Wave height prediction method for multi-strategy snow ablation optimization

Abstract

The invention provides a wave height prediction method for multi-strategy snow ablation optimization, and belongs to the field of ocean engineering. The wave height prediction method comprises the following five steps: preprocessing sea condition data, performing feature extraction on the sea condition data by using a CNN model, training the model, optimizing parameters and evaluating the model. According to the method, the multi-strategy snow ablation optimization algorithm, the convolutional neural network, the bidirectional gating circulation unit and the multi-head attention mechanism are comprehensively utilized, and efficient processing and accurate prediction of the sea condition data are achieved; the introduction of the multi-strategy snow ablation optimization algorithm increases the diversity of the population, prevents the model from falling into local optimum, and improves the robustness of the algorithm. The CNN and the bidirectional GRU are combined, so that the model has excellent performance in the aspect of processing high-dimensional, nonlinear and time sequence data; the application of the attention mechanism can highlight important information, perform weighting processing on features, and further improve the prediction performance of the model.

Description

Technical Field

[0001] The invention relates to a wave height prediction method for multi-strategy snow melting optimization, which is applicable to the field of marine engineering technology. Background Art

[0002] In the field of marine engineering, accurate wave height prediction is crucial to the safety and efficiency of offshore operations. Wave height prediction can help plan marine transportation, construction and various marine activities, reduce risks and improve operational efficiency. In recent years, with the development of artificial intelligence and deep learning technologies, data-driven methods have shown broad application prospects in wave height prediction. These methods try to use historical data and various algorithms to improve the accuracy of predictions.

[0003] Although data-driven methods have made some progress in wave height prediction, existing methods still suffer from insufficient prediction accuracy when dealing with multi-dimensional, nonlinear sea state data. This may be due to the complexity of sea state data, including high-dimensional, nonlinear characteristics and dynamic changes in time series. A single optimization strategy and model may have difficulty in fully capturing all relevant features and patterns in the data, resulting in insufficient accuracy and robustness of the prediction results. Especially in complex marine environments, more advanced methods are needed to improve the reliability of predictions.

[0004] Existing methods may rely too much on traditional statistical or machine learning techniques, which may not be able to effectively handle large-scale, high-dimensional data sets, or have difficulty capturing deep features in the data. In wave height prediction, there may be a lack of capture of long-term dependencies in time series data, resulting in the prediction model being unable to accurately reflect the impact of past and future moments on the current moment.

[0005] In order to solve the above problems, this patent proposes a wave height prediction method that combines multiple optimization strategies and deep learning technologies, aiming to improve prediction accuracy, more effectively process sea condition data, and achieve accurate prediction. Summary of the invention

[0006] The present invention provides a wave height prediction method for multi-strategy snow melting optimization, aiming to address the shortcomings of existing methods in wave height prediction accuracy and robustness. The method comprehensively utilizes a multi-strategy snow melting optimization algorithm, a convolutional neural network, a bidirectional gated recurrent unit and a multi-head attention mechanism to achieve efficient processing and accurate prediction of sea condition data.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A multi-strategy snowmelt optimized wave height prediction method includes the following steps:

[0009] S100, sea condition data preprocessing: collecting sea condition data, first processing missing data and outliers, and then normalizing the data; the sea condition data includes ocean wave height, period, wave direction, wind speed and wind direction data;

[0010] S200, using the CNN model to extract features from the sea state data: setting CNN model parameters, feeding the normalized sea state data into the CNN network structure from the input layer, and obtaining a stable sea state data vector sequence;

[0011] S300, model training: performing wave height prediction based on the bidirectional GRU deep learning model, setting bidirectional GRU model parameters, inputting the stable sea state data vector sequence into the bidirectional GRU deep learning model, extracting the global features of the stable sea state data vector sequence through the bidirectional GRU deep learning model, and expanding the forward and backward wave heights for model training, that is:

[0012]

[0013] The ATTENTION mechanism is introduced to evaluate the impact of different global features on the wave height prediction results, and different weights are set for the global features, namely

[0014]

[0015] Where X t Represents the input data, i.e., a vector sequence of steady sea state data;

[0016] represents the forward wave height of the hidden state, represents the backward wave height of the hidden state;

[0017] α t Represents probability distribution, and softmax represents a function that converts a real number vector into a probability distribution;

[0018] W a Represents the weight matrix, which is used to weight the attention scores;

[0019] S400, parameter optimization: optimizing the bidirectional GRU model parameters by using a snow melting algorithm, and searching for an optimal parameter combination, including the following steps:

[0020] S401, determining a search space, initializing individuals in the search space using a good point set, and generating an initial population, wherein a single individual in the initial population represents a set of bidirectional GRU model parameters;

[0021] S402, introducing a dual population mechanism of the snow melting algorithm, respectively exploring and utilizing the initial population, where the total number of individuals in the initial population is N;

