Tide prediction method and system based on neural network
Through the tide prediction method based on neural network, using frequency domain computing and multi-layer neural network to process historical and observation data, the problems of high tide simulation calculation cost and poor stability in the existing technology are solved, and efficient and accurate tide waveform prediction is achieved. .
Patent Information
- Application Number
- CN202411505088.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2024-11-26
AI Technical Summary
The existing tidal simulation technology has high calculation costs, poor numerical stability, and insufficient adaptability when dealing with complex terrain and nonlinear fluctuations, making it difficult to meet the needs of large-scale tidal simulations.
Using the tide prediction method based on neural network, the tide prediction model is constructed by acquiring and preprocessing historical tide data, extracting periodic features, and tide periodic prediction model is constructed, and the frequency domain calculation is used by Fourier transform and multi-layer neural network to generate high-precision Tide waveform prediction.
Maintaining high-precision tidal simulation at low resolution significantly reduces calculation costs, improves numerical stability and adaptability, improves the calculation efficiency and accuracy of tidal simulation, and is suitable for applications under large-scale tidal simulation and complex terrain conditions.
Smart Images

Figure CN119026484A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tide prediction, and in particular to a tide prediction method and system based on a neural network. Background Art
[0002] Tidal simulation plays an important role in oceanographic and hydrological research, especially in the prediction and analysis of tidal fluctuations, coastal changes and marine engineering. Tides are periodic fluctuations of water bodies caused by the gravitational interaction of the earth, moon and sun. Tidal simulation simulates these complex fluctuations through mathematical models and numerical methods.
[0003] Tidal simulation is usually based on shallow water wave equations. By considering factors such as water depth, topography, and fluid dynamics, it accurately predicts the propagation of tidal waves in oceans, estuaries, bays, and other areas. Since tidal waves are significantly affected by topography in shallow waters, nonlinear and dissipative effects become particularly important. Through tidal simulation, these nonlinear effects can be captured and realistic tidal behavior can be simulated.
[0004] Tidal simulation is not only used to predict tidal times and heights, but also has practical applications in navigation, marine engineering, and disaster prevention and mitigation. For example, in port and waterway design, accurate tidal simulation helps improve the safety of ships entering and leaving ports. It can also help predict extreme tidal events such as storm surges or floods, provide early warning information, and reduce the impact of disasters.
[0005] Existing tidal simulation techniques, such as the finite difference method and the finite element method, can provide a certain degree of accuracy, but face significant challenges when dealing with complex terrain and nonlinear fluctuations. These methods often rely on high-resolution grids, resulting in high computational costs, poor numerical stability, and insufficient adaptability when dealing with multi-scale problems and complex boundary conditions. In addition, the traditional methods have a complex structure, involving a variety of stabilization techniques and fine grid division, which not only increases the difficulty of implementation, but may also lead to a decrease in the robustness of the algorithm.
[0006] Among the existing tidal simulation technologies, traditional numerical methods such as the finite difference method and the finite element method have obvious deficiencies in accuracy and efficiency. These methods usually rely on high-resolution grids to improve calculation accuracy, but this also leads to a significant increase in computational costs, especially in large-scale tidal simulations. In addition, when faced with complex terrain and nonlinear fluctuations, traditional methods have poor numerical stability, are prone to computational errors, and lack adaptability in multi-scale problems and complex boundary conditions, making it difficult to meet the needs of practical applications. Summary of the invention
[0007] In order to solve the deficiencies of the prior art, the present invention provides a tidal prediction method and system based on a neural network; it provides a tidal simulation technology that can maintain high accuracy at low resolution, greatly reducing the computational cost while improving numerical stability and adaptability. The calculation is performed in the frequency domain, which simplifies the complex structure of the traditional method, avoids artificial dissipation and stabilization technology, and makes the overall calculation more efficient and robust, which is suitable for large-scale tidal simulation and applications under complex terrain conditions.
[0008] On the one hand, a tidal prediction method based on a neural network is provided, including: obtaining historical tidal data of a target sea area; preprocessing the historical tidal data; performing time series decomposition on the historical tidal data to obtain periodic characteristics; determining parameters of a tidal periodicity prediction model based on the periodic characteristics; obtaining tidal observation data, inputting the tidal observation data into a tidal periodicity prediction model to obtain periodic characteristics of the tidal observation data; fitting the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data; performing modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; inputting the subsequences of different frequency components into a trained neural network to obtain predicted values of tidal data of the target sea area, and obtaining tidal changes within the prediction period.
[0009] On the other hand, a tidal prediction system based on a neural network is provided, including: an acquisition module, which is configured to: acquire historical tidal data of a target sea area; pre-process the historical tidal data; a decomposition module, which is configured to: perform time series decomposition on the historical tidal data to obtain periodic characteristics; determine the parameters of a tidal periodicity prediction model based on the periodic characteristics; an input module, which is configured to: acquire tidal observation data, input the tidal observation data into a tidal periodicity prediction model, and obtain the periodic characteristics of the tidal observation data; a fitting module, which is configured to: fit the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data; a decomposition module, which is configured to: perform modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; a prediction module, which is configured to: input the subsequences of different frequency components into a trained neural network to obtain predicted values of tidal data of the target sea area, and obtain tidal changes within the prediction period.
[0010] The above technical solution has the following advantages or beneficial effects: The present invention relates to a tidal prediction method and system based on a neural network, which aims to solve the problems of low calculation accuracy, slow simulation speed, and poor simulation accuracy under complex terrain in the existing tidal simulation technology. The method ensures the quality of input data by comprehensive preprocessing of tidal data and terrain data, including data acquisition, normalization, noise filtering and other steps. Subsequently, the neural network performs tidal prediction on the processed data, especially extracting frequency domain features through the Fourier transform layer, and then simulating complex physical processes through a multi-layer neural network, and finally generating an accurate tidal waveform prediction.
