Pressure sensing inverse echo measurement LSTM short-time sea wave significant wave height prediction method

Abstract

The invention relates to a pressure sensing inverse echo measurement LSTM short-time sea wave significant wave height prediction method, and belongs to the technical field of marine environment parameter inversion and marine dynamics observation. The method comprises the following steps: firstly, preprocessing multi-source observation data, and matching and integrating data from different sources; establishing a model, and improving the performance of the model by adjusting the structure and parameters of the model; and finally, evaluating the precision of the model, and verifying the precision and stability of the model in a complex environment. According to the method provided by the invention, a plurality of long and short-term memory network layers are combined with a plurality of full-connection layers to realize hierarchical fusion of time features, so that the prediction precision of the significant wave height of the short-term sea wave is improved, and important data support is provided for the directions of ocean dynamics, environmental evaluation and the like.

Description

Technical Field

[0001] This invention belongs to the field of marine exploration technology and sea surface parameter inversion, specifically involving the pressure-sensing inverse echo wave measurement LSTM short-time effective wave height prediction method. Background Technology

[0002] Significant wave height is a crucial parameter characterizing sea state strength and ocean dynamics, significantly impacting maritime safety, risk assessment for offshore operations, and marine weather forecasting. Accurate forecasting of significant wave height is fundamental for climate model validation, ocean circulation research, and marine ecosystem simulation.

[0003] In recent years, PIES-based inversion techniques have attracted widespread attention. PIES inverts the overall density structure and dynamic changes of a sea column by emitting acoustic pulses and measuring their round-trip time (tau). After dispersion analysis, a dynamic relationship can be established between the tau sequence and sea surface waves, thereby obtaining sea surface parameters such as the significant wave height. This method has advantages such as not relying on floating platforms, being unaffected by severe sea conditions, long-term stable operation, and suitability for deep-sea environments, and therefore has gradually become an important means of wave monitoring in deep-sea areas.

[0004] With the development of observation technology, more and more studies are attempting to extract temporal features from multi-source observations to improve the accuracy of short-term wave prediction. However, existing research mainly focuses on wave sequences directly measured by buoys, with less attention paid to short-term prediction methods based on PIES inversion of significant wave height. Existing methods often struggle to effectively integrate the coupling characteristics between Tau inversion of significant wave height and multiple meteorological observations, resulting in insufficient response to rapid changes in future waves and limitations in prediction performance under sudden sea state conditions such as rapid changes in wind and waves.

[0005] Current ocean wave prediction technologies suffer from limitations on three main levels: First, at the data acquisition level, while traditional buoy observations can provide direct measurement data, their deployment costs are high, maintenance is difficult, and they are easily damaged in harsh sea conditions, resulting in inconsistent data continuity. In contrast, while satellite remote sensing technology offers wide coverage, its temporal resolution is limited, making it difficult to meet the demand for high-frequency data in short-term predictions. Although PIES technology compensates for these shortcomings to some extent, its data inversion accuracy is significantly affected by the complexity of the marine environment.

[0006] Secondly, at the model building level, traditional statistical models such as the Autoregressive Integral Moving Average (ARIMA) model, while computationally simple, struggle to capture the nonlinear characteristics of wave evolution. Physical models such as WaveWatch III, although based on fluid dynamics equations, have high computational complexity and are extremely sensitive to initial and boundary conditions, thus limiting their practical application.

[0007] Third, at the level of multi-source data fusion, existing methods often simply use linear weighting or interpolation to integrate observation data from different sources, failing to fully consider the spatiotemporal correlations and physical mechanisms between various parameters. In particular, there is a lack of effective feature extraction and fusion mechanisms in the collaborative prediction of wind speed, wave period, and wave height.

[0008] With the development of deep learning technology, time series models such as LSTM are increasingly being applied in the fields of meteorology and oceanography. However, research on combining LSTM with PIES inversion data is still in its early stages, especially in short-term forecasting, where a systematic technical solution has not yet been developed. This invention presents an innovative solution to address the shortcomings of existing methods against this technological backdrop.

[0009] From an international research perspective, wave prediction technology is developing in three directions: multi-source data fusion, artificial intelligence-driven approaches, and operational applications. Regarding data fusion, the EU's Copernicus marine environmental monitoring service has begun experimenting with assimilating multi-source observation data from satellites, buoys, and radar with numerical models to improve prediction accuracy. In terms of artificial intelligence applications, the US National Oceanic and Atmospheric Administration (NOAA) is exploring the use of deep learning models for extreme sea state warnings. Regarding operational applications, the Japan Meteorological Agency has established an operational wave prediction system based on machine learning.

