A sea wave missing data prediction method based on a bivariate long short-term memory algorithm
By constructing a wave missing data forecasting method based on bivariate long short-term memory algorithm, the problem of inaccurate forecasting caused by missing buoy data is solved. By using regression model and machine learning to impute data and combining it with a wave variable association database, fast and accurate wave missing data forecasting is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2023-01-12
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies struggle to effectively address the issue of missing buoy data in wave forecasting, leading to inaccurate forecast results and an inability to effectively utilize wave generation factors and correlations, thus affecting the reliability of forecasts.
A bivariate long short-term memory algorithm-based approach is adopted. By imputing single missing data through regression model and continuously missing data through machine learning, a bivariate LSTM model is constructed. The ocean wave variable association database is used to match relevant variables for data imputation and prediction.
It improves the accuracy and reliability of forecasting missing wave data, reduces model complexity and computational load, and enables rapid and accurate forecasting of missing data.
Smart Images
Figure CN116245018B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ocean wave missing data forecasting technology, specifically to an ocean wave missing data forecasting method based on a bivariate long short-term memory algorithm. Background Technology
[0002] Ocean waves are a common marine phenomenon, primarily powered by wind. Wind transmits energy to the sea surface, causing the water to rise and fall, forming waves. Besides wind, wave formation is also influenced by other marine and atmospheric factors, making the wave propagation process highly complex, and certain constraints exist between wave elements. Ocean waves possess immense destructive power and are closely related to human life; therefore, understanding their distribution characteristics and patterns of change is crucial. Observational methods such as buoys and radar can obtain accurate and reliable sea surface data, allowing us to reconstruct the spatiotemporal distribution and changes of ocean waves.
[0003] Beyond observing ocean waves, forecasting them is even more crucial. The earliest developed numerical wave forecasting methods are based on observational data and theoretical research, combining current wave conditions within a region to calculate and predict future wave conditions. Numerical wave forecasting has become a widely adopted method in wave forecasting research. Based on the laws governing wave generation, dissipation, and propagation, numerical wave models can simulate and forecast waves within a study area. Numerical models use partial differential equations for calculations. Partial differential equations describing ocean processes are complex. Improving the accuracy of model forecasts requires adding more influencing factors, which increases model complexity and computation time. Although researchers both domestically and internationally widely use numerical models to study ocean processes, physical modeling is not conducive to improving forecast efficiency.
[0004] With the development of technology, computer performance has been greatly improved, creating a favorable environment for the development of machine learning algorithms. On the one hand, artificial intelligence calculations are simpler than numerical methods, do not require a clear understanding of the changing mechanisms of physical processes, and to some extent compensate for the shortcomings of numerical models. Furthermore, they can complement numerical models, greatly helping to improve the timeliness and accuracy of wave forecasts. On the other hand, artificial intelligence methods provide a new means of wave forecasting, expanding the scope of marine research and having far-reaching significance for marine observation and marine disaster early warning. Currently, many researchers at home and abroad have applied this research method to wave forecasting and have made many discoveries. For example, in Gao Libin's research, "Research on Wave Forecasting in the Taiwan Strait and Surrounding Waters Based on Deep Learning," wave height, wind speed, wind direction, and wind speed at the time of forecast were used as four input factors in an LSTM model to train buoy data. The results showed that the more input variables and the more input periods (historical time series), the better the prediction results. However, if there is a significant amount of missing buoy data, the forecast results will be greatly reduced.
[0005] Meanwhile, numerous studies have shown that wave generation is closely related to wind speed and other factors. The generated waves and their periods satisfy the wave dispersion relation. For example, the "Vincentian wave spectrum" takes into account the relationship between wind and wave growth and overcomes the dissipation problem of wave patterns by adding external conditions. This indicates that there is a strong correlation between waves and wind, and the influence of wind cannot be ignored when studying wave changes.
