. The network structure setting of the KAN model can be found in Figure 7 As shown:

[0095] Figure 7 The input layer of the model is The dimension feature data XL, the output layer is Dimension label data YL, KAN adopts a three-layer network structure design, the dimensions of each layer are H1, H2 and H3, and the KAN model directly outputs the predicted data with the dimension

[0096] Step 4: Model training. Parameter tuning and training methods are the same as those in the TCN part, generating model structure, parameter, and weight files.

[0097] Step 5: Model prediction. Run the optimal model obtained from the above training, input the test set, and calculate the predicted value for each feature record in the test set, which is recorded as YL' a,n . YL' a,n In the formula, a is the starting time of the forecast product, n is the forecast validity period of the model applied for,

[0098] Exemplarily, the storm surge observation data of the previous 72 hours from the start time of each numerical prediction file is selected as the start time of the sample data of the model, 120 prediction values in the corresponding time range are selected to organize feature data with a dimension of (None, 192). The 3-hour storm surge observation data from the start time of the model to the 120th hour after the start time is selected as the label data. The generated feature data and label data are normalized by the extreme value standardization method, and then the sample data set is divided into a training set, a validation set and a test set in the same way as in the TCN model. The KAN model is constructed, the input dimension is (None, 192), the three hidden layer dimensions H1, H2 and H3 are determined to be 128, 32 and 8 respectively by the grid search algorithm, the third-order B-spline activation function is used to connect between different layers, the grid size of the B-spline interval is defined as 5, and the output dimension is (None, 40). During the model training process, the maximum number of iterations is 1000, the minimum batch size is 100, the learning rate is 0.0001, the optimizer selects the Adam algorithm, the loss function uses MSE, and no early stopping mechanism is set. The training set data is input into the model for training, and the training results are verified by the validation set until the training is completed, and the optimal model structure, parameters and weights are saved. After the training is completed, the model file is read and run, the feature data in the test set is input into the model, the output result is obtained and de-normalized to obtain the prediction value.

[0099] Alternatively, for the prediction results of the above two models, the following method can be used for evaluation:

[0100] The root mean square error (RMSE) and the correlation coefficient (CC) are used to evaluate the model results, and the evaluation is divided according to different prediction times, and the calculation formulas are as follows (4) and (5):

[0101]

[0102] In the formula, m is the total number of samples in the prediction set of the model, YL i is the storm surge observation value of sample i, and YL' is the storm surge prediction value corresponding to sample i.

[0103] In addition, the RMSE and CC indicators are used to test and evaluate the model. In order to better compare the advantages of the application in multi-step storm surge prediction accuracy, two models are selected as references: one is the LSTM model with only storm surge time series data as input, to examine the effectiveness of the preliminary prediction module in extracting meteorological features; the other is the existing PCA-CL-LSTM model, to examine the superiority of the TCN-KAN combination method. In order to enhance the comparability of the prediction results, the LSTM model uses 72 hours of storm surge time series data as feature data, the PCA-CL-LSTM model has the same prediction time limit as the application model, and the model structure and parameter selection are the same as in the TCN model.

[0104] Within 120 hours of prediction time, the 6-hourly RMSE and CC indicators between the application model, the PCA-CL-LSTM model, the LSTM model and the observed values are calculated, and the statistical results are shown in the table. As can be seen from the table, the application model has higher prediction accuracy at each time, with an average RMSE error of 10.3% lower than the PCA-CL-LSTM model and 33.5% lower than the LSTM model; the average CC is 0.06 lower than the PCA-CL-LSTM and 0.61 lower than the LSTM model. The prediction results of the LSTM model after 24 hours are poor, indicating that using only time series observation data as feature data cannot predict long-term storm surge changes, and verifying the effectiveness of the preliminary prediction module in extracting meteorological features. On the other hand, the application has a significant improvement in prediction accuracy over the PCA-CL-LSTM model, indicating that the TCN-KAN combination method has more superior fitting and generalization ability than the CL-LSTM combination method.