[0011] The present invention can maintain high-precision tidal simulation at low resolution, effectively reduce computational costs, and significantly improve numerical stability. In addition, neural networks show good adaptability when dealing with complex terrain and multi-scale tidal characteristics, ensuring that simulation results are more accurate and reliable. The completed experimental results show that neural networks are superior to traditional methods in both accuracy and computational efficiency of tidal simulation, providing strong support for large-scale applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0013] Figure 1 This is a flow chart of the method of embodiment 1.
[0014] Figure 2 This is a diagram of the internal structure of the neural network after training in Example 1.
[0015] Figure 3 FIG. 4 is a diagram showing the internal structure of the Fourier layer with a hypernetwork according to the first embodiment. DETAILED DESCRIPTION
[0016] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0017] like Figure 1As shown, embodiment 1: This embodiment provides a tidal prediction method based on a neural network, including: S101: obtaining historical tidal data of a target sea area; preprocessing the historical tidal data; S102: performing time series decomposition on the historical tidal data to obtain periodic characteristics; based on the periodic characteristics, determining the parameters of a tidal periodicity prediction model; S103: obtaining tidal observation data, inputting the tidal observation data into a tidal periodicity prediction model, and obtaining the periodic characteristics of the tidal observation data; S104: fitting the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data; S105: performing modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; S106: inputting the subsequences of different frequency components into the trained neural network to obtain the predicted value of the tidal data of the target sea area, and obtaining the tidal changes within the prediction period.
[0018] Furthermore, the S101: obtaining historical tidal data of the target sea area includes: using a tidal monitoring station, a sensor, and a buoy to collect historical tidal data of the target sea area.
[0019] It should be understood that tidal data: Tidal data is collected from multiple sources such as satellite observations, ocean buoys, tidal stations, etc. Tidal data includes: historical tidal data, environmental parameters such as sea surface height, wind speed, air pressure, and coastline topographic data. Topographic data: The topographic information of the simulation area is obtained through high-precision topographic measurement data or seabed topographic maps. These data are particularly critical in complex terrain because they determine the propagation path and morphology of tidal waves.
[0020] It should be understood that real-time tidal observation data of the target sea area is obtained from tidal monitoring stations, sensors, buoys or other observation equipment. At the same time, historical tidal data of the sea area is collected, which usually includes factors affecting tidal changes such as tide level, air pressure, wind speed, wind direction, etc.
[0021] Furthermore, the S101: preprocessing the historical tidal data includes: missing value completion, data normalization and noise filtering.
[0022] It should be understood that missing value filling refers to interpolation of missing tidal data to ensure the integrity of input data.
[0023] Data normalization refers to the normalization of data such as tidal height and time series to make the data distribution more suitable for the input of subsequent models and prevent model bias caused by dimensional differences.
[0024] Noise filtering refers to the use of filters or other methods to remove high-frequency noise in the data and retain valid tidal fluctuation information to improve simulation accuracy.
[0025] It should be understood that the raw data is converted into a tidal time series data format suitable for time series analysis. This step may include standardizing tidal data at different time intervals to ensure data consistency and stability.
[0026] Furthermore, the S102: performing time series decomposition on the historical tidal data to obtain periodic characteristics, wherein the time series decomposition includes: Fourier Transform (Fourier Transform) or Empirical Mode Decomposition (EMD).
[0027] When using Empirical Mode Decomposition (EMD) to decompose the historical tidal data into time series, the original tidal time series data is decomposed into several intrinsic mode functions (IMFs) and a residual trend term The several intrinsic mode functions (IMFs) represent different periodic characteristics, and the residual trend term To capture the long-term trend of the data, the mathematical expression corresponding to the empirical mode decomposition (EMD) is: ;in, represents the time series data of the original tide, Represents the decomposition of IMF, each IMF represents a different periodic component (short-term or long-term periodic characteristics). is the remaining long-term trend part, usually a non-oscillating stable signal, is the number of IMFs decomposed. The more IMFs there are, the more refined the periodic characteristics captured.
[0028] Through empirical mode decomposition (EMD), different periodic components of tidal data can be extracted and the short-term and long-term periodic characteristics of tidal phenomena can be analyzed.
[0029] It should be understood that the historical tidal time series data is decomposed into multiple periodic components through time series decomposition technology. Common periodic components include semi-diurnal tides (12-hour cycle), diurnal tides (24-hour cycle) and other tidal changes with longer cycles. The purpose of this step is to identify and extract key periodic features in historical tidal data.
[0030] Furthermore, the S102: determining the parameters of the tidal periodic prediction model based on the periodic characteristics, includes: S102-1: analyzing the time series data, and using correlation calculation, spectrum analysis or wavelet analysis to find the main periodic components of the tidal data; S102-2: determining the parameters required for the periodic prediction model, including: trend terms, period terms and seasonal terms; S102-3: determining the periodic prediction model based on the main periodic components of the tidal data and the parameters required for the periodic prediction model.
[0031] Furthermore, the correlation calculation in S102-1 specifically includes: analyzing the correlation between two time series data, which is a lagged correlation, to help find the periodicity, autocorrelation function It is expressed as follows: ;in, is time series data, is the lag period, is the data length, It is time series data In the time delay lag period The value after.
[0032] Furthermore, the spectrum analysis in S102-1 specifically includes: analyzing the frequency components of the signal by performing Fourier transform on the signal; the spectrum analysis is divided into discrete Fourier transform and power spectral density.
[0033] Discrete Fourier transform means: ;in, is time series data, is the number of data points, is the frequency domain representation, It's the frequency.
[0034] Power spectral density refers to: ;in, is the Fourier transform result, Indicates the power of each frequency.
[0035] The wavelet analysis in S102-1 specifically includes: analyzing the frequency components of the time series data that change with time, wherein the continuous wavelet transform function as follows: ;in, is time series data, is the conjugate of the wavelet function, Indicates scale, Represents the translation parameter.
