Prediction method suitable for short-term rapid wave forecasting during construction period

Abstract

The invention provides a prediction method suitable for short-term rapid wave prediction during a construction period, which is suitable for the field of sea wave prediction and comprises the steps of planning a construction area, planning typical point locations, obtaining wave element information, constructing a data set, processing data, extracting features, constructing a time convolutional network prediction model, training, testing and optimizing the model and applying the model. According to the method, a point location latitude and longitude-wave element data set of an engineering sea area is constructed through an SWAN / WAVEWATCH mode, the influence of factors such as water depth topography, complex shorelines, islands, shoreline wave breakage and bottom friction on wave propagation can be considered, and it is guaranteed that data contain enough physical parameter information influencing waves; through the time convolution network, diversified activation function selection, systematic data processing and feature extraction, the analysis capability of the model is enhanced, the accuracy and rapidity of wave element prediction are improved, selection of a construction window period is facilitated, reliable support is provided for construction decision making, and thus the construction risk is reduced.

Description

Technical Field

[0001] The invention relates to a prediction method suitable for short-term rapid prediction of waves during a construction period, and is suitable for the field of sea wave prediction. Background Art

[0002] Accurate short-term forecast of ocean waves has an important impact on production planning and engineering safety of offshore vessel operations.

[0003] At present, the third-generation numerical prediction models are more commonly used for simulating waves, such as the SWAN (simulating waves nearshore) model and the WAVEWATCH model. These models comprehensively consider the influence of multiple factors such as topography, ocean currents, sea temperature difference, and shallow water deformation of waves. They are widely used in the fields of wave generation and propagation, wave energy resource assessment, etc. However, these models have large computational complexity, long time consumption, and high complexity when predicting waves.

[0004] Machine learning methods avoid iterative solutions of physical equations in the third-generation numerical forecasting model by modeling and predicting data, and have the advantage of fast calculation, such as long short-term memory neural network LSTM (long short-term memory), NLP (natural language-processing), CNN (Convolutional Neural Networks), etc. However, these methods are mostly used for training and prediction of single-point and single-item wave elements. In actual sea area wave prediction, engineering operations are mostly carried out in a sea area, and the operation points change in space. At the same time, it is necessary to know the prediction information of multiple wave elements at the operation points, such as significant wave height, average period, spectral peak period, average wave direction, swell component in waves, wind and wave component in waves, etc. At the same time, the data change forms of different wave elements are different, and targeted adjustments need to be made.

[0005] In order to quickly forecast waves in an operating sea area during construction and simultaneously forecast multiple wave elements at multiple locations, a prediction method suitable for short-term rapid forecasting of waves during the construction period is needed. Summary of the invention

[0006] The purpose of the present invention is to propose a prediction method suitable for short-term rapid forecast of waves during the construction period.

[0007] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0008] A prediction method suitable for short-term rapid forecasting of waves during construction period comprises the following steps:

[0009] S101 Planning construction area;

[0010] The planned construction area includes a sea surface construction area planned according to construction needs. The construction area is rectangular with dimensions of a km × b km;

[0011] S102 Plan typical points;

[0012] The planned typical points include planning n typical points marked as D1 to D according to the shape of the construction area, n and at the same time obtaining the longitude and latitude information of the typical points D1 to D, n The typical points are evenly distributed in the construction area, and n is calculated by formula (1),

[0013]

[0014] In the formula, ceiling is the ceiling function and L is the spacing;

[0015] S103 Obtain wave element information and construct a data set;

[0016] The obtaining wave element information and constructing a data set includes using a wave model to predict the wave element information of the construction area, and constructing a data set according to the prediction result of the wave element information. The data set includes the longitude and latitude information of the typical points and the corresponding wave element information. The wave element information includes parameters such as average wave height, significant wave height, average period, spectral peak period, and average wave direction;

[0017] S104 Data processing;

[0018] The data processing includes performing data processing on the data set;

[0019] S105 Extract features;

[0020] The extracting features includes performing data feature extraction on the wave element information in the data set;

[0021] S106 Construct a temporal convolutional network prediction model;

[0022] The constructing a temporal convolutional network prediction model includes constructing a temporal convolutional network prediction model for the wave element information;

[0023] S107 Training, testing, and optimizing the model;

[0024] The training, testing, and optimizing the model includes using the data in the data set to train and test the temporal convolutional network prediction model, and optimizing according to the test results to obtain the optimized temporal convolutional network prediction model;

[0025] S108 Application of the model;

[0026] The application of the model includes applying the time convolutional network prediction model after optimization to the prediction of wave element information at typical points in the construction area.