[0010] However, these advancements still face numerous challenges. Particularly in the application of PIES data, the limited number of deployed devices and insufficient data samples make model training difficult; simultaneously, the significant differences in marine environments across different sea areas severely test the model's generalization ability. This invention effectively solves these technical problems through innovative model structures and training strategies, providing a new technical approach for short-term wave prediction. Summary of the Invention

[0011] The purpose of this invention is to provide a pressure-sensing inverse echo wave measurement LSTM method for predicting short-time effective wave height.

[0012] To achieve the above objectives, this invention employs the following technical solution: a pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method, comprising the following steps:

[0013] S1. Acquire multi-source sea surface observation data with an hourly observation frequency, including: multi-source observation average wind speed (WSPD), instantaneous wind speed (GST), dominant wave period (DPD), average wave period (APD), and effective wave height (WVHT) time series calculated based on the dispersion of PIES high-frequency sampling echo time (tau).

[0014] S2: Match the spatiotemporal information with the corresponding multi-source observation data and the WVHT obtained by PIES inversion. The sample includes 5 parameters: WSPD, GST, DPD, APD and WVHT, and samples with null values ​​are deleted.

[0015] S3: Arrange the samples in chronological order, divide all samples into training set, test set and validation set in a ratio of 5:3:2, and select WSPD, GST, DPD, APD and WVHT time series of 12 hours before the measurement time as 5 features as input to the model, and WVHT data at a future time as label as output to the model.

[0016] S4: Build the model and use the training set data to input into the model for training;

[0017] S5: Using the trained model, use the test dataset to adjust the model parameters until the model accuracy reaches the optimal level, that is, to obtain the optimal model;

[0018] S6: Use the validation dataset to evaluate the accuracy of the optimal model and obtain the accuracy of the model in predicting the significant wave height of ocean waves;

[0019] Preferably, in step S1, in order to construct a high temporal resolution short-time effective wave height prediction model, the acquisition frequency of multi-source observation data used for model construction should be greater than or equal to once per hour.

[0020] Preferably, in step S1, WVHT is calculated based on the dispersion of high-frequency sampling echo time (tau). Specifically, it is known that there is a linear relationship between the dispersion σ of tau and WVHT. For the tau in the acquired PIES observation data, a suitable time interval is selected, the dispersion of the dataset is calculated, and WVHT is obtained by linear fitting.

[0021] WVHT=aσ+b

[0022] Preferably, in step S2, the matching of multi-source observation data and WVHT obtained by PIES inversion is performed using spatiotemporal information. The main principle of spatiotemporal proximity is used to first find the WSPD, GST, DPD, and APD observation data that are in the same time period as each WVHT, and then use spatial distance to find the nearest grid as its corresponding WSPD, GST, DPD, and APD observation data.

[0023] Preferably, in step S2, if there are null values ​​for parameters in each sample, the entire sample is deleted.

[0024] Preferably, in step S4, the structure of the model used in this invention is as follows: Figure 2 As shown, the model consists of m LSTM layers and n fully connected layers connected in series. The ReLU function is used as the activation function between each layer, and the data is output after passing through the final fully connected layer. The LSTM model has strong temporal memory capabilities and can efficiently capture the spatiotemporal nonlinear coupling relationship of the significant wave height.

[0025] The expression for the ReLU function is:

[0026] ReLU(x) = max(0, w T +b)

[0027] Where x is the input vector, w is the weight vector, and b is the bias term.

[0028] The loss function uses the mean squared error function:

[0029]

[0030] Among them, Y i For the true value, These are predicted values.

[0031] All weights ω in each layer k Update using gradient descent:

[0032]

[0033] Where η is the learning rate.

[0034] Preferably, in step S4, the values ​​of m and n are adjusted using a grid search method. Specifically, the following steps are taken: first, determine the range of values ​​for m and n, and build a network structure based on different values ​​of m and n; then, train each model using the same training strategy, and calculate its RMSE on the test set as an evaluation index of the model's performance; finally, select the parameter combination with the smallest RMSE on the validation set as the optimal structure of the model.

[0035] Preferably, in step S4, the distribution of the number of neurons in each layer of the model is an expansion-contraction structure, that is, the number of neurons first increases and then decreases with increasing depth. By gradually increasing the number of neurons, the model can learn richer features. Subsequently, by gradually reducing the number of neurons, the model can compress and simplify features, effectively preventing overfitting. The number of neurons in each layer can be set according to the complexity of the samples.