[0006] Therefore, it is particularly important to understand how to process missing buoy data to improve the accuracy of forecasts, how to use machine learning algorithms to mine and analyze the changing patterns of data based on the characteristics of wave generation, the factors that influence wave generation, and the correlation between waves, and how to obtain a series of complex and nonlinear ocean features through training and learning, so as to improve the reliability of forecasts based on missing wave data. Summary of the Invention
[0007] The purpose of this invention is to provide a method for forecasting missing wave data based on a bivariate long short-term memory algorithm. This method interpolates and corrects buoy data and constructs a bivariate LSTM model based on the relationship between wave generation characteristics and buoy data to achieve reliable forecasting of missing wave data.
[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for forecasting missing wave data based on machine learning, comprising:
[0009] S1. Obtain the buoy dataset A;
[0010] S2. Use regression model to impute single missing buoy data, and use machine learning to impute continuously missing buoy data, thereby forming a new buoy dataset B;
[0011] S3. Using the wave variable correlation database, find the correlation between the predicted variable and the wave variable in the buoy dataset B. Match the variables and construct a bivariate set V, and determine the training set and validation set for each bivariate in the bivariate set V.
[0012] S4. Substitute the training sets of each bivariate model into the Long Short-Term Memory algorithm for training to obtain the corresponding bivariate LSTM model.
[0013] S5. Input the validation set of each variable into the corresponding bivariate LSTM model to predict the missing variables, and compare and analyze each prediction result with the corresponding validation set data to evaluate the prediction performance of each bivariate LSTM model and obtain the prediction results. Determine the optimal prediction model based on the preset conditions and prediction results.
[0014] The process of imputing consecutively missing buoy data using machine learning and imputing single missing buoy data using regression models follows a specific order. First, missing buoy data in buoy dataset A is imputed using regression models. Then, machine learning is used to imput consecutively missing buoy data, forming a new buoy dataset B. Buoy dataset A contains variable data such as wind speed and wave period. Imputing missing data in the acquired buoy dataset A first helps to avoid the impact of missing buoy data on subsequent model building.
[0015] Based on the above technical solution, the regression model is as follows:
[0016] ;
[0017] in, For the single interpolated buoy data at time t, Represented as and Implicit functions between Indicates time; Represented as Buoy data at any given time Represented as Buoy data at any given time This is the error term.
[0018] According to the above technical solution, the interpolation process for the continuously lost buoy data is as follows:
[0019] by As input, continuously missing buoy data For the target label, in the artificial mask Calculate the loss within the region and use a hybrid loss function. Supervised network training; in this process, the network will learn from buoy data containing two consecutive missing blocks. To observe missing buoy data The mapping relationship aims to be found in artificial masks. Missing buoy channels in the interpolation area;
[0020] After training is completed, Substitute In this process, the goal is to interpolate the original mask using a trained network. The original missing buoy channels in the region are used to interpolate continuously missing buoy data.
[0021] in, This represents buoy data containing two consecutive missing blocks. Indicated as an artificial mask, For focus frequency loss, This is expressed as L1 norm loss. For masking operators, The interpolation result is for continuously missing buoy data. To train the network, For continuously missing buoy data, To optimize parameters, This is the complete buoy dataset.
[0022] The interpolation of the continuous data uses a self-supervised learning method, wherein... Focus frequency loss This allows the network to adaptively focus on frequency domain components, improving interpolation performance; the L1 norm loss is used here because it's uncertain whether the buoy dataset contains outliers, and the L1 norm loss is unaffected by outliers, where the focus frequency loss... L1 norm loss is existing technology and will not be described in detail here.
[0023] According to the above technical solution, the process of determining the training set for each bivariate is as follows:
[0024] The wave variable association database was searched to find the predicted variable. Related set of variables ;
[0025] Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate;
[0026] The first part of the buoy dataset corresponding to the bivariate As a bivariate training set;
[0027] The process for determining the validation sets for each bivariate is as follows:
[0028] The wave variable association database was searched to find the predicted variable. Related set of variables ;
[0029] Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate;
[0030] The latter part of the corresponding buoy dataset in the bivariate model. As a bivariate validation set; and the bivariate validation set and The corresponding buoy data are respectively placed into Validation set and Validation set; where, This indicates the search for and prediction variables in buoy dataset B. The number of variables with high correlation.