[0036] Furthermore, the trend item in S102-2 describes the long-term change direction of the time series data, the overall upward or downward trend of the tidal data over time, and reflects the global characteristics of the time series data, which usually has no periodicity but a long-term change that changes slowly.
[0037] Further, the trend item in S102-2 is used Indicates that Represents a trend that changes slowly over time t. For example, in tide prediction, the trend term in tide data can reflect the overall change in tide level due to climate change or sea level rise.
[0038] The periodic term in S102-2 is approximately expressed by a sine wave function: ;in, is the amplitude, is the cycle length, is the phase of the periodic term.
[0039] Periodic terms represent components in the data that have obvious periodic fluctuations, but unlike seasonality, the period of periodic components is usually not fixed and may exceed one year. Periodic terms reflect the cyclical pattern in the data and can be long-term (such as multi-year periodic changes) or short-term. In tidal prediction, the periodic components in tidal data may be related to changes in the gravitational pull of the Earth and the Moon or other astronomical factors.
[0040] Furthermore, the seasonal term in S102-2 is a periodic fluctuation with a fixed period, which is usually related to seasonal changes, climate changes or rhythmic changes within a year. Its period is stable and shows similar patterns in each period. For example, in tide prediction, the tide fluctuates seasonally in the four seasons of the year or in relation to the phases of the moon.
[0041] Seasonal term, expressed as a sinusoidal function with a fixed period S : ;in, is the amplitude, is the cycle length, is the phase of the seasonal term.
[0042] Further, S102-3: based on the main periodic components of the tidal data and the parameters required by the periodic prediction model, determine the periodic prediction model, the periodic prediction model is a time series model; the time series model includes: autoregressive integrated moving average model ARIMA (Autoregressive Integrated Moving Average), seasonal autoregressive integrated moving average model SARIMA (Seasonal Autoregressive Integrated Moving Average), or autoregressive model. The output of this model will be used to capture the periodic fluctuations of future tides.
[0043] Among them, the Autoregressive Integrated Moving Average (ARIMA) model is a statistical model used for time series analysis and prediction. The ARIMA model combines the autoregressive (AR) part, the difference (I) part, and the moving average (MA) part, and is suitable for non-stationary time series.
[0044] Among them, the seasonal autoregressive integrated moving average model SARIMA (Seasonal Autoregressive Integrated Moving Average) is an extension of the ARIMA model, which takes into account the seasonal factors in the time series. SARIMA adds seasonal autoregressive and seasonal moving average components to the ARIMA model to better capture the impact of seasonal changes on time series.
[0045] Taking the ARIMA time series model as an example, the ARIMA (Autoregressive Integrated Moving Average) model is used to process stationary time series data. It combines autoregression (AR), difference (I) and moving average (MA) components.
[0046] Among them, the autoregressive model AR (Autoregressive) is a statistical model that uses the past values of a time series to predict future values. In the AR model, the current value is a linear combination of its previous values. Its basic assumption is that the time series has a certain degree of linear dependence, which is usually modeled through regression analysis. The order of the AR model ( ) indicates how many past values to use.
[0047] Differentiation (I) refers to the transformation of a time series to eliminate trends and seasonality and make it stationary. The basic idea of differentiation is to calculate the difference between adjacent observations of the series to eliminate non-stationarity. In the ARIMA model, the order of "I" ( ) indicates that several difference operations are required to make the time series stable. The differenced series can better fit the model and predict. MA (Moving Average) is a statistical model that predicts future values by calculating the random error term (ie, noise) of the time series. In the moving average MA model, the current value is a linear combination of its past random error terms. The MA model assumes that the change in the current observation value is affected by the previous errors. The order of the model ( ) indicates how many past random error terms are used. The form of the ARIMA model is: ;in, is the observed value at time point t, is the autoregressive (AR) coefficient, is the moving average coefficient (MA coefficient), is the white noise error term (random error). is the past observation value, which is used for the calculation of the autoregressive part (AR part) is the past error term, used in the calculation of the moving average part (MA part), The coefficient reflects the continuity and periodicity of the tidal values over time. If the tidal data exhibits a certain periodicity (such as diurnal tides, lunar tides, etc.), the corresponding autoregressive coefficient will reflect this feature.
[0048] In the ARIMA model, the trend term is indirectly represented by the difference operation, and the number of differences is determined by the parameter The form of the ARIMA model is ARIMA(p,d,q), where represents the order of the autoregressive term, Indicates the number of differences, used to remove trends, Indicates the order of the sliding average term.
[0049] If you need to differentiate The obtained difference time series is expressed as: ;in, Indicates passing When using the ARIMA model for modeling, the original time series is first differentiated to obtain a stable time series before model fitting.
[0050] For example, suppose that after one difference the stationary time series is obtained , then the ARIMA model is expressed as: ; is the autoregression (AR) coefficient, is the moving average coefficient (MA coefficient), is a white noise error term (random error); through the difference operation, the ARIMA model can indirectly represent the trend term. It reflects the degree of trend removal and ensures that the time series is stationary before modeling. This method makes time series analysis more effective and can better capture the intrinsic characteristics of the data. The ARIMA model is extended to seasonal ARIMA (SARIMA).
[0051] .
[0052] in, is the autoregression (AR) coefficient, is the seasonal autoregressive coefficient, is the length of the seasonal cycle (e.g., a 365-day annual cycle).
[0053] Model parameters include: autoregressive order, difference number, moving average order, seasonal autoregressive order, seasonal difference and seasonal moving average order, etc. Combining these parameters with the preset model formula, a periodic prediction model suitable for tidal prediction is constructed.
[0054] Furthermore, S103: acquiring tidal observation data, inputting the tidal observation data into a tidal periodicity prediction model, and obtaining periodic characteristics of the tidal observation data, wherein the periodic characteristics of the tidal observation data include: tidal height data, timestamp, daily cycle, monthly cycle, annual cycle, seasonal component, amplitude, phase and long-term trend.