[0022] S403, in the exploration phase, the snow melting algorithm uses Brownian motion to simulate the irregular movement of steam to guide the search agent to explore potential areas, and the control equation of the center of mass position is:

[0023]

[0024] in, is the centroid position;

[0025] BMi(t) is a vector containing random numbers based on Gaussian distribution, representing Brownian motion; is the step-by-step multiplication;

[0026] θ1 is a randomly selected number between [0,1];

[0027] G(t) is the current best solution obtained by the search agent;

[0028] Elite(t) is an individual randomly selected from several elite groups of the initial population;

[0029] Z i (t) represents the i-th excellent individual in the initial population after iteration t; Z second (t) and Z third (t) represent the second and third excellent individuals in the initial population after exploration;

[0030] Z i (t+1) represents the i-th excellent individual in the initial population after t+1 iterations;

[0031] Z c (t) represents the centroid position of the individuals whose fitness values ​​are in the top 50%;

[0032] S404, in the utilization stage, the snow melting algorithm uses the degree-day method to simulate the snow melting process, and guides the search agent to find a high-quality solution near the current optimal solution. The snow melting rate control equation is:

[0033] M=DDF×(T-T1)

[0034] Among them, M is the snowmelt rate, which is the key parameter for simulating snowmelt behavior in the development stage; DDF is the degree-day factor, which represents the contribution of daily temperature to the snowmelt process; T represents the daily average temperature;

[0035] T1 is the reference temperature, usually set to 0, that is:

[0036]

[0037] Among them, t maxrepresents the termination time, and DDF represents the degree-day factor;

[0038] S405, overall position update: combining the strategies of the exploration phase and the utilization phase, as well as the dynamic changes of the population, to optimize the position of each individual in the initial population;

[0039] S406. Use the periodic oscillation mutation strategy to introduce mutations in the search process, escape from local traps and obtain the correct search mode, find the global optimal solution, and model the particle trajectory by solving the particle path equation and applying the second-order differential equation:

[0040]

[0041] Among them, v i,j (t) represents the velocity of the i-th particle in the j-th dimension at the t-th iteration; v i,j (t-1) represents the velocity of the i-th particle in the j-th dimension at the t-1-th iteration; φ1 and φ2 represent constant coefficients used to adjust the movement of particles; p i represents the individual best position of the i-th particle; p g represents the global optimal position; x i,j (t) represents the position of the i-th particle in the j-th dimension at the t-th iteration; x i,j (t-1) represents the position of the i-th particle in the j-th dimension at the t-1-th iteration;

[0042] By updating the position and velocity of the particles, and then optimizing the bidirectional GRU model parameters by simulating natural group behavior, an optimal bidirectional GRU model parameter combination is obtained;

[0043] S500, Evaluation Model: The mean absolute error and root mean square error are used to evaluate the performance of the bidirectional GRU deep learning model on the tide prediction task. The calculation formula is as follows:

[0044]

[0045] Where: n is the number of wave height data test sets; y i is the true value of wave height; is the predicted value of the wave height data test set sample.

[0046] The preferred technical solution of the present invention, the CNN model parameters include that the number of convolution kernels in the convolution layer is 64 and the size is 2×2; the activation function is the ReLU function; the size of the pooling layer is 1×1.

[0047] The preferred technical solution of the present invention, the bidirectional GRU model parameters include GRU parameters, ATTENTION parameters, sequence length, batch size and number of iterations; the GRU parameters include the number of hidden layer units and the number of GRU layers; the ATTENTION parameters include the number of attention heads and attention dimensions.

[0048] In the preferred technical solution of the present invention, the control equation of the coordinate value of a single initialized individual of the initial population is:

[0049]

[0050] Among them, a is the lower bound of the search space; b is the upper bound of the search space; x i is the coordinate value of the i-th initialization point, that is, the initial value of the i-th parameter; N is the total number of individuals in the initial population.

[0051] The preferred technical solution of the present invention is that the centroid position Z of the individuals whose fitness values ​​are in the top 50% is c The governing equation for (t) is,

[0052]

[0053] Among them, N1 is the number of leaders, that is, the number of the top N1 individuals with the best fitness in the current population.

[0054] In summary, compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. The present invention preprocesses the sea state data to ensure the quality of the input data and lays the foundation for subsequent feature extraction and model training; the CNN model is used to extract features from the preprocessed sea state data. The CNN model automatically learns local features in the data through its convolution layer and pooling layer, and converts these features into a stable sea state data vector sequence, with the purpose of converting the original data into a feature representation useful for the prediction model; the bidirectional GRU model is used to predict wave height, which can capture the long-term dependency in the time series data, extract global features through the forward and backward network structures, and introduce the attention mechanism to further enhance the model's attention to key features and optimize the prediction results; after the model training, that is, after the first iteration is completed, the snow ablation algorithm is used to optimize the parameters of the bidirectional GRU model, and the search process is guided by simulating natural phenomena such as Brownian motion and snow melting process to find the optimal combination of model parameters, combining the two-stage strategy of exploration and utilization, and the periodic oscillation mutation strategy to enhance the global search ability of the algorithm and avoid local optimality; finally, the mean absolute error and root mean square error indicators are used to evaluate the prediction performance of the model to verify the effectiveness and accuracy of the model.