[0031] The wave variable association database consists of many variable association sets. The variable association sets are constructed based on the characteristics of wave generation. Specifically, wave variable d will have multiple variables associated with it, and all variables associated with variable d constitute a variable association set.
[0032] in, This indicates the search for and prediction variables in buoy dataset B. There are I highly correlated variables. The main characteristic of wave generation is that wave generation is closely related to wind speed. The generated waves and their periods satisfy the wave dispersion relationship. Therefore, the relationship between various variables (wind speed-wave height, period-wave height, etc.) can be considered. This relationship can also be given by the correlation between data. This relationship is obvious in this field, so it has not been elaborated in detail.
[0033] According to the above technical solution, the steps for establishing the bivariate LSTM model are as follows:
[0034] The training set of the two variables is an n-row, 2-column matrix, which is then input into the Long Short-Term Memory (LSTM) algorithm for training to construct a bivariate LSTM model.
[0035] The first column is represented as The buoy data, the second column represents the forecast variables. For buoy data, the LSTM model has two transmission states, one , and a ; Changes occur very slowly during online dissemination, representing a relatively stable form of information over a long period; while Changes occur rapidly during network propagation, representing short-term, localized information; each layer of the LSTM network needs to adapt to the input at the current moment. and the short-term memory of the previous moment To update the cellular state representing long-term memory The update is achieved through a three-gate structure, which includes a forget gate, a remember gate, and an output gate.
[0036] Based on the above technical solution, each The data from the validation set is input into the corresponding bivariate LSTM model to obtain the corresponding prediction variables. The forecast results;
[0037] The correlation coefficient, root mean square error, and mean absolute percentage error of the forecast results obtained from each bivariate model were calculated separately, and the correlation coefficient results were put into the newly constructed set. In the process, the root mean square error calculation results are placed into the newly constructed set. In the process, the mean absolute percentage error calculation results are placed into the newly constructed set. middle;
[0038] Use the if function to determine the set Maximum value, set and set The minimum value in the set is used to determine the set. Maximum value, set and set The bivariate LSTM model corresponding to the minimum value in the set is then added to the newly constructed set. middle;
[0039] Determine set The most frequently occurring bivariate LSTM model, and the use of the if function to determine the set. Is the bivariate LSTM model that appears most frequently unique?
[0040] If set If the bivariate LSTM model that appears most frequently is unique, then it proves that the bivariate LSTM model is the optimal prediction model.
[0041] If set If the most frequently occurring bivariate LSTM model is not unique, then the set will be... The bivariate LSTM model corresponding to the maximum value in the middle value is identified as the optimal prediction model;
[0042] Among them, set Maximum value, set and set The minimum value in the range is not unique.
[0043] In this process, m points are selected from the validation set data for K-hour prediction. That is, the first m data points in the validation set are taken to predict the wave height at time m+k-1, and so on, until the complete time is predicted. During this process, the wave height validation set data is not required to complete the prediction. In contrast, a similar univariate process requires wave height validation set data, so the univariate LSTM model actually implements post-prediction.
[0044] Among them, the correlation coefficient can reflect the degree of correlation between variables, the root mean square error characterizes the deviation between the simulation results and the measured values, and is more sensitive to extreme values. The mean absolute percentage error is expressed as a percentage to show the degree of deviation of the predicted value from the measured value. The methods of correlation coefficient, root mean square error and mean absolute percentage error are existing technologies and will not be described in detail here.
[0045] A wave data missing forecasting system based on a bivariate long short-term memory algorithm, characterized by: including
[0046] The data acquisition module is used to acquire the buoy dataset A.