[0055] The tidal height data (tidal level) is the actual measured tidal water level height, usually time series data, such as daily or hourly tidal level records.
[0056] Timestamp, which is the time information associated with each tide height data point, including date and time.
[0057] The diurnal cycle is the daily tidal cycle caused by the rotation of the Earth. It usually includes a semi-diurnal cycle (semi-diurnal tides) and a full-day cycle (full-day tides).
[0058] The lunar cycle is the monthly tidal cycle caused by the gravitational force of the moon, including the synodic month (about 29.53 days) and the lunar cycle (about 27.32 days).
[0059] The annual cycle is the annual tidal cycle caused by the movement of the Earth around the Sun, usually 365.25 days.
[0060] The seasonal component is the tidal pattern that reflects seasonal changes, such as the difference in tides between summer and winter.
[0061] Amplitude is the strength or amplitude of each periodic component, indicating the magnitude of tidal changes.
[0062] Phase is the phase of each periodic component, which indicates the time position of the periodic change, that is, the time when the peak or trough of the periodic wave occurs.
[0063] Long-term trends are long-term changes that may exist in tidal data, such as long-term sea level changes caused by climate change.
[0064] Furthermore, the step S104: fitting the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data includes: fitting the periodic characteristics with the historical tidal data so that the fitted tidal time series data conforms to the observed historical tidal pattern. The fitted data is used for further tidal change analysis and prediction.
[0065] It should be understood that fitting refers to selecting a mathematical model in statistics and machine learning to make the relationship between a set of data points as close as possible. The goal of fitting is to find the optimal parameters so that the model can well describe the pattern or trend of the data. The present invention uses ARIMA and SARIMA in the fitting process.
[0066] Furthermore, the S105: performing modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components, includes: using a time series decomposition method to perform modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components, and the time series decomposition includes: Fourier transform, short-time Fourier transform or wavelet transform.
[0067] Fourier transform can convert time series data into a superposition of sine waves of different frequencies, thereby separating the main tidal cycle components. In this way, characteristic data of different frequencies such as semi-diurnal tides and diurnal tides can be extracted to prepare for further tidal predictions.
[0068] This decomposition method helps to identify important periodicity and trend information in tidal data and convert it into subsequences suitable for subsequent neural network model processing.
[0069] Further, the S106: input the subsequences of different frequency components into the trained neural network to obtain the tidal data prediction value of the target sea area and the tidal change situation within the prediction period, wherein the trained neural network, such as Figure 2As shown, its network structure includes: an input layer, a first fully connected layer, several Fourier layers with a hypernetwork connected in series, a second fully connected layer and an output layer, the second fully connected layer is also connected in parallel with the third fully connected layer, the input end of the third fully connected layer is connected to the output end of the last Fourier layer with a hypernetwork, and the output end of the third fully connected layer is connected to the input end of the output layer.
[0070] Furthermore, the input layer is used to receive tidal observation data , indicating that the system is at the initial time The tidal conditions at different observation locations at different times are used as the input of the neural network model. Normalize and denoise to reduce noise interference in the input data and ensure the accuracy of subsequent calculations.
[0071] Furthermore, the first fully connected layer transforms the low-dimensional tidal observation data Project into a high-dimensional space to generate intermediate variables By increasing the dimensionality, the system is able to lift the tidal data to a higher-dimensional feature space, thereby better capturing complex nonlinear relationships. This step is critical because it provides richer feature information for subsequent network layers.
[0072] Furthermore, if Figure 3 As shown, the Fourier layer with a hypernetwork includes: an input end of the Fourier layer with a hypernetwork, and the input end of the Fourier layer with a hypernetwork is respectively connected to the input end of the first branch and the input end of the second branch; the first branch includes: a first linear layer and a second linear layer connected in sequence; the input end of the first linear layer is the input end of the first branch, and the output end of the second linear layer is the output end of the first branch; the second branch includes: a Fourier transform layer, a signal processing layer and an inverse Fourier transform layer connected in sequence; the input end of the Fourier transform layer is the input end of the second branch, and the output end of the inverse Fourier transform layer is the output end of the second branch; the output end of the first branch and the output end of the second branch are both connected to the input end of the activation function layer, and the output end of the activation function layer is the output end of the Fourier layer with a hypernetwork.
[0073] Furthermore, the first linear layer and the second linear layer are implemented by a linear transformation (i.e., a fully connected layer, or affine transformation), which is completed by applying matrix multiplication to the input data and adding a bias, and the formula is as follows: ;in, is the input data (the signal extracted from the frequency domain features after Fourier transformation), is a learnable weight matrix, is a learnable bias vector.
[0074] Furthermore, the signal processing layer processes the features after Fourier transformation in the frequency domain (Fourier space), performs multiplication operations on the Fourier coefficients, and adjusts the performance of the signal in the frequency domain.
[0075] Furthermore, the Fourier layer with the hypernetwork includes: using the Fourier neural operator FNO to extract frequency domain features from the input through multiple Fourier layers. The output of each layer It is calculated by the following update formula: ;in, Represents the linear transformation part, through the learned weights Tidal prediction output for the current layer Perform a linear transformation. Represents the convolution operation, kernel function is parameterized and expressed as an operation in the frequency domain via the convolution theorem: . It indicates that the nonlinear activation function is used to enhance the nonlinear expression ability of the network and improve the capture effect of complex tidal characteristics. Represents a convolution kernel or operator, which is used to perform some kind of convolution or kernel transformation on the input signal. is an operator that depends on and parameters , Represents the core The set of parameters is the learnable weights in the neural network. Adjusted through training, helping the model capture patterns or dynamic behaviors in input features, represents time-dependent input, is some external signal or some spatiotemporal feature, Considered as convolution kernel Part of the input, Determines the parameterization of the convolution kernel, It means that at the time step The feature representation or state variable at a certain moment is an intermediate result obtained after the input signal is processed by several layers of neural networks. represents the parameters of Fourier transform, It means that the neural network is used from time Update to next time The characteristics or state of being.