[0105] Table 2 Comparison of prediction errors between the application model and the existing invention

[0106]

[0107] Note: The decrease in the correlation coefficient cc is the difference between the absolute values of the corresponding values

[0108] For example, as shown in Figure 9 and Figure 10 , the two cold air processes on October 16, 2024 and December 19, 2024 are selected as example periods to compare the differences in prediction accuracy between the application model and the existing model, and the prediction curve graphs are shown in Figure 8 and 9 . The two cold air processes caused serious storm surge and storm surge processes, respectively. As shown in Figure 9 and Figure 10It can be seen that the prediction results of the model of the application are closer to the observed values in the two cold air processes, and the extreme value prediction is more accurate. According to statistics, in the cold air process on October 16, 2024, the prediction RMSE of the model of the application, the PCA-CL-LSTM model and the LSTM model are 23.05 cm, 28.22 cm and 58.45 cm respectively; in the cold air process on December 19, 2024, the prediction RMSE of the model of the application, the PCA-CL-LSTM model and the LSTM model are 16.70 cm, 23.81 cm and 40.00 cm respectively. The LSTM model based on time series observation data prediction cannot simulate the change of storm surge for 120 hours, and the model of the application and the PCA-CL-LSTM model can predict the change process of storm surge to a certain extent, but the prediction error of the model of the application in the two cold air processes is reduced by 18.3% and 28.9% respectively compared with the PCA-CL-LSTM model.

[0109] In summary, the application proposes an improved intelligent prediction method for site storm surge, which has high prediction accuracy. The application learns from the modeling idea of the PCA-CL-LSTM model "two steps", processes the meteorological background field characteristic data by the PCA method, uses the same data organization method to construct the sample data set, but uses the TCN model instead of the Conv1D LSTM model for preliminary prediction, and uses the KAN model instead of the LSTM model for final prediction, and constructs a PCA-TCN-KAN hybrid model. The examples prove that the model of the application can realize more accurate long-time storm surge prediction, and the RMSE is reduced by more than 10% compared with the PCA-CL-LSTM model within 120 hours, and can show more excellent fitting ability in the disaster process, to a certain extent, to alleviate the mean regression problem of machine learning model.

[0110] Optionally, the programming language used by the application is Python 3.9.13, the machine learning framework is Pytorch1.12.1, and the operating system is Windows Server 2019Standard. The server hardware conditions used by the application are as follows: Intel(R) Xeon(R) Gold 6226 CPU @ 2.70 GHz, 256 GB memory.

[0111] The application provides a storm surge intelligent prediction combination model based on TCN and KAN, aiming to improve the prediction accuracy of storm surge. The model mainly solves the following two problems: first, the training data of storm surge itself has deviation, and has obvious long-term dependence and multi-scale characteristics, which leads to the problem of easy gradient disappearance / explosion in the training process of the existing model (such as LSTM); second, the sparsity of extreme value samples of storm surge may lead to the phenomenon of over-regression of mean value in the training result. Although the existing technology often uses regularization method, dropping strategy or early stopping mechanism to alleviate the above problems, it leads to underfitting of the model to a certain extent, and then affects the convergence accuracy of the model. Therefore, the application learns from the "two-step" modeling strategy, introduces more advanced TCN and KAN models, and constructs an improved storm surge intelligent prediction combination model: first, the original meteorological numerical prediction data is processed by PCA method, and the dimensionality-reduced data is organized as meteorological spatio-temporal feature samples, and the TCN model is used for preliminary prediction; then, the historical storm surge observation data and the TCN model preliminary prediction result are spliced in time sequence to form a hybrid feature sample, and the KAN model is used for multi-step storm surge final prediction.

