A method for predicting ocean significant wave height based on deep learning model of spatiotemporal series

Abstract

The present invention provides a method for predicting significant wave height based on a deep learning model of spatiotemporal sequences, which relates to the technical field of significant wave height prediction. The method specifically comprises the following steps: obtaining historical wind speed data and historical significant wave height data from a test area; quality screening the acquired data, and then performing spatiotemporal registration to align the historical wind speed data and historical significant wave height data with a one-hour time offset in the temporal dimension; using a sliding window algorithm to generate spatiotemporal sequence pairs and construct a training set; constructing a deep learning model for spatiotemporal sequences, inputting the training set into the deep learning model for iterative training, and obtaining a significant wave height prediction model based on the deep learning model for spatiotemporal sequences; and inputting the acquired future wind speed data into the prediction model based on an externally provided numerical forecast wind field to obtain a forecast result. The technical solution of the present invention overcomes the low accuracy and reliability of wave forecasts in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean wave significant wave height prediction, and in particular to an ocean wave significant wave height prediction method based on a spatiotemporal sequence deep learning model. Background Art

[0002] Ocean waves are undulations caused by wind acting on the sea surface, encompassing wind waves and swells. Their formation and development are influenced by a variety of factors, including wind direction, wind speed, and topography. Wave heights range from a few centimeters to over 20 meters, and even reach over 30 meters under unusual terrain conditions. Significant wave height is a crucial parameter for measuring ocean conditions and is crucial for planning and safety assessments of marine engineering and related activities. Accurate ocean wave forecasts are crucial for addressing climate change, ensuring navigation safety, and protecting coastal areas.

[0003] Numerical simulation is the primary method for ocean wave forecasting in existing technologies. While traditional empirical statistical methods offer rapid computational speed, they fail to accurately capture the complex characteristics of ocean waves, particularly swells, due to model simplifications and assumptions, resulting in significant deviations in forecast results. On the other hand, numerical wave models based on physical laws can simulate wave processes more accurately, but they are computationally intensive, slow, and can lead to cumulative errors over long forecast periods, compromising forecast accuracy.

[0004] To improve the performance of ocean wave simulation and forecasting, researchers are constantly innovating and refining statistical and numerical models. However, current technologies still have significant limitations when it comes to rapidly and accurately forecasting ocean waves. With the rapid increase in observational and simulated data, we are entering an era of big data. New methods are urgently needed to effectively process this data and delve deeper into the physical processes and complex relationships underlying ocean waves, thereby improving the accuracy and reliability of ocean wave forecasts.

[0005] Therefore, there is a need for a method for predicting significant wave height based on a spatiotemporal sequence deep learning model to improve the accuracy and reliability of wave forecasts. Summary of the Invention

[0006] The main purpose of the present invention is to provide a method for predicting the significant wave height of waves based on a deep learning model of spatiotemporal sequences, so as to solve the problems of low accuracy and reliability of wave forecasting in the prior art.

[0007] To achieve the above objectives, the present invention provides a method for predicting significant wave height based on a deep learning model of spatiotemporal sequences, which specifically includes the following steps:

[0008] S1, obtain historical wind speed data and historical wave significant wave height data of the test area.

[0009] S2, the acquired historical wind speed data and historical wave significant wave height data are quality screened, and then spatiotemporal registration is performed to align the historical wind speed data and historical wave significant wave height data with a 1-hour offset in the time dimension.

[0010] S3, uses the sliding window algorithm to generate spatiotemporal sequence pairs and construct a training set.

[0011] S4 builds a spatiotemporal sequence deep learning model, which includes an encoding layer and a prediction layer. The encoding layer is responsible for encoding and memorizing historical wave conditions, while the prediction layer is responsible for forecasting waves based on historical conditions and future wind speeds.

[0012] S5, inputting the training set constructed in step S3 into the spatiotemporal sequence deep learning model for iterative training, and obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model.

[0013] S6, based on the numerical forecast wind field provided by the outside, inputs the acquired future wind speed data into the trained ocean wave significant wave height prediction model based on the spatiotemporal sequence deep learning model to obtain the forecast results.

[0014] Furthermore, step S2 specifically includes the following steps:

[0015] S2.1, perform quality screening on historical wind speed data and historical wave significant wave height data, including missing value processing and land marker masking.

[0016] S2.2, perform spatiotemporal registration of historical wind speed data and historical significant wave height data.

[0017] Furthermore, step S2.2 specifically includes the following steps:

[0018] S2.2.1. Perform interpolation processing to make the spatial resolution of historical wind speed data and historical wave significant height data consistent.

[0019] S2.2.2. Align the historical wind speed data and the historical wave significant wave height data with a 1-hour offset in the time dimension, and integrate the historical wind speed data and the historical wave significant wave height data into a T×C×H×W spatiotemporal series data, where T, C, H, and W represent time, channel, height, and width, respectively.