[0076] Specifically, is the new state after one step of updating at the spatial position x, which is determined by the current moment Status Generated through linear transformation, kernel operation, etc. Represents the first and second linear layers, the time dimension or resolution of the input data, the width dimension, which represents the number of sampling points of the data on the time coordinate, that is, the time dimension of the tide observation data, which represents the number of sampling points of the time in the observation data, Represents the spatial coordinate variable, which indicates the location of the problem in the physical space, that is, different tide observation locations or sea areas. Calculate the input data by fast Fourier transform (FFT) Frequency domain representation of , and apply the convolution theorem to perform convolution operations. Inverse Fourier transform The convolution result is converted back to the space-time domain. By performing operations in the frequency domain, the system is able to perform the convolution operation more efficiently. Frequency domain operations are more efficient than directly calculating in the space-time domain, and the inverse transform ensures that the output results can be returned to the space-time domain for subsequent physical analysis.
[0077] The high-dimensional data after frequency domain processing is converted into a tide prediction function through the second and third fully connected layers. The second and third fully connected layers are responsible for decoding the high-dimensional frequency domain features back to the low-dimensional spatiotemporal domain. Through this decoding operation, the final tide prediction results are generated.
[0078] Final output of predicted tidal data , which represents the physical system at time and location The output can be used for further physical analysis or engineering applications, such as solitary wave propagation prediction and coastal management in marine engineering.
[0079] The trained neural network is responsible for generating important parameters in the main neural network. The main neural network is the remaining part of the trained neural network excluding the hypernetwork part. These parameters are generated by the hypernetwork function, and the generation process is controlled by the trainable parameters. and the global context impact.
[0080] ; It is the first The weight matrix of the layer is obtained by the hypernetwork function Generate. According to the tidal data, it is specified later The weights of each layer are dynamically generated, ensuring that the model adapts to different tidal inputs or forecasting requirements.
[0081] ; It is the first The linear transformation matrix of the layer, similar to the weight matrix , through the hypernetwork function Generation. The hypernetwork dynamically generates linear transformation matrices to enhance the model’s tidal data processing capabilities.
[0082] ; is the parameter matrix used to map the low-dimensional tidal input to a high-dimensional space, which is obtained by the hypernetwork function Generation; By lifting the tidal input data into a high-dimensional space, the model can capture more complex tidal characteristic information.
[0083] ; It is a decoding matrix in the main neural network, responsible for converting high-dimensional tidal features into the final output. This step generates corresponding parameters for the decoding layer through the hypernetwork, enabling the model to return from high-dimensional tidal features to low-dimensional predicted output.
[0084] ; is the parameter matrix of another decoding layer in the main neural network. similar, It is used to further transform the output of the tidal solution to ensure that the final tidal prediction conforms to the tidal spatiotemporal structure.
[0085] It should be understood that the input layer is used to receive pre-processed multi-dimensional tidal and topographic data; the Fourier transform layer performs Fourier transform on the input data, converts the spatial domain data into the frequency domain, and extracts key frequency features. This step can effectively capture the periodic characteristics of tidal waves. The signal processing layer is used to process the frequency domain features through a multi-layer neural network, extracting and combining feature information at different frequencies layer by layer. This process can simulate physical processes such as the propagation, refraction and reflection of tidal waves. Inverse Fourier transform layer: converts the processed frequency domain data back to the spatial domain to generate a predicted tidal waveform.
[0086] The structure of the trained neural network realizes the frequency domain information extraction of the spatial dimension through the Fourier layer. Each layer of Fourier transform operation processes different frequency components of the input data. The model also includes weights and bias parameters in each layer to capture complex nonlinear features.
[0087] The feature extraction process in the model mainly includes the frequency domain branch (Fourier layer) and the nonlinear activation branch. The frequency domain branch is used to capture the frequency components of the signal, while the nonlinear activation branch converts the frequency domain features into a time domain solution through a weight matrix (such as a convolutional layer).
[0088] Through multiple Fourier layers and nonlinear operations, the model iteratively calculates the features of the input data and finally outputs the predicted solution through the linear layer.
[0089] Furthermore, S106: input subsequences of different frequency components into the trained neural network to obtain the predicted value of tidal data of the target sea area, and obtain the tidal changes within the prediction period, wherein the training process of the trained neural network includes: constructing a training set, the training set is the tidal data of the tidal changes within a known period; inputting the training set into the neural network, training the neural network, and stopping the training when the total loss function value of the neural network no longer decreases to obtain the trained neural network.
[0090] Furthermore, the total loss function of the trained neural network is expressed as the sum of the SmoothL1Loss loss function and the mean square error loss function.
[0091] The SmoothL1Loss loss function is a combination of Loss and A mixed loss function of loss is used to deal with regression problems. In tidal simulation, it can effectively deal with the difference between the model prediction value and the true value.
[0092] .
[0093] In the formula of SmoothL1Loss, is the predicted value of the model, is the real tidal data value, is a hyperparameter that controls Loss and Loss balance. SmoothL1Loss uses Loss is smoothed; when it is far from the true value, use The loss reduces sensitivity to outliers.
[0094] The Mean Squared Error (MSE) is another commonly used loss function, especially for regression tasks. In tidal data prediction, MSE is used to measure the average squared difference between the model prediction value and the actual observation value.
[0095] The formula for MSE is as follows: ;in, It is Prediction values, It is True value, is the number of samples. MSE evaluates the performance of the model by calculating the squared difference between the predicted value and the true value and averaging it over all samples. MSE gives a larger penalty for large errors, so in tide prediction, it helps to reduce large prediction errors.
[0096] These two loss functions help the model adjust parameters during training to improve prediction accuracy and ensure the robustness of the model when facing tidal data.
[0097] Loss function design: Define a suitable loss function (such as mean square error, MSE) to measure the difference between the model prediction value and the actual tidal data, and guide the model optimization process.