[0036] Preferably, in step S5, the learning rate (lr) and batch size (batch_size) are adjusted using a grid search method. Specifically, the range of values ​​for lr and batch_size is first determined, then different parameter combinations are used to train the model obtained in step S4, and the RMSE on the test set is calculated as an evaluation index of the model's performance. Finally, the parameter combination with the smallest RMSE on the validation set is selected as the optimal parameters for the model.

[0037] Preferably, in step S6, the optimal model is applied to the validation set, and the calculated RMSE and R... 2 To verify the performance of the model. RMSE and R 2 The expression is as follows:

[0038]

[0039] In the formula, n is the sample size, X i Y is the independent variable and Y is the dependent variable. i f(X) is the actual value. i The value is the predicted value. A smaller RMSE value indicates a smaller overall model error, higher accuracy, and more accurate predictions. 2 A higher numerical value indicates a better model fit and stronger interpretability.

[0040] This invention provides a pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method, which has the following advantages:

[0041] Deep learning models based on the LSTM architecture possess unique temporal memory capabilities, effectively capturing long-term dependencies in the evolution of ocean waves. Compared to traditional machine learning models, the model proposed in this invention exhibits significant advantages in temporal feature extraction. Specifically, these advantages are manifested in the following aspects:

[0042] First, the gating mechanism of LSTM (input gate, forget gate, output gate) can selectively memorize important temporal features, effectively solving the gradient vanishing problem in traditional RNN models. In wave prediction scenarios, the model can memorize sea state change patterns from several hours or even tens of hours ago, thereby more accurately predicting future wave height trends.

[0043] Secondly, the model employs a multi-layered LSTM structure to achieve hierarchical extraction of temporal features. The bottom-layer LSTM captures short-term fluctuation features, such as sudden changes in wave height caused by gusts; the middle-layer LSTM extracts periodic variation patterns, such as tidal influences; and the top-layer LSTM learns long-term evolution trends, such as seasonal variation features. This hierarchical feature extraction mechanism enables the model to adapt to the prediction needs of different time scales.

[0044] Experimental results show that, in the 2023 West Atlantic measured data validation, the RMSE of the model of this invention for 1-hour predictions reached 0.1216, R0.05 2 The RMS value was 0.9747, significantly better than the traditional ARIMA model (RMSE = 0.2154, R...). 2 =0.8923) and SVM model (RMSE=0.1987, R 2 =0.9015).

[0045] The model's modular design gives it good adaptability and scalability. By adjusting the network depth (number of LSTM layers m) and width (number of fully connected layers n), it can flexibly adapt to the prediction needs of different sea areas and seasons. Specific advantages include:

[0046] Cross-ocean applicability: Tests in three different sea areas—the South China Sea, the East China Sea, and the western Atlantic Ocean—show that the model can achieve good predictive results simply by adjusting parameters while keeping the core structure unchanged. In the South China Sea, where tropical cyclones are influential, the model's prediction lead time for sudden large waves is advanced to 3-5 hours; in the East China Sea, where monsoons are influential, the predictive correlation for periodic waves reaches over 0.95.

[0047] Multi-timescale adaptability: The model supports wave height prediction at multiple timescales, including 1 hour, 3 hours, and 6 hours. By adjusting the input sequence length and output layer structure, it can meet the needs of different application scenarios. For example, shipping safety requires short-term predictions within 1 hour, offshore engineering operations require medium-term predictions of 3-6 hours, and marine disaster early warning requires trend predictions over a longer period.

[0048] Data Missing Data Tolerance: The model employs an expanding-contracting neuron distribution structure, enabling effective predictions even when some observational data is missing, based on existing features. Tests show that when 30% of wind speed data is missing, the model's prediction accuracy decreases by only 8.7%, significantly better than the 23.5% accuracy loss of traditional methods.

[0049] From a computational efficiency perspective, the model uses GPU-accelerated training, with a single training session taking no more than 30 minutes on an NVIDIA V100 graphics card, while traditional numerical models often require several hours of computation. Furthermore, the model supports online learning, allowing for continuous optimization of predictive performance as new data is acquired.

[0050] From an engineering application perspective, the model can be deployed on edge computing devices to provide real-time wave prediction services for offshore platforms. Combined with satellite communication systems, it enables remote data transmission and model updates, meeting the operational needs of deep-sea areas.