[0027] Further, in the above step S103, the types of the wave element information include the one-tenth wave height, average wave height, significant wave height, average period, spectral peak period, average wave direction, swell component in the wave, and wind wave component in the wave of the typical points. The content of the wave element information includes the time history change data corresponding to each type for each typical point of the wave element information.

[0028] Further, in the above step S104, the data processing includes performing kernel principal component analysis to reduce the dimension of the data of each typical point in the dataset. The steps of performing kernel principal component analysis to reduce the dimension of the data of each typical point in the dataset are as follows: a) Select a Gaussian kernel function and calculate the kernel matrix between data points; b) Perform eigenvalue decomposition on the kernel matrix, extract the non-linear features of the data, and retain the eigenvectors corresponding to the largest several eigenvalues as the principal components, and finally project the data onto these principal components to achieve dimensionality reduction.

[0029] Further, in the above step S105, the feature extraction includes performing statistical feature extraction on the wave element information in the dataset. The extracted indicators include mean, variance, extreme value, wave band, zero-crossing wave, long and short peak values of the wave band, sum of squares of the time series, approximate entropy, autoregressive model coefficients, autocorrelation coefficients at lag, number of numbers greater than the mean, number of numbers less than the mean, position where the maximum value first appears, length of the longest continuous subsequence greater than the mean, length of the longest continuous subsequence less than the mean, mean of the absolute values of the continuously changing values, mean of the continuously changing values, number of values that appear more than once / total number, proportion of values greater than n times the standard deviation, entropy, standard deviation, number of points that appear multiple times, variance of the values that appear multiple times.

[0030] Further, in the above step S106, the construction of the time convolutional network prediction model includes designing the model structure, selecting the loss function and the optimization algorithm. The time convolutional network prediction model is constructed separately according to the types of wave element information to be predicted. The activation function of the constructed time convolutional network prediction model selects different activation functions according to the types of wave element information to be predicted.

[0031] Even further, the steps of selecting different activation functions according to the types of wave element information to be predicted are as follows:

[0032] a) Obtain the typical history data change curve of the types of wave elements to be predicted;

[0033] b) Determine whether the process data change curve is binary data. If it is, select the linear enhanced growth curve activation function and do not perform the following steps. If not, proceed to the next step. The expression of the linear enhanced growth curve activation function is Equation (2).

[0034]

[0035] In the formula, a is a fixed parameter with a value of 25;

[0036] c) For non-binary data, select the piecewise enhanced activation function. The expression of the piecewise enhanced activation function is Equation (3).

[0037]

[0038] In the formula, a is the control leakage parameter, b is the saturation threshold, and c is a fixed parameter and c = b(1 - ab).

[0039] Further, in the above step S107, the model training includes data splitting of the dataset into a training set and a test set, and the model optimization includes hyperparameter optimization.

[0040] The present invention has the following beneficial effects: The time convolution network prediction model can effectively capture the time series changes of wave elements. Using deep learning technology can improve the accuracy and rapidity of prediction. This will help reduce construction risks and enhance construction safety; Using kernel principal component analysis for data dimensionality reduction can effectively extract the non-linear characteristics of wave elements, reduce feature redundancy, and retain key information, thereby improving the training efficiency and performance of the model; Selecting appropriate activation functions according to different types of wave elements can optimize the performance of the model in specific tasks, enhance the adaptability of the model, and ensure that the model can select the best performance method according to different data types. By predicting the wave parameters of a region through the SWAN / WAVEWATCH model and constructing a dataset of longitude and latitude - wave elements at the points in the engineering sea area, the influence of factors such as water depth topography, complex shorelines, islands, breaking of shore waves, and bottom friction on wave propagation can be considered. Combining the advantages of rapidity of machine learning prediction, this method simultaneously ensures the physical accuracy and temporal rapidity of the prediction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flowchart of a prediction method for short-term rapid prediction of waves during the construction period according to the present invention;