[0098] Optimization algorithm: Use advanced optimization algorithms (such as the Adam optimizer) to iteratively update model parameters and gradually reduce the error on the training set.
[0099] Hyperparameter tuning: Optimize the model’s hyperparameters (such as learning rate, number of network layers, number of nodes, etc.) through methods such as cross-validation and grid search to achieve the best prediction effect.
[0100] Early stopping mechanism: Set up an early stopping mechanism during the training process to prevent the model from overfitting and improve the generalization ability of the model.
[0101] Validation set evaluation: After each training cycle, the validation set is used to evaluate the model performance and adjust the model parameters.
[0102] Test set evaluation: Finally, a comprehensive evaluation is performed on the test set to confirm the prediction accuracy and robustness of the model.
[0103] By introducing the trained neural network model, the Fourier neural operator (FNO) was expanded and a hypernetwork was added. The design of the hypernetwork enables the model to be optimized in a higher-dimensional parameter space, effectively improving the model's ability to express tidal fluctuations. This improvement not only ensures that the model remains highly accurate under low-resolution conditions, but also significantly reduces the calculation time, providing a more efficient solution for large-scale tidal simulations.
[0104] Furthermore, in step S106, the subsequences of different frequency components are input into the trained neural network to obtain the tidal data prediction value of the target sea area and the tidal change situation within the prediction period. The working process of the trained neural network includes: the input layer Input into the trained neural network, receiving space and time The combined input signal , which represents the initial conditions of the tidal data or historical observations. Perform preliminary processing to embed the input data into a higher-dimensional feature space. This step is achieved through a linear layer or embedding layer to convert it into a feature representation of a specific dimension.
[0105] Fourier Layer with HyperNetwork Input data is processed through a series of Fourier layers, each of which performs a frequency domain transform of the signal. Specifically: F: Perform Fourier transform, convert it into frequency space representation, and extract frequency domain features.
[0106] R: Perform operations in the frequency domain, usually by frequency filtering or convolution, to capture space-time dependencies. The Fourier coefficients are processed by a linear transform R in the complex domain.
[0107] :The processed frequency domain signal is transformed by inverse Fourier transform Convert back to the time domain, thereby generating an updated representation of the spatiotemporal signal.
[0108] Stacking multiple layers: The process of Fourier transform and inverse transform can be repeated at multiple Fourier layers to form a deep model that gradually extracts complex tidal patterns and space-time dependencies.
[0109] Linear transformation: Before each Fourier layer outputs the result, an additional linear combination of the spatiotemporal signal is performed through a linear transformation W. The role of this step is to perform linear operations on the time domain signal to further extract feature information at different time steps.
[0110] Non-linear activation function layer , the signal after linear transformation is input into the nonlinear activation function In the above example, the signal is nonlinearly mapped by the multi-element vector (such as ReLU, Sigmoid or Tanh). This step can enhance the model's ability to represent complex tidal patterns.
[0111] Spatiotemporal state update: At each time step The input signal is processed by the Fourier layer, linear layer and nonlinear activation layer to obtain the updated tidal state This process recursively passes the historical tide data to the next time step.
[0112] Output layer Q, the signal processed by multiple Fourier layers is finally input to the output layer Q, which maps the feature into the actual tidal prediction value The output result can be the tide height or other physical quantities related to the tide in a certain period of time in the future.
[0113] The trained neural network can effectively process complex spatiotemporal data. Through the dual-channel model structure, the first branch is used to process the periodic changes of tides, and the second branch is used to process the interaction between tides and environmental factors (meteorological factors: such as temperature, air pressure, humidity, wind speed and wind direction, etc. These factors affect the climate conditions of the ocean, thereby indirectly affecting the tides. Ocean factors: such as sea surface temperature, ocean salinity, ocean currents, etc. These factors can change the density and flow of water bodies, thereby affecting tides. Topographic factors: such as coastline shape, seabed topography, etc. Changes in topography will affect the propagation and deformation of tides. Astronomical factors: such as moon phases, sun position, etc., these factors determine the periodic changes of tides.). Combining the results of these two channels, the trained neural network generates the tidal changes of the target sea area within the prediction period.
[0114] Specifically, the original space-time domain problem is first converted into an operation in the frequency domain through Fourier transform processing, thereby reducing the need for high-resolution grids and alleviating computational complexity. During the model training process, the trained neural network can better capture the nonlinear characteristics of tidal fluctuations by dynamically adjusting network parameters using a large amount of tidal data. Especially in the simulation of complex terrain and multi-scale tidal characteristics, the Fourier layer with a hypernetwork demonstrates its flexibility, enabling the model to adaptively adjust parameters to ensure the accuracy and stability of the simulation results.
[0115] By comparing and verifying with actual observation data, the trained neural network significantly improved the computational efficiency and prediction accuracy of tidal simulation. This technical solution provides an efficient tidal simulation tool that can cope with complex terrain conditions, bringing important technological progress to the practical application of tidal prediction.
[0116] The tidal simulation technology of the trained neural network of the present invention significantly improves the efficiency and accuracy of tidal prediction based on traditional numerical methods. Traditional tidal simulation methods, such as finite difference method and finite element method, usually require high-resolution grids to ensure simulation accuracy, which leads to a significant increase in computational costs, and at the same time exhibits numerical instability and insufficient accuracy in the processing of complex terrain and nonlinear fluctuations. In contrast, the trained neural network converts tidal data from the spatiotemporal domain to frequency domain operations through Fourier transform, and uses a hypernetwork to optimize model parameters, so that high-precision simulation can still be maintained under low-resolution grids. This improvement increases the computational efficiency by about 10%-20%, and the error of the simulation results is reduced by 10%-30% compared to traditional methods.