[0051] The technological innovation of this invention is not only reflected in the model algorithm level, but also in pioneering a new technological path that combines PIES data with deep learning. This innovation has significant implications for the development of marine observation technology: First, it provides technical support for the in-depth development and utilization of PIES data, enhancing the application value of equipment observation data; second, it promotes the application of artificial intelligence technology in the marine field, providing key technical support for the construction of smart oceans; finally, the resulting technical solution can be replicated and extended to the prediction of other marine parameters, such as ocean currents, water temperature, and salinity, showing broad application prospects. Attached Figure Description

[0052] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0053] Figure 1 This is a flowchart of the short-time effective wave height prediction method of the present invention;

[0054] Figure 2 This is a model structure diagram for the short-time effective wave height prediction method.

[0055] Figure 3 This is a diagram of the model structure in the embodiment;

[0056] Figure 4 This is a comparison chart of the predicted and actual values ​​of short-term sea wave significant wave height. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0059] like Figure 1 As shown, a method for predicting the significant wave height of short-time ocean waves based on pressure-sensing inverse echo sounder wave measurement and long short-time memory network includes the following steps:

[0060] S1. Acquire multi-source sea surface observation data with an hourly observation frequency, including: multi-source observation average wind speed (WSPD), instantaneous wind speed (GST), dominant wave period (DPD), average wave period (APD), and effective wave height (WVHT) time series calculated based on the dispersion of PIES high-frequency sampling echo time (tau).

[0061] S2: Match the spatiotemporal information with the corresponding multi-source observation data and the WVHT obtained by PIES inversion. The sample includes 5 parameters: WSPD, GST, DPD, APD and WVHT, and samples with null values ​​are deleted.

[0062] S3: Arrange the samples in chronological order, divide all samples into training set, test set and validation set in a ratio of 5:3:2, and select WSPD, GST, DPD, APD and WVHT time series of 12 hours before the measurement time as 5 features as input to the model, and WVHT data at a future time as label as output to the model.

[0063] S4: Build the model and use the training set data to input into the model for training;

[0064] S5: Using the trained model, use the test dataset to adjust the model parameters until the model accuracy reaches the optimal level, that is, to obtain the optimal model;

[0065] S6: Use the validation dataset to evaluate the accuracy of the optimal model and obtain the accuracy of the model in predicting the significant wave height of ocean waves;

[0066] Example:

[0067] S1. Acquire multi-source sea surface observation data with an hourly observation frequency, including: multi-source observed mean wind speed (WSPD), instantaneous wind speed (GST), dominant wave period (DPD), mean wave period (APD), and significant wave height (WVHT) time series calculated based on the dispersion of PIES high-frequency sampling echo time (tau). The observation data area in this embodiment is located in the PIES deployment area in the western Atlantic Ocean, with a time range from January 1, 2023 to December 31, 2024. The WSPD, GST, DPD, and APD data in this embodiment are obtained from buoy observation data from the National Buoy Data Center (NDBC). WSPD is the average wind speed within 8 minutes of the measurement time, and APD is the average wave period of all waves within 20 minutes of the measurement time. For the tau in the acquired PIES observation data, the time range is selected as 24 hours, the dispersion σ is calculated, and WVHT is calculated through linear fitting.

[0068] S2. Match the spatiotemporal information with the corresponding multi-source observation data and WVHT. The sample contains 5 parameters: WSPD, GST, DPD, APD and WVHT. Remove the samples with null values ​​to obtain the dataset. The total number of samples in the dataset is 27,596.

[0069] S3. Arrange the samples in chronological order and divide all samples into training set, test set and validation set in a ratio of 5:3:2. Select WSPD, GST, DPD, APD and WVHT time series of 12 hours before the measurement time as 5 features as input to the model. Use WVHT data of the next 1 hour, 3 hours and 6 hours as labels as output to the model.

[0070] S4. Based on the LSTM framework, a model is constructed using a sequential concatenation method, and the model is trained using a training dataset. The parameters m and n are adjusted using a grid search method. Specifically, based on the characteristics of the dataset, the range of m is selected as {1,2,3,4,5}, and the range of n is also selected as {1,2,3,4,5}, resulting in 25 possible model structures. The distribution of the number of neurons in each layer follows an expanding-contracting structure, meaning that the number of neurons first increases and then decreases with increasing depth. Each epoch is set to 2000, and the Adam optimizer is used to train each model separately. Finally, m=4 and n=3 are determined to be the optimal parameter combination. The established LSTM model structure is as follows: Figure 3 As shown, after the data is input through one input layer, it passes through four LSTM layers with 32, 64, 128 and 256 LSTM units respectively, three fully connected layers with 128, 64 and 32 neurons respectively, and finally outputs the wave height prediction result through an output layer with one neuron. The optimizer is Adam and the loss function is MSE.