[0056] 2. Parameter Optimization The snow melting algorithm is used to introduce a dual population mechanism, which divides the population into two sub-populations. One is responsible for "exploring" new areas to find potential solutions, and the other is responsible for "utilizing" the area near the currently found best solution to find a higher quality solution, which helps the algorithm achieve balance in the solution space and avoid falling into the local optimum. In the exploration phase, the algorithm uses Brownian Motion to simulate the irregular movement of steam or search agents, so that the search agents have highly dispersed characteristics, so that they can explore potential areas extensively. This stage promotes the search process by updating the centroid position and guiding elite individuals; in the utilization phase, the degree-day method is used to simulate the snow melting process, encouraging the search agent to conduct a detailed search near the current best solution to find a higher quality solution. Parameters such as snow melting rate and daily average temperature play a key role in the simulation process.

[0057] Since the snow melting algorithm uses an overall position update equation, combined with exploration and utilization strategies and population dynamics, the position of each individual is optimized, thereby continuously improving the quality of the solution. In order to overcome local optimality and escape local traps, a periodic oscillation mutation strategy is also introduced. By solving the particle path equation and applying the second-order differential equation to model the particle trajectory, the algorithm can grasp the general trend of convergence and escape from the local optimality.

[0058] 3. The present invention integrates multiple optimization strategies and deep learning technologies to improve the prediction accuracy of the model and effectively reduce the wave height prediction error; the introduction of a multi-strategy snow melting optimization algorithm increases the diversity of the population, avoids the model from falling into local optimality, and improves the robustness of the algorithm; the combination of CNN and bidirectional GRU makes the model perform well in processing high-dimensional, nonlinear and time series data; the application of the attention mechanism can highlight important information, perform weighted processing on features, and further improve the prediction performance of the model; at the same time, the application effect on measured data is good, and it has high practical value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a multi-strategy snow melting optimization wave height prediction method of the present invention;

[0060] Figure 2 A comparison diagram of wave height prediction and wave height measurement in a specific embodiment of the present invention;

[0061] Figure 3 This is a wave height prediction error analysis diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0062] The specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments given here are only used to illustrate and explain the present invention, and cannot be used to limit the present invention.

[0063] It should be noted that many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention may also have other implementations and variations thereof. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0064] The wave height prediction method of multi-strategy snow melting optimization of the present invention is used to predict the wave height of a certain sea area. As a specific embodiment of the present invention, a buoy is used to automatically record and transmit ocean data regularly according to design requirements. The collected ocean data is used as the basic data for wave height prediction to prepare for subsequent model training.

[0065] Refer to the flow chart, Figure 1 As shown, the present embodiment provides a multi-strategy snow melting optimization wave height prediction method, comprising the following steps:

[0066] S100, sea condition data preprocessing: collecting sea condition data, first processing missing data and outliers, and then normalizing the data; the sea condition data includes ocean wave height, period, wave direction, wind speed and wind direction data;

[0067] Since the acquired sea state data has data missing and data duplication problems, the data needs to be preprocessed. The missing data is processed by taking the average of adjacent data, and the duplicate data is deleted. Since the results of the sea state data are quite different, the data needs to be normalized before entering the neural network model, that is,

[0068]

[0069] Among them, x' is the normalized value; x is the actual value; x max 、x min are the maximum and minimum values ​​in the sea state data respectively;

[0070] S200, using the CNN model to extract features from the sea state data: setting CNN model parameters, feeding the normalized sea state data into the CNN network structure from the input layer, and obtaining a stable sea state data vector sequence;

[0071] The normalized sea state data is input into the CNN model. When processing data, CNN passes through the five parts of input layer, convolution layer, pooling layer, fully connected layer and output layer. The normalized sea state data enters the network structure from the input layer, and outputs multiple feature surfaces through the convolution layer. The value of each neuron in the feature surface is calculated by the corresponding convolution kernel. Then it passes through the pooling layer for data compression, and the compression result is a one-dimensional vector; the output result of the pooling layer is passed to the fully connected layer, which integrates the features extracted by the pooling layer of the previous convolution layer, and then updates the weights to achieve back propagation; by adjusting the size of the convolution kernel, the most suitable CNN model parameters are selected, and the CNN model parameters include the number of convolution kernels in the convolution layer is 64 and the size is 2×2; the activation function is the ReLU function; the size of the pooling layer is 1×1. Then use CNN to extract high-dimensional and nonlinear features of the