[0020] Furthermore, step S3 is specifically as follows: using a sliding window algorithm, sliding sampling generates training samples, and integrating the training sample data set into N×T×C×H×W, where N, T, C, H and W represent the number of samples, time, channel, height and width respectively.

[0021] Furthermore, the encoding layer in step S4 includes: an encoded spatiotemporal sequence of multiple time steps connected in series, and the encoded spatiotemporal sequence of each time step includes: a first input layer, a first CBAM layer, and a first ConvLSTM unit based on convolutional long short-term memory units connected in series in sequence, and the first input layer inputs the wave significant wave height data SWH and the wind speed data of the next time step.

[0022] Furthermore, the prediction layer in step S4 includes: a prediction spatiotemporal sequence of multiple time steps connected in series, and the prediction spatiotemporal sequence of each time step includes: a second input layer, a second CBAM layer, a second ConvLSTM unit based on convolutional long and short-term memory units connected in series three times, and a convolutional layer Conv2D, and the hidden layer state of the second ConvLSTM unit based on convolutional long and short-term memory units connected in series three times for the time step is formed through the convolutional layer Conv2D to form the wave significant wave height data SWH of the next time step, and the wave significant wave height data SWH output by the convolutional layer Conv2D of each time step is used as the input data of the second input layer of the next time step, and the second input layer also inputs the wind speed data of the previous time step, and outputs the prediction result after the prediction spatiotemporal sequence of multiple time steps connected in series.

[0023] Furthermore, step S5 is specifically as follows: inputting the training set of historical wind speed data and historical wave data obtained in step S3 into the model, performing iterative training, and obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model.

[0024] The present invention has the following beneficial effects:

[0025] The present invention combines the advantages of the attention mechanism CBAM, historical wind speed data, historical wave significant wave height data, and the coding layer-prediction layer. By introducing future wind speed data, the present invention can quickly and accurately predict the wave significant wave height, which plays an important role in ensuring the safety of navigation and coastal activities. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0027] Figure 1 A flow chart of a method for predicting significant wave height of ocean waves based on a deep learning model of spatiotemporal sequences according to the present invention is shown.

[0028] Figure 2The structure diagram of the spatiotemporal sequence deep learning model of the present invention is shown.

[0029] Figure 3 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 1 hour is shown.

[0030] Figure 4 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 3 hours is shown.

[0031] Figure 5 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 6 hours is shown.

[0032] Figure 6 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 12 hours is shown.

[0033] Figure 7 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 24 hours is shown.

[0034] Figure 8 A two-dimensional scatter density plot of the predicted and actual significant wave heights for the next 48 hours is shown. DETAILED DESCRIPTION

[0035] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0036] like Figure 1 The method for predicting significant wave height based on a deep learning model of spatiotemporal sequences is shown, and specifically includes the following steps:

[0037] S1. Obtain historical wind speed data and historical significant wave height data for the test area. For example, obtain historical wind speed data at 10 meters above the sea surface and historical significant wave height data for the test area.

[0038] S2, the acquired historical wind speed data and historical wave significant wave height data are quality screened, and then spatiotemporal alignment is performed to align the historical wind speed data and historical wave significant wave height data with a 1-hour offset in the time dimension, that is, the wind speed data at time T+1 matches the wave data at time T.

[0039] S3, uses the sliding window algorithm to generate spatiotemporal sequence pairs and construct a training set.

[0040] S4, builds a spatiotemporal sequence deep learning model, including an encoding layer and a prediction layer. The encoding layer is responsible for encoding and memorizing historical wave conditions, while the prediction layer is responsible for forecasting waves based on historical conditions and future wind speeds.

[0041] S5, inputting the training set constructed in step S3 into the spatiotemporal sequence deep learning model for iterative training, and obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model.

[0042] S6, based on the numerical forecast wind field provided by the outside, inputs the acquired future wind speed data into the trained ocean wave significant wave height prediction model based on the spatiotemporal sequence deep learning model to obtain the forecast results.

[0043] Specifically, step S2 includes the following steps:

[0044] S2.1, perform quality screening on historical wind speed data and historical wave significant wave height data, including missing value processing and land marker masking.

[0045] S2.2, perform spatiotemporal registration of historical wind speed data and historical significant wave height data.

[0046] Specifically, step S2.2 includes the following steps:

[0047] S2.2.1. Perform interpolation processing to make the spatial resolution of historical wind speed data and historical wave significant height data consistent.

[0048] S2.2.2. Align the historical wind speed data and the historical wave significant wave height data with a 1-hour offset in the time dimension, and integrate the historical wind speed data and the historical wave significant wave height data into a T×C×H×W spatiotemporal series data, where T, C, H, and W represent time, channel, height, and width, respectively.