[0071] S5. Using the trained model and the test dataset, perform model parameter tuning using the network search method until the model accuracy reaches its optimal level, i.e., obtain the optimal model. Specifically, select the batch_size range as {32, 64, 128, 256, 512, 1024, 2048} and the lr range as {10-2, 10-3, 10-4}, forming 21 parameter combinations. Set the epoch to 2000 for each combination. Use the Adam optimizer to train the model obtained in step S4. Finally, determine that batch_size = 1024 and lr = 10-4 is the optimal parameter combination, thus obtaining the optimal model.

[0072] S6. Evaluate the accuracy of the optimal model using the validation dataset, calculate RMSE and R², and base the predictions on the validation set as follows: Figure 4 As shown.

[0073] The effects of the present invention will be described below with reference to the accompanying drawings. To visually demonstrate the accuracy of the model calculation results, a scatter plot is drawn with the actual values ​​on the x-axis and the predicted values ​​on the y-axis, resulting in... Figure 4 The results shown indicate that, based on the model calculations, the RMSE value for the 1-hour prediction is 0.1216, and the R² value is 0.9747; the RMSE value for the 3-hour prediction is 0.1947, and the R² value is 0.9350; and the RMSE value for the 6-hour prediction is 0.3160, and the R² value is 0.8258. The scattered points in the figure are mainly distributed around the straight line y = x. The results demonstrate that the model in this invention can efficiently capture the temporal nonlinear coupling relationship of the significant wave height, exhibiting suitability for short-term significant wave height prediction and possessing high application value and promising prospects for wider adoption.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method, characterized in that, Includes the following steps: S1. Acquire multi-source sea surface observation data with an hourly observation frequency, including: multi-source observation average wind speed (WSPD), instantaneous wind speed (GST), dominant wave period (DPD), average wave period (APD), and significant wave height (WVHT) time series calculated based on the dispersion of PIES high-frequency sampling echo time for 12 hours prior to the measurement time; S2: Match the spatiotemporal information with the corresponding multi-source observation data and the WVHT obtained by PIES inversion. The sample includes 5 parameters: WSPD, GST, DPD, APD and WVHT, and samples with null values ​​are deleted. S3: Arrange the samples in chronological order, divide all samples into training set, test set and validation set in a ratio of 5:3:2, and select WSPD, GST, DPD, APD and WVHT time series of 12 hours before the measurement time as 5 features as input to the model, and WVHT at a future time as label as output to the model. S4: Build the model and use the training set data to input into the model for training; S5: Using the trained model, use the test dataset to adjust the model parameters until the model accuracy reaches the optimal level, that is, to obtain the optimal model; S6: Use the validation dataset to evaluate the accuracy of the optimal model and obtain the accuracy of the model in predicting the significant wave height of ocean waves.

2. The method according to claim 1, characterized in that, In step S4, the model consists of m sets of LSTM layers and n fully connected layers connected in series, with ReLU function used as the activation function between each layer.

3. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 2, characterized in that, m represents the number of LSTM layers, which affects the network depth and the ability to extract features from the dataset. The more complex the features of the dataset samples, the larger the value of m. n represents the number of fully connected layers, which affects the network's final mapping ability. The more features predicted, the larger the value of n.

4. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 3, characterized in that, The values ​​of m and n are adjusted using a grid search method. The specific steps are as follows: First, determine the range of values ​​for m and n, and build network structures based on different values ​​of m and n; then, train each model using the same training strategy, and calculate its RMSE on the test set as an evaluation metric for model performance; finally, select the parameter combination with the smallest RMSE on the validation set as the optimal model structure.

5. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 1, characterized in that, In step S1, WVHT is calculated based on the dispersion of the high-frequency sampling echo time tau, specifically satisfying the relationship: WVHT=aσ+b, where σ is the dispersion of tau, and 2 and b are parameters obtained through linear fitting.

6. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 1, characterized in that, In step S2, the principle of spatiotemporal proximity is used to match data. First, the WSPD, GST, DPD and APD observation data in the same time period as each WVHT are found, and the nearest grid is found as the corresponding observation data using spatial distance.

7. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 1, characterized in that, In step S4, the distribution of the number of neurons in each layer of the model is an expansion-contraction structure, that is, the number of neurons first increases and then decreases with increasing depth.

8. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 1, characterized in that, In step S5, the learning rate and batch size are adjusted using a grid search method, and the optimal parameters are determined by the RMSE performance of different parameter combinations on the test set.

9. The pressure-sensing inverse echo wave measurement LSTM short-time significant wave height prediction method according to claim 1, characterized in that, Step S6 uses RMSE and R 2 The calculation formulas for the model evaluation metrics are as follows:

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1-9.

Images