[0117] In addition, the trained neural network enhances the adaptability and robustness of tidal simulation. Traditional methods often show insufficient adaptability when faced with complex terrain and multi-scale tidal characteristics, and it is difficult to maintain a stable simulation effect under different conditions. The trained neural network can effectively cope with complex terrain and tidal scale changes through the adaptive adjustment of the Fourier layer with a hypernetwork in deep learning, and its adaptability is improved by about 15%-20%. At the same time, the present invention simplifies the implementation structure of the model, reduces the dependence on fine grids and complex stabilization techniques, and makes the implementation and debugging process of the model more efficient. In summary, the application of trained neural networks in tidal simulation provides a more efficient and accurate solution, which significantly improves the defects in traditional methods.
[0118] In tidal simulation, the improvement in accuracy leads to more accurate predictions, especially under complex coastal terrain and nonlinear fluctuation conditions. High-precision simulation enables users to more accurately predict tidal rise and fall, tidal range and extreme tide levels, which is crucial for coastal management, shipping, fisheries, and disaster prevention and mitigation. In addition, the improvement in accuracy reduces the accumulation of errors in long-term predictions, ensuring the reliability of prediction results over a long period of time, which is of great significance for planning marine engineering and assessing the impact of climate change.
[0119] At the same time, the shortened calculation time significantly enhances the real-time performance of tidal simulation, enabling users to obtain tidal prediction results more quickly and respond in a timely manner in emergencies or extreme weather. This also improves the ability to process large-scale data, allowing a wider area to be covered or multi-region predictions to be made simultaneously in the same time, greatly supporting the needs of coastal management projects and ocean observation systems. By improving accuracy and reducing calculation time, the practical application value of tidal simulation has been significantly enhanced, helping users make more scientific and effective judgments in ocean management and safety decisions.
[0120] The tidal simulation system in this embodiment has efficient data processing and analysis capabilities, and can improve the accuracy of tidal predictions through precise historical data analysis and periodic model construction, providing reliable technical support for marine environment monitoring and early warning.
[0121] The trained neural network is expressed as: ,in, It represents the output of the L+1th layer in the main neural network, and represents the next layer result of tide prediction. Represents a nonlinear activation function, which is used to enhance the model's ability to express tidal data. and Represents the weight matrix and linear transformation matrix generated by the hypernetwork. Represents the kernel operator applied in each layer, which is operated by Fourier transform. This formula describes the update process of each layer. The model extracts the characteristics of tidal input data through linear transformation, nonlinear activation and frequency domain operation.
[0122] The trained neural network is passed through trainable parameters and the global context Dynamically generate weights and parameter matrices to ensure that the model is highly adaptable in tidal simulation tasks. Specifically: Trainable parameters Controls the model's expressiveness in tidal simulations. Global Context The background information related to the tidal characteristics is provided, and the model adjusts the parameter generation according to different input tasks. Through this mechanism of dynamically generating parameters, the trained neural network shows strong adaptability and prediction accuracy in tidal simulation tasks.
[0123] This paper proposes a tidal simulation technology solution based on a trained neural network, focusing on solving the efficiency and accuracy problems in tidal prediction. Traditional numerical simulation methods often face the challenges of low computational efficiency and insufficient accuracy when dealing with large-scale tidal predictions, especially when dealing with complex terrain and multi-scale tidal fluctuations. The technical solution for tidal simulation first involves the processing of input data to ensure the quality and applicability of the data.
[0124] Embodiment 2. This embodiment provides a tidal prediction system based on a neural network, including: an acquisition module, which is configured to: acquire historical tidal data of a target sea area; pre-process the historical tidal data; a decomposition module, which is configured to: perform time series decomposition on the historical tidal data to obtain periodic characteristics; determine the parameters of a tidal periodicity prediction model based on the periodic characteristics; an input module, which is configured to: acquire tidal observation data, input the tidal observation data into a tidal periodicity prediction model, and obtain the periodic characteristics of the tidal observation data; a fitting module, which is configured to: fit the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data; a decomposition module, which is configured to: perform modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; a prediction module, which is configured to: input the subsequences of different frequency components into a trained neural network to obtain predicted values of tidal data of the target sea area, and obtain tidal changes within the prediction period.
[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. The tidal prediction method based on neural network is characterized by: include: Obtain historical tidal data of the target sea area; Preprocessing of historical tidal data; Perform time series decomposition on historical tidal data to obtain periodic characteristics; Based on the periodic characteristics, the parameters of the tidal periodicity prediction model are determined; Acquire tidal observation data, input the tidal observation data into a tidal periodicity prediction model, and obtain the periodic characteristics of the tidal observation data; Fit the periodic characteristics of tidal observation data with historical tidal data to obtain fitted tidal time series data; Perform modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; The subsequences of different frequency components are input into the trained neural network to obtain the predicted values of tidal data of the target sea area and the tidal changes within the prediction period.
2. The tidal prediction method based on a neural network as claimed in claim 1, characterized in that: Performing time series decomposition on historical tidal data to obtain periodic characteristics, wherein the time series decomposition includes: Fourier transform or empirical mode decomposition; When using empirical mode decomposition to decompose historical tidal data into time series, the time series data It is decomposed into several intrinsic mode functions IMFs and a residual trend term ; The several intrinsic mode functions IMFs represent different periodic characteristics, and the residual trend term To capture the long-term trend of the data, the mathematical expression corresponding to the empirical mode decomposition EMD is: ;in, Represents the decomposition of IMF, each IMF represents a different periodic component. is the remaining long-term trend part, which is a non-oscillating stable signal. is the decomposed IMF quantity.
3. The tidal prediction method based on a neural network as claimed in claim 1, characterized in that: Based on the periodic characteristics, the parameters of the tidal periodic prediction model are determined, including: analyzing the time series data, using correlation calculation, spectrum analysis or wavelet analysis to find the main periodic components of the tidal data; determining the parameters required for the periodic prediction model, including: trend terms, periodic terms and seasonal terms; determining the periodic prediction model based on the main periodic components of the tidal data and the parameters required for the periodic prediction model.