[0042] Figure 2 It is the data information of the significant wave height of the wave element data in the embodiment of the present invention;

[0043] Figure 3Data information of wave element data - spectral peak period in the embodiments of the present invention;

[0044] Figure 4 Data information of wave element data - mean wave direction in the embodiments of the present invention;

[0045] Figure 5 Basic data and predicted values of significant wave height in the embodiments of the present invention;

[0046] Figure 6 Basic data and predicted values of spectral peak period in the embodiments of the present invention. Detailed implementation manners

[0047] The following further elaborates on the specific embodiments of the present invention in conjunction with the accompanying drawings; it should be understood that the specific embodiments provided herein are only for the purpose of illustrating and explaining the present invention and should not be used to limit the present invention.

[0048] The following is a specific embodiment of a prediction method applicable to short - term and rapid prediction of waves during the construction period.

[0049] A prediction method applicable to short - term and rapid prediction of waves during the construction period includes the following steps:

[0050] S101 Plan the construction area;

[0051] The planning of the construction area includes planning the sea - surface construction area according to the construction needs. The construction area is rectangular with dimensions of a km × b km;

[0052] S102 Plan typical points;

[0053] The planning of typical points includes planning the number of typical points as n, labeled as D1 to D n , and simultaneously obtaining the longitude and latitude information of typical points D1 to D n . The typical points are evenly distributed in the construction area, and n is calculated by formula (1),

[0054]

[0055] where ceiling is the ceiling function and L is the spacing;

[0056] S103 Obtain wave element information to construct a data set;

[0057] The acquisition of wave element information to construct a data set includes using a wave model to predict the wave element information of the construction area, and constructing a data set according to the prediction results of the wave element information. The data set includes the longitude and latitude information of typical points and the corresponding wave element information. The wave element information includes parameters such as mean wave height, significant wave height, mean period, spectral peak period, and mean wave direction;

[0058] Further, in the above step S103, the types of the wave element information include the one-tenth wave height, mean wave height, significant wave height, mean period, spectral peak period, mean wave direction, swell component in the wave, and wind wave component in the wave of typical points. The content of the wave element information includes the time history change data of each type corresponding to each typical point of the wave element information.

[0059] In this embodiment, the wave parameters of a region are predicted through the SWAN / WAVEWATCH model, and a data set of longitude and latitude of points - wave elements in the engineering sea area is constructed. The third-generation WAVEWATCH wave model completely describes the physical process of wave evolution and can consider the influence of factors such as water depth topography, complex shorelines, and islands on wave propagation. SWAN considers nearshore physical processes such as breaking of shore-breaking waves and bottom friction in the nearshore area. This patent adopts the method of mutual nesting of the WAVEWATCH and SWAN models to simulate the wave field of a certain construction sea area in the East China Sea. The specific area and parameters are shown in Table 1, and at the same time, wave element data such as Figure 2-4 shown.

[0060] Table 1 Parameter settings for the mutual nesting of the WAVEWATCH and SWAN models

[0061]

[0062] S104 Data processing;

[0063] The data processing includes performing data processing on the data set;

[0064] Further, in the above step S104, the data processing includes performing dimensionality reduction on the data of each typical point in the data set by adopting kernel principal component analysis. The steps of performing dimensionality reduction on the data of each typical point in the data set by adopting kernel principal component analysis are as follows: a) Select a Gaussian kernel function and calculate the kernel matrix between data points; b) Perform eigenvalue decomposition on the kernel matrix, extract the non-linear features of the data, and retain the eigenvectors corresponding to the largest several eigenvalues as the principal components, and finally project the data onto these principal components to achieve dimensionality reduction.

[0065] S105 Feature extraction;

[0066] The feature extraction includes performing data feature extraction on the wave element information in the data set;

[0067] Further, in the above step S105, the feature extraction includes performing statistical feature extraction on the wave element information in the dataset, and the extracted metrics include mean, variance, extreme value, wave band, zero-crossing wave, long and short peaks of the wave band, sum of squares of the time series, approximate entropy, autoregressive model coefficients, lagged autocorrelation coefficients, number of numbers greater than the mean, number of numbers less than the mean, position where the maximum value first appears, length of the longest consecutive subsequence greater than the mean, longest consecutive subsequence less than the mean, mean of the absolute values of the continuously changing values, mean of the continuously changing values, number of values that appear more than once / total number, proportion of values greater than n times the standard deviation, entropy, standard deviation, number of points that appear multiple times, variance of the values that appear multiple times.