[0049] Specifically, step S3 is as follows: using the sliding window algorithm, sliding sampling generates training samples, and integrating the training sample data set into N×T×C×H×W, where N, T, C, H and W represent the number of samples, time, channel, height and width respectively.

[0050] Specifically, the encoding layer in step S4 includes: a coded spatiotemporal sequence of multiple time steps connected in series, and the coded spatiotemporal sequence of each time step includes: a first input layer, a first CBAM layer, and a first ConvLSTM unit based on convolutional long short-term memory units connected in series in sequence. The first input layer inputs the wave significant wave height data SWH and the wind speed data of the next time step.

[0051] Specifically, the prediction layer in step S4 includes: a prediction spatiotemporal sequence of multiple time steps connected in series, and the prediction spatiotemporal sequence of each time step includes: a second input layer, a second CBAM layer, a second ConvLSTM unit based on convolutional long and short-term memory units connected in series three times, and a convolutional layer Conv2D, and the hidden layer state of the second ConvLSTM unit based on convolutional long and short-term memory units connected in series three times for the time step is formed through the convolutional layer Conv2D to form the wave significant wave height data SWH of the next time step, and the wave significant wave height data SWH output by the convolutional layer Conv2D of each time step is used as the input data of the second input layer of the next time step, and the second input layer also inputs the wind speed data of the previous time step, and outputs the prediction result after the prediction spatiotemporal sequence of multiple time steps connected in series.

[0052] Among them, CBAM is used to enhance the characteristics of input data, the convolutional layer Conv2D is responsible for characterizing the forced motion and spatial correlation of the significant wave height data, and the ConvLSTM unit is responsible for characterizing the temporal memory of the wave field;

[0053] like Figure 2 As shown in the figure, in the encoding layer, at time T-11, the significant wave height data SWH and the wind speed data Wind (T-10) of the next time step are input into the first attention mechanism CBAM layer, and then pass through the first ConvLSTM unit of the first layer, the second layer and the third layer in sequence. The data processing process of other time steps is similar to that of time T-11, and is input into the prediction layer after being encoded by the encoding layer.

[0054] In the prediction layer, at time T, the significant wave height data SWH of the previous time step and the wind speed data Wind(T+1) of the next time step are input into the second attention mechanism CBAM layer, and then pass through the second ConvLSTM unit of the first, second and third layers in sequence. The data processing process for other time steps is similar to that at time T, and the prediction results are output after passing through the prediction layer.

[0055] At the encoding stage, the input elements include wind speed data and significant wave height data. During the prediction stage, the significant wave height data is updated from the previous prediction step, while the wind speed data is updated from an external numerical forecast wind field. At the prediction layer, the second ConvLSTM unit of the third layer at that time step is passed through a Conv2D convolutional layer to form the significant wave height data SWH forecast for the next time step, which serves as the input for the next prediction.

[0056] Specifically, step S5 is as follows: inputting the training set of historical wind speed data and historical wave data obtained in step S3 into the model, performing iterative training, and obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model.

[0057] The effectiveness of the method provided by the present invention is verified below:

[0058] To ensure representative training data, wind and wave data from a specific region were used as the training set. These data were obtained from the European Centre for Medium-Range Weather Forecasts' Fifth Generation Atmospheric Reanalysis (ERA5) dataset. The wind and wave data were preprocessed, including spatiotemporal interpolation, normalization, land marking, and feature vectorization. The feature fields were spatiotemporally interpolated over a spatial range of 4°N–26°N and 100°E–123°E, with a spatial resolution of 0.5°×0.5°, a temporal resolution of 1 hour, and a grid size of 45×45. The maximum and minimum values ​​of each feature were counted and normalized to 0–1. Values ​​in land areas were set to 0. The features are vectorized into spatiotemporal series data of T×C×H×W, where T, C, H, and W represent time, channel, height, and width, respectively. A sliding window algorithm is used to generate training samples by sliding sampling, and the dataset is integrated into N×T×C×H×W, where N, T, C, H, and W represent the number of samples, time, channel (number of features), height, and width, respectively.

[0059] The key parameters in the ocean significant wave height prediction model based on the spatiotemporal sequence deep learning model (such as the number of channels in each ConvLSTM unit, the size of the convolution kernel, the hidden layer dimension, etc.) need to be determined through iterative training experience to minimize the prediction error.

[0060] The performance of the ocean significant wave height prediction model based on the spatiotemporal sequence deep learning model is verified and evaluated by the following formula.