4. The tidal prediction method based on a neural network as claimed in claim 3, characterized in that: Correlation calculation, specifically including: analyzing the correlation between two time series data, which is lagged correlation, helping to find periodicity, autocorrelation function It is expressed as follows: ;in, is the data length, It is time series data In the time delay lag period The value after Spectral analysis includes: analyzing the frequency components of a signal by performing Fourier transform on it; Spectral analysis is divided into discrete Fourier transform and power spectral density; discrete Fourier transform refers to: ;in, is the number of data points, is the frequency domain representation, is the frequency; Power spectral density refers to: ;in, is the Fourier transform result, Indicates the power of each frequency; Wavelet analysis, specifically including: analyzing the frequency components of time series data that change over time, where the continuous wavelet transform function as follows: ;in, is the conjugate of the wavelet function, Indicates scale, Represents the translation parameter.
5. The tidal prediction method based on a neural network as claimed in claim 3, characterized in that: The periodic term is approximated by a sine wave function: ;in, is the amplitude, is the cycle length, is the phase of the periodic term; the seasonal term, expressed as a sinusoidal function with a fixed period length S : ;in, is the phase of the seasonal term.
6. The tidal prediction method based on a neural network as claimed in claim 3, characterized in that: Based on the main periodic components of tidal data and the parameters required by the periodic prediction model, a periodic prediction model is determined, wherein the periodic prediction model is a time series model; the time series model includes: an autoregressive integrated moving average model, a seasonal autoregressive integrated moving average model or an autoregressive model.
7. The tidal prediction method based on a neural network as claimed in claim 1, characterized in that: Acquire tidal observation data, input the tidal observation data into a tidal periodicity prediction model, and obtain periodic characteristics of the tidal observation data, wherein the periodic characteristics of the tidal observation data include: tidal height data, timestamp, daily cycle, monthly cycle, annual cycle, seasonal component, amplitude, phase and long-term trend; The fitted tidal time series data is modally decomposed to obtain subsequences of different frequency components, including: modal decomposition using a time series decomposition method, wherein the time series decomposition includes: Fourier transform, short-time Fourier transform or wavelet transform.
8. The tidal prediction method based on a neural network as claimed in claim 1, characterized in that: The subsequences of different frequency components are input into the trained neural network to obtain the tidal data prediction value of the target sea area and the tidal change situation within the prediction period. The network structure of the trained neural network includes: An input layer, a first fully connected layer, a plurality of Fourier layers with a hypernetwork connected in series, a second fully connected layer, and an output layer connected in sequence, wherein the second fully connected layer is also connected in parallel with a third fully connected layer, an input end of the third fully connected layer is connected to an output end of the last Fourier layer with a hypernetwork, and an output end of the third fully connected layer is connected to an input end of the output layer; The Fourier layer with a hypernetwork comprises: an input end of the Fourier layer with a hypernetwork, wherein the input end of the Fourier layer with a hypernetwork is respectively connected to an input end of a first branch and an input end of a second branch; the first branch comprises: a first linear layer and a second linear layer connected in sequence; the input end of the first linear layer is the input end of the first branch, and the output end of the second linear layer is the output end of the first branch; the second branch comprises: a Fourier transform layer, a signal processing layer and an inverse Fourier transform layer connected in sequence; the input end of the Fourier transform layer is the input end of the second branch, and the output end of the inverse Fourier transform layer is the output end of the second branch; the output end of the first branch and the output end of the second branch are both connected to the input end of an activation function layer, and the output end of the activation function layer is the output end of the Fourier layer with a hypernetwork.
9. The tidal prediction method based on a neural network as claimed in claim 8, characterized in that: Fourier layer with hypernetwork, including: using Fourier neural operator, extracting frequency domain features from input through multiple layers of Fourier layers, and the output of each layer It is calculated by the following update formula: ; Represents the linear transformation part, through the learned weights Tidal prediction output for the current layer Perform a linear transformation, Represents the convolution operation, kernel function is parameterized, , represents a nonlinear activation function, Represents a convolution kernel or operator, kernel is an operator that depends on and parameters , Represents the core The parameter set, parameter Adjust through training, represents time-dependent input, It means that at the time step The characteristic representation or state variable at the moment, represents the parameters of Fourier transform, It means that the neural network is used from time Update to next time The characteristics or state of being.
10. A tidal prediction system based on a neural network, characterized in that: include: An acquisition module is configured to: acquire historical tidal data of a target sea area; Preprocessing of historical tidal data; A decomposition module is configured to: perform time series decomposition on historical tidal data to obtain periodic features; Based on the periodic characteristics, the parameters of the tidal periodicity prediction model are determined; An input module is configured to: obtain tidal observation data, input the tidal observation data into a tidal periodicity prediction model, and obtain periodic characteristics of the tidal observation data; A fitting module is configured to: fit the periodic characteristics of the tidal observation data with the historical tidal data to obtain fitted tidal time series data; A decomposition module is configured to: perform modal decomposition on the fitted tidal time series data to obtain subsequences of different frequency components; The prediction module is configured to: input subsequences of different frequency components into the trained neural network, obtain the tidal data prediction value of the target sea area, and obtain the tidal change situation within the prediction period.
Citation Information
Patent Citations
Tide forecasting method based on multi-model combination
CN116822336A
River hydrodynamic agent model establishment method based on Fourier neural operator
CN117669424A
Module self-organizing tide combination forecasting method
CN117669835A
Sea area wave data deduction method and system based on neural network model
CN118503927A
New 4D (x, y, z, t) CO2-EOR rapid simulation method based on Fourier neural operator
CN118780068A
Cited By
Tidal flat shallow stratum deformation prediction device and method
CN119642774A
All-working-condition operation method and device of bidirectional tidal unit, electronic equipment and unit
CN119878435A
Ocean tide level simulation method and device based on hybrid neural network model
CN120354749A
Bridge safety assessment method based on multi-intelligent algorithm fusion
CN120408326A
Method for quantifying tidal non-stationarity
CN121614716A