[0068] S106 Construct a time convolutional network prediction model;

[0069] The constructing of the time convolutional network prediction model includes constructing a time convolutional network prediction model for the wave element information.

[0070] Further, in the above step S106, the constructing of the time convolutional network prediction model includes designing the model structure, selecting the loss function and the optimization algorithm. The time convolutional network prediction model is constructed separately according to the types of wave element information to be predicted, and different activation functions are selected for the time convolutional network prediction model according to the types of wave element information to be predicted.

[0071] Even further, the step of selecting different activation functions according to the types of wave element information to be predicted is as follows:

[0072] d) Obtain the typical history data change curve of the types of wave elements to be predicted;

[0073] e) Determine whether the history data change curve is binary data. If it is, select the linear enhanced growth curve activation function and do not perform the subsequent steps. If not, proceed to the next step. The expression of the linear enhanced growth curve activation function is Equation (2).

[0074]

[0075] In the formula, a is a fixed parameter with a value of 25;

[0076] f) For non-binary data, select the piecewise enhanced activation function. The expression of the piecewise enhanced activation function is Equation (3).

[0077]

[0078] Wherein, a is a leakage control parameter, b is a saturation threshold, and c is a fixed parameter and c = b(1 - ab).

[0079] Training, testing, and optimization of the S107 model;

[0080] The training, testing, and optimization of the model include using the data in the dataset to train and test the temporal convolutional network prediction model, and optimizing according to the test results to obtain the optimized temporal convolutional network prediction model;

[0081] Further, in the above step S107, the model training includes data splitting of the dataset into a training set and a test set, and the model optimization includes hyperparameter optimization.

[0082] Application of the S108 model;

[0083] The application of the model includes applying the optimized temporal convolutional network prediction model to the forecasting of wave element information at typical points in the construction area.

[0084] In this embodiment, 2 typical wave element data (significant wave height and spectral peak period) are selected, and the predicted values obtained by using this method are compared with the basic data as Figure 5 and Figure 6 shown. The blue part is the predicted part, and the predicted values obtained through data analysis have achieved the expected effect and meet the requirements of the engineering practice.

[0085] In the above embodiment, the present invention discloses a prediction method applicable to short-term and rapid forecasting of waves during the construction period, including planning the construction area, planning typical points, obtaining wave element information to construct a dataset, data processing, feature extraction, constructing a temporal convolutional network prediction model, training, testing, and optimization of the model, and application of the model; predicting the wave parameters of an area through the SWAN / WAVEWATCH model, constructing a dataset of longitude and latitude - wave elements of points in the engineering sea area, and considering factors such as water depth topography, complex shorelines, islands, breaking of shore - pounding waves, and bottom friction, etc., on the wave propagation, ensuring that the data contains sufficient physical parameter information affecting waves. Using kernel principal component analysis for data dimensionality reduction, effectively extracting the non - linear features of wave elements, selecting appropriate activation functions according to different types of wave elements, can optimize the performance of the model in specific tasks, enhance the adaptability of the model, and using the temporal convolutional network prediction model can effectively capture the time - series changes of wave elements. Using deep learning technology can improve the accuracy and rapidity of prediction. This method can provide reliable support for the selection of the construction window period and construction decision - making, and reduce the construction risk.

[0086] The above is a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A prediction method suitable for short-term rapid forecasting of waves during construction period, characterized in that: The following steps are involved: S101 Planning construction area; S102 planning typical points; S103 obtains wave element information to construct a data set; S104 data processing; S105 extracts features; S106 constructs a temporal convolutional network prediction model; Training, testing and optimization of the S107 model; Application of the S108 model; The planned construction area includes a sea surface construction area planned according to construction needs, and the construction area is rectangular with a size of akm×bkm; The planned typical points include n planned typical points marked as D1 to D2 according to the shape of the construction area. n , and obtain typical points D1~D n The latitude and longitude information of the typical points are evenly distributed in the construction area, and n is calculated by formula (1). In the formula, ceiling is the upward rounding function, and L is the spacing; The step of obtaining the wave element information and constructing a data set includes using a wave model to predict the wave element information of the construction area, and constructing a data set according to the wave element information prediction result, wherein the data set includes the latitude and longitude information of typical points and the wave element information corresponding thereto, and the wave element information includes parameters such as average wave height, significant wave height, average period, spectrum peak period, average wave direction, etc.; The data processing includes performing data processing on the data set; The extracting of features comprises extracting data features from the wave element information in the data set; The constructing of the temporal convolutional network prediction model comprises constructing a temporal convolutional network prediction model for wave element information; The training, testing and optimization of the model includes using the data of the data set to perform model training and testing on the temporal convolutional network prediction model, optimizing according to the test results, and obtaining the optimized temporal convolutional network prediction model; The application of the model includes applying the optimized temporal convolutional network prediction model to the forecasting of wave element information at typical points in the construction area.

2. A prediction method suitable for short-term rapid forecast of waves during construction period according to claim 1, characterized in that: In step S103, the types of wave element information include one-tenth wave height, average wave height, significant wave height, average period, spectral peak period, average wave direction, swell component and wind wave component of waves at typical points, and the content of the wave element information includes time history change data of each type corresponding to each typical point of the wave element information.

3. A prediction method suitable for short-term rapid forecast of waves during construction period according to claim 1, characterized in that: In step S104, the data processing includes taking kernel principal component analysis to reduce the dimension of the data of each typical point in the data set, and the steps of taking kernel principal component analysis to reduce the dimension of the data of each typical point in the data set are: a) selecting a Gaussian kernel function and calculating the kernel matrix between data points, b) performing eigenvalue decomposition on the kernel matrix to extract the nonlinear characteristics of the data, and retaining the eigenvectors corresponding to the largest eigenvalues ​​as principal components, projecting the data onto the principal component to finally achieve dimensionality reduction.

4. A prediction method suitable for short-term rapid forecast of waves during construction period according to claim 1, characterized in that: In the step S105, the feature extraction includes performing statistical feature extraction on the wave element information in the data set, and the extracted indicators include mean, variance, extreme value, band, zero-crossing wave, long and short peaks of the band, sum of squares of time series, approximate entropy, autoregressive model coefficient, lagged autocorrelation coefficient, the number of numbers greater than the mean, the number of numbers less than the mean, the position of the first occurrence of the maximum value, the length of the longest continuous subsequence greater than the mean, the longest continuous subsequence less than the mean, the mean of the absolute value of continuous length change values, the mean of continuous change values, the number of values ​​that appear more than once / total number, the proportion of values ​​greater than n times the standard deviation, entropy, standard deviation, the number of points that appear multiple times, and the variance of values ​​that appear multiple times.

5. A prediction method suitable for short-term rapid forecast of waves during construction period according to claim 1, characterized in that: In step S106, the constructing of the temporal convolutional network prediction model includes designing a model structure, selecting a loss function and an optimization algorithm. The constructing of the temporal convolutional network prediction model is constructed according to the type of wave element information that needs to be predicted. The activation function of the constructing of the temporal convolutional network prediction model selects different activation functions according to the type of wave element information that needs to be predicted.

6. A prediction method suitable for short-term rapid forecast of waves during construction period according to claim 5, characterized in that: The step of selecting different activation functions according to the type of wave element information to be predicted is: a) Obtaining the typical history data change curve of the wave element type to be predicted; b) Determine whether the history data change curve is binary data. If yes, select the linear enhancement growth curve activation function and do not proceed to the following steps. If no, proceed to the next step. The expression of the linear enhancement growth curve activation function is formula (2): In the formula, a is a fixed parameter with a value of 25; c) For non-binary data, a piecewise enhancement activation function is selected. The piecewise enhancement activation function is expressed as formula (3): Wherein, a is the control leakage parameter, b is the saturation threshold, c is a fixed parameter and c=b(1-ab).

7. According to a prediction method suitable for short-term rapid forecast of waves during construction period according to claim 1, it is characterized in that: In step S107, the model training includes segmenting the data set into a training set and a test set, and the model optimization includes hyperparameter optimization.

Images