[0047] The data interpolation module is used to interpolate single missing buoy data using regression models and to interpolate consecutively missing buoy data using machine learning methods, thereby forming a new buoy dataset B.
[0048] The correlation variable determination module is used to correlate ocean wave variables with a database and identify variables in buoy dataset B that are related to the predicted variables. Match the variables and construct a bivariate set V, and determine the training set and validation set for each bivariate in the bivariate set V.
[0049] The bivariate model construction module is used to input the training sets of each bivariate model into the Long Short-Term Memory algorithm for training, and obtain the corresponding bivariate LSTM model.
[0050] The optimal prediction model determination module inputs the validation set of each variable into the corresponding bivariate LSTM model to predict missing variables, compares and analyzes each prediction result with the corresponding validation set data, evaluates the prediction performance of each bivariate LSTM model to obtain the prediction result, and determines the optimal prediction model based on preset conditions and prediction results.
[0051] A wave data forecasting device based on a bivariate long short-term memory algorithm is characterized by comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the wave data forecasting method based on the bivariate long short-term memory algorithm.
[0052] This approach can also be applied to situations where a single variable in a buoy dataset has been missing for an extended period of time. However, this requires a year of continuous observation data, which can then be combined with relevant variables to form a bivariate model for training. Long-term predictions can then be made using the relevant variables.
[0053] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: For missing continuous buoy data, a self-supervised machine learning approach is used for buoy data imputation. This self-supervised network uses the missing buoy data as the target label, making it more applicable to real-world situations and resulting in imputed values closer to the true values. For single missing buoy data, a regression model is used for imputation, considering the correlation between the missing point and its two neighboring points, making the imputed buoy data more accurate. For buoy data with a single variable missing for a continuous long period, a bivariate short-time memory model is used. Variables highly correlated with the missing variable are used to form a bivariate model for machine learning. Finally, by inputting the values of the relevant variables n moments before the forecast time, the time change sequence of the missing variable is predicted, thus achieving the goal of quickly and accurately forecasting missing buoy data. Furthermore, compared with other forecasting models, this model does not require adding more influencing factors, reducing model complexity and computational load. Attached Figure Description
[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0055] Figure 1 This is a schematic diagram of the process for forecasting missing wave data based on a bivariate long short-term memory algorithm according to the present invention.
[0056] Figure 2 This is a diagram of the LSTM model structure;
[0057] Figure 3 This is a comparison chart of the 6-hour forecast results for buoy No. 46083. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0059] Please see Figures 1-3 The present invention provides a technical solution: a method for forecasting missing ocean wave data based on machine learning, comprising:
[0060] S1. Obtain the buoy dataset A;
[0061] S2. Use regression model to impute single missing buoy data, and use machine learning to impute continuously missing buoy data, thereby forming a new buoy dataset B;
[0062] The regression model is as follows:
[0063] ;
[0064] in, For the single interpolated buoy data at time t, Represented as and Implicit functions between Indicates time; Represented as Buoy data at any given time Represented as Buoy data at any given time This is the error term.
[0065] The interpolation process for continuously lost buoy data is as follows:
[0066] by As input, continuously missing buoy data For the target label, in the artificial mask Calculate the loss within the region and use a hybrid loss function. Supervise network training; after training is complete, Substitute In this process, interpolation of continuously lost buoy data is achieved;
[0067] in, This represents buoy data containing two consecutive missing blocks. Indicated as an artificial mask, For focus frequency loss, This is expressed as L1 norm loss. For masking operators, The interpolation result is for continuously missing buoy data. To train the network, For continuously missing buoy data, To optimize parameters, This is the complete buoy dataset.
[0068] The method used for interpolating continuous data is self-supervised learning, in which... Focus frequency loss This allows the network to adaptively focus on frequency domain components, improving interpolation performance; the L1 norm loss is used here because it's uncertain whether the buoy dataset contains outliers, and the L1 norm loss is unaffected by outliers, where the focus frequency loss... The L1 norm loss is existing technology and will not be described in detail here.