[0112] Compared with the prior art PCA-Conv1DLSTM-LSTM model (hereinafter referred to as PCA-CL-LSTM model), the advantage of the application mainly lies in the accuracy of storm surge prediction. In the prior art, the Conv1DLSTM model mainly captures local patterns, and its receptive field is relatively limited, and the long-term dependence is mainly captured by LSTM, but the ability of LSTM to capture very long distance dependence will decrease with the increase of sequence length. On the other hand, although LSTM alleviates the gradient disappearance problem of simple RNN through the gating mechanism, the gradient problem still occurs when dealing with very long sequences. Compared with this, TCN flexibly controls the receptive field of the model by stacking convolutional layers, adjusting the size of the convolution kernel and using a hollow convolution, and through the exponential growth of the hollow factor, TCN uses fewer layers to obtain a larger receptive field, thereby efficiently capturing long-term dependence. At the same time, KAN has a learnable activation function (such as a spline function) located at the edge of the neural network, which has stronger function representation ability than the network with fixed activation function, so in the modeling task related to the physical process, KAN can achieve higher accuracy with fewer parameters. In summary, the application adopts TCN and KAN combination modeling, uses TCN as a powerful spatio-temporal sequence feature extractor, and then inputs the features extracted by TCN and the time series observation data into KAN for final mapping to obtain more accurate function fitting.

[0113] In an exemplary embodiment, a computer device, which can be a server or a terminal, is provided, and its internal structure diagram can be seen from Figure 11As shown in the figure. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store site storm surge prediction related data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor, and a site storm surge prediction method can be realized.

[0114] Those skilled in the art can understand that, Figure 11 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0115] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in each of the above method embodiments.

[0116] In one exemplary embodiment, a computer readable storage medium is provided, storing a computer program, which is executed by a processor to implement the steps in each of the above method embodiments.

[0117] In one exemplary embodiment, a computer program product is provided, including a computer program, which is executed by a processor to implement the steps in each of the above method embodiments.

[0118] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0119] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. In the embodiments provided in the present application, any reference to the memory, database or other medium can include at least one of the non-volatile and volatile memories. The non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. The volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

[0120] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0121] The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0122] The principles and implementation modes of the present application are described by using specific examples in the present application. The above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A site storm surge prediction method, characterized in that: include: Obtaining n eigenvectors and sending them to a pre-trained TCN model so that the pre-trained TCN model outputs n preliminary storm surge prediction values ​​after the onset time a, wherein the n eigenvectors are eigenvectors corresponding to the N forecast time limits of the target meteorological element within a predetermined range centered on the target site after the onset time a; Obtain the previous storm surge values ​​of the starting time a and its M preceding time periods, and obtain M+1 preceding storm surge values; splicing the M+1 previous storm surge values ​​and the n storm surge preliminary prediction values ​​to obtain a mixed feature; The mixed features are sent to a pre-trained KAN model, so that the KAN model outputs n final storm surge prediction values ​​after the onset time a.

2. The site storm surge prediction method according to claim 1, characterized in that: The target meteorological element includes: any one of sea level pressure, latitudinal wind speed at a first predetermined height, and longitudinal wind speed at a second predetermined height.

3. The site storm surge prediction method according to claim 1, characterized in that: The predetermined range includes a predetermined number of latitude and longitude grid areas centered on the target site.

4. The site storm surge prediction method according to claim 1, characterized in that: The n eigenvectors are generated by the following method: For the target element, obtain weather forecast data for N forecast periods after the start time a; For the N weather forecast data of the forecast time, slide according to the window size W and the sliding step size L, and generate a feature vector each time you slide, to obtain the n feature vectors.

5. The site storm surge prediction method according to claim 4, characterized in that: For the target element, before obtaining the weather forecast data of N forecast periods after the start time a, the method further includes the following method: The meteorological forecast data for the N forecast periods corresponding to a first predetermined number of latitude and longitude grid areas are sent to a principal component analysis model, so that the principal component analysis model performs dimensionality reduction processing on the meteorological forecast data for the N forecast periods to obtain the meteorological forecast data for the N forecast periods of a second predetermined number of latitude and longitude grid areas after dimensionality reduction, wherein the second predetermined number is smaller than the first predetermined number.

6. The site storm surge prediction method according to claim 2, characterized in that: The first predetermined height is 10 meters, and the second predetermined height is 10 meters.

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the site storm surge prediction method according to any one of claims 1 to 7.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the site storm surge prediction method according to any one of claims 1 to 7 is implemented.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the site storm surge prediction method according to any one of claims 1 to 7 is implemented.

Images