[0061] The first is the root mean square error, which represents the deviation between the model results and the observed values. The smaller the root mean square error RMSE, the better the model performance. The formula is as follows:

[0062] ;

[0063] in, represents the land mask matrix, represents the root mean square error, Represents the amount of data involved in verification, Representative ocean area sample points, To predict the results, For real data.

[0064] The correlation coefficient indicates the correlation between the model results and the observed values. The correlation coefficient CC formula is as follows:

[0065] ;

[0066] in, is the average value of the prediction results, is the average value of the real data.

[0067] Mean absolute error It is an objective function that represents the mean absolute difference between the predicted value and the true value. It can measure the average error between the predicted value and the true value and has good robustness.

[0068] ;

[0069] The experimental data were processed according to the method of the present invention, and the experimental results were analyzed as follows:

[0070] For forecasts of 1 to 48 hours into the future, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) remained low, ranging from 0.0806m to 0.2213m, 0.0469m to 0.1515m, and 5.1487% to 15.6429%, respectively. The correlation coefficient (CC) also remained between 0.9959 and 0.9728, demonstrating that the method provided by this invention can accurately predict future wave field changes.

[0071] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for predicting significant wave height based on a deep learning model of spatiotemporal sequences, characterized in that: The specific steps include: S1, obtain historical wind speed data and historical significant wave height data of the test area; S2, the acquired historical wind speed data and historical wave significant wave height data are quality screened, and then spatiotemporal registration is performed to align the historical wind speed data and historical wave significant wave height data with a time offset of 1 hour; S3, using the sliding window algorithm to generate spatiotemporal sequence pairs and construct a training set; S4, builds a spatiotemporal sequence deep learning model, including an encoding layer and a prediction layer. The encoding layer is responsible for encoding and memorizing historical wave conditions, while the prediction layer is responsible for forecasting waves based on historical conditions and future wind speeds. S5, inputting the training set constructed in step S3 into the spatiotemporal sequence deep learning model for iterative training, thereby obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model; S6, based on the numerical forecast wind field provided by the external party, inputs the acquired future wind speed data into the trained ocean wave significant wave height prediction model based on the spatiotemporal sequence deep learning model to obtain the forecast result; The encoding layer in step S4 includes: a coded spatiotemporal sequence of multiple time steps connected in series, and the coded spatiotemporal sequence of each time step includes: a first input layer, a first CBAM layer, and a first ConvLSTM unit based on convolutional long short-term memory units connected in series in sequence, and the first input layer inputs the wave significant wave height data SWH and the wind speed data of the next time step; The prediction layer in step S4 includes: a prediction spatiotemporal sequence of multiple time steps connected in series, and the prediction spatiotemporal sequence of each time step includes: a second input layer, a second CBAM layer, a second ConvLSTM unit based on convolutional long and short-term memory units connected in series in sequence, and a convolutional layer Conv2D. The hidden layer state of the second ConvLSTM unit based on convolutional long and short-term memory units connected in series in the time step is formed through the convolutional layer Conv2D to form the wave significant wave height data SWH of the next time step. The wave significant wave height data SWH output by the convolutional layer Conv2D of each time step is used as the input data of the second input layer of the next time step. The second input layer also inputs the wind speed data of the previous time step. After the prediction spatiotemporal sequence of multiple time steps is connected in series, the prediction result is output.

2. The method for predicting significant wave height based on a spatiotemporal sequence deep learning model according to claim 1, characterized in that: Step S2 specifically includes the following steps: S2.1, perform quality screening on historical wind speed data and historical significant wave height data, including missing value processing and land mark masking; S2.2, perform spatiotemporal registration of historical wind speed data and historical significant wave height data.

3. The method for predicting significant wave height based on a spatiotemporal sequence deep learning model according to claim 2, characterized in that: Step S2.2 specifically includes the following steps: S2.2.1, perform interpolation processing to make the spatial resolution of historical wind speed data and historical wave significant height data consistent; S2.2.

2. Align the historical wind speed data and the historical wave significant wave height data with a 1-hour offset in the time dimension, and integrate the historical wind speed data and the historical wave significant wave height data into a T×C×H×W spatiotemporal series data, where T, C, H, and W represent time, channel, height, and width, respectively.

4. The method for predicting significant wave height based on a spatiotemporal sequence deep learning model according to claim 1, characterized in that: Step S3 is specifically as follows: using the sliding window algorithm, sliding sampling generates training samples, and integrating the training sample data set into N×T×C×H×W, where N, T, C, H and W represent the number of samples, time, channel, height and width respectively.

5. The method for predicting significant wave height based on a spatiotemporal sequence deep learning model according to claim 1, characterized in that: Step S5 specifically includes: inputting the training set of historical wind speed data and historical wave data obtained in step S3 into the model, performing iterative training, and obtaining a wave significant wave height prediction model based on the spatiotemporal sequence deep learning model.

Images