[0069] S3. Using the wave variable correlation database, find the correlation between the predicted variable and the wave variable in the buoy dataset B. Match the variables and construct a bivariate set V, and determine the training set and validation set for each bivariate in the bivariate set V.
[0070] The process of determining the training set for each bivariate variable is as follows:
[0071] The wave variable association database was searched to find the predicted variable. Related set of variables ;
[0072] Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate;
[0073] The first part of the buoy dataset corresponding to the bivariate As a bivariate training set;
[0074] The process for determining the validation sets for each bivariate is as follows:
[0075] The wave variable association database was searched to find the predicted variable. Related set of variables ;
[0076] Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate;
[0077] The latter part of the corresponding buoy dataset in the bivariate model. As a bivariate validation set; and the bivariate validation set and The corresponding buoy data are respectively placed into Validation set and Validation set; where, This indicates the search for and prediction variables in buoy dataset B. The number of highly correlated variables, such as those found in a wave variable association database that are highly correlated with forecast variables. Set of variables associated with wave height ; and the variables and variable sets in the buoy dataset B. The variables in the data were matched, and the variables that were successfully matched were wind speed and period.
[0078] S4. Substitute the training sets of each bivariate model into the Long Short-Term Memory (LSTM) algorithm for training to obtain the corresponding bivariate LSTM model. The steps for establishing the bivariate LSTM model are as follows:
[0079] The training sets of the two variables are input into the Long Short-Term Memory algorithm for training to construct a bivariate LSTM model;
[0080] The bivariate training set is an n x 2 matrix, and the LSTM model has two transmission states, one... , and a ; Changes occur very slowly during online dissemination, representing a relatively stable form of information over a long period; while Changes occur rapidly during network propagation, representing short-term, localized information; each layer of the LSTM network needs to adapt to the input at the current moment. and the short-term memory of the previous moment To update the cellular state representing long-term memory The update is achieved through a three-gate structure.
[0081] S5. Input the validation set of each variable into the corresponding bivariate LSTM model to predict the missing variables, and compare and analyze each prediction result with the corresponding validation set data to evaluate the prediction performance of each bivariate LSTM model and obtain the prediction results. Based on the preset conditions and prediction results, determine the optimal prediction model, specifically as follows:
[0082] Each The data from the validation set is input into the corresponding bivariate LSTM model to obtain the corresponding prediction variables. The forecast results;
[0083] The correlation coefficient, root mean square error, and mean absolute percentage error of the forecast results obtained from each bivariate model were calculated separately, and the correlation coefficient results were put into the newly constructed set. In the process, the root mean square error calculation results are placed into the newly constructed set. In the process, the mean absolute percentage error calculation results are placed into the newly constructed set. middle;
[0084] Use the if function to determine the set Maximum value, set and set The minimum value in the set is used to determine the set. Maximum value, set and set The bivariate LSTM model corresponding to the minimum value in the set is then added to the newly constructed set. middle;
[0085] Determine set The most frequently occurring bivariate LSTM model, and the use of the if function to determine the set. Is the bivariate LSTM model that appears most frequently unique?
[0086] If set If the bivariate LSTM model that appears most frequently is unique, then it proves that the bivariate LSTM model is the optimal prediction model.
[0087] If set If the most frequently occurring bivariate LSTM model is not unique, then the set will be... The bivariate LSTM model corresponding to the maximum value in the middle value is identified as the optimal prediction model;
[0088] Among them, set Maximum value, set and set The minimum value in the range is not unique.
[0089] In this process, m points are selected from the validation set data for K-hour prediction. That is, the first m data points in the validation set are taken to predict the wave height at time m+k-1, and so on, until the complete time is predicted. During this process, the wave height validation set data is not required to complete the prediction. A similar univariate process, however, requires wave height validation set data. Therefore, the univariate LSTM model actually performs a post-prediction; for example, with m=6 and k=1, the wave height values are predicted for 6 times.
[0090] Example:
[0091] This invention selects wave height data of buoy 46083 in the North Pacific Ocean from 2019 to 2020, with a time interval of 1 hour, from 00:00 on January 1, 2019 to 23:00 on December 31, 2020. The buoy data set contains wind speed data, periodic data, etc.
[0092] Machine learning methods are used to impute continuously missing buoy data in buoy dataset A, and regression models are used to impute individually missing buoy data in buoy dataset A, thus forming a new buoy dataset B.
[0093] Among the forecast variables For wave height, a query was performed in the wave variable association database to find a value related to the forecast variable. Related set of variables ; and the variables and variable sets in the buoy dataset B. The variables in the model are matched. If the matched variables are wind speed and period, then wind speed-wave height and period-wave height bivariate models are established. The wind speed-wave height bivariate model uses the wind speed and wave height time series observed in 2019 as the training set and the wind speed and wave height time series observed in 2020 as the validation set. The period-wave height bivariate model uses the period and wave height time series observed in 2019 as the training set and the period and wave height time series observed in 2020 as the validation set.
[0094] The period-wave height training set data (n×2 matrix) is substituted into the bivariate LSTM model for training; the wind speed validation set data is input into the trained wind speed-wave height model, and the wave height validation set data is input into the trained period-wave height model. Finally, the wave height prediction result is obtained. In this process, m points are selected from the validation set data for K-hour prediction. That is, the first m data points in the validation set are taken to predict the wave height value at time m+k-1, and so on, until the complete time is predicted. In this process, the wave height validation set data does not need to be input to complete the prediction. A similar univariate process requires the wave height validation set data, so the univariate LSTM model actually implements post-prediction. In this example, m=6, k=1, 3, 6, 12.
[0095] The prediction results obtained from the bivariate LSTM model are compared with the univariate prediction results to obtain the following: Figure 3 The comparison chart of the 6-hour forecast results for buoy 46083 is shown. The calculation results of the three forecast models for buoy 46083 are obtained by using the correlation coefficient, root mean square error and mean absolute percentage error method as shown in Table 1. This further verifies the feasibility and superiority of the method in handling missing data, and then obtains the optimal prediction model - the period-wave height bivariate LSTM model.
[0096] Table 1
[0097]
[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0099] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for forecasting missing ocean wave data based on a bivariate long short-term memory algorithm, characterized in that: include: Obtain buoy dataset A; The regression model is used to impute single missing buoy data, and the machine learning method is used to impute continuously missing buoy data, thus forming a new buoy dataset B; By using the ocean wave variable to correlate with the database, find the correlation between the predicted variable and the ocean wave variable in the buoy dataset B. Match the variables and construct a bivariate set V, and determine the training set and validation set for each bivariate in the bivariate set V. The training sets of each bivariate model are substituted into the Long Short-Term Memory algorithm for training to obtain the corresponding bivariate LSTM model. The validation sets of each variable are input into the corresponding bivariate LSTM model to predict missing variables. The prediction results are compared and analyzed with the corresponding validation set data to evaluate the prediction performance of each bivariate LSTM model and obtain the prediction results. The optimal prediction model is determined based on the preset conditions and prediction results. The process for determining the training sets for each bivariate variable is as follows: The wave variable association database was searched to find the predicted variable. Related set of variables ; Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate; The first part of the buoy dataset corresponding to the bivariate As a bivariate training set; The process for determining the validation sets for each bivariate is as follows: The wave variable association database was searched to find the predicted variable. Related set of variables ; Combine the variables in the buoy dataset B with the variable set. Match the variables in the data to construct a set. In the middle, the predicted variables will be With sets Each element in the set is combined to form a bivariate set. In, the set Represents the buoy dataset B and the variable set. The set of variables that match in Represented as the i-th variable in the buoy dataset B that is related to the prediction variable. Related variables, Represented as the I-th bivariate; The latter part of the corresponding buoy dataset in the bivariate model. As a bivariate validation set; And bivariate validation set and The corresponding buoy data are respectively placed into Validation set and Validation set; where, This indicates the search for and prediction variables in buoy dataset B. The number of variables with high correlation.
2. The method for forecasting missing wave data based on a bivariate long short-term memory algorithm according to claim 1, characterized in that: The regression model is as follows: ； in, For the single interpolated buoy data at time t, Represented as and Implicit functions between Indicates time; Represented as Buoy data at any given time Represented as Buoy data at any given time This is the error term.
3. The method for forecasting missing wave data based on a bivariate long short-term memory algorithm according to claim 1, characterized in that: The interpolation process for the continuously lost buoy data is as follows: by As input, continuously missing buoy data For the target label, in the artificial mask Calculate the loss within the region and use a hybrid loss function. Supervise network training; after training is complete, Substitute In this process, the interpolation of continuously missing buoy data is achieved; in, This represents buoy data containing two consecutive missing blocks. Indicated as an artificial mask, For focus frequency loss, This is expressed as L1 norm loss. For masking operators, The interpolation result is for continuously missing buoy data. To train the network, For continuously missing buoy data, To optimize parameters, This is the complete buoy dataset.
4. The method for forecasting missing wave data based on a bivariate long short-term memory algorithm according to claim 1, characterized in that: The training set for the bivariate dataset is an n x 2 matrix; The first column is represented as The buoy data, the second column represents the forecast variables. Buoy data.
5. The method for forecasting missing wave data based on a bivariate long short-term memory algorithm according to claim 1, characterized in that: The steps for determining the optimal prediction model include: Each The data from the validation set is input into the corresponding bivariate LSTM model to obtain the corresponding prediction variables. The forecast results; The correlation coefficient, root mean square error, and mean absolute percentage error of the forecast results obtained from each bivariate model were calculated separately, and the correlation coefficient results were put into the newly constructed set. In the process, the root mean square error calculation results are placed into the newly constructed set. In the process, the mean absolute percentage error calculation results are placed into the newly constructed set. middle; Use the if function to determine the set Maximum value, set and set The minimum value in the set is used to determine the set. Maximum value, set and set The bivariate LSTM model corresponding to the minimum value in the set is then added to the newly constructed set. middle; Determine set The most frequently occurring bivariate LSTM model, and the use of the if function to determine the set. Is the bivariate LSTM model that appears most frequently unique? If set If the bivariate LSTM model that appears most frequently is unique, then it proves that the bivariate LSTM model is the optimal prediction model. If set If the most frequently occurring bivariate LSTM model is not unique, then the set will be... The bivariate LSTM model corresponding to the maximum value in the middle value is identified as the optimal prediction model; Among them, set Maximum value, set and set The minimum value in the range is not unique.
6. A system for implementing the wave missing data forecasting method based on the bivariate long short-term memory algorithm as described in claim 1, characterized in that: include The data acquisition module is used to acquire the buoy dataset A. The data interpolation module is used to interpolate continuously lost buoy data using machine learning methods and to interpolate individually lost buoy data using regression models, thereby forming a new buoy dataset B. The correlation variable determination module is used to correlate ocean wave variables with a database and identify variables in buoy dataset B that are related to the predicted variables. Match the variables and construct the bivariate set V, while determining the training and validation sets for each bivariate; The bivariate model construction module is used to substitute each bivariate training set into the Long Short-Term Memory algorithm for training, and then construct the corresponding bivariate LSTM model. The optimal prediction model determination module is used to input the validation set of each variable into the corresponding bivariate LSTM model to predict the missing variables, and compare and analyze each prediction result with the corresponding validation set data to evaluate the prediction performance of each bivariate LSTM model and obtain the optimal prediction model.
7. A wave data missing prediction device based on a bivariate long short-term memory algorithm, characterized in that: It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the wave missing data forecasting method based on the bivariate long short-term memory algorithm as described in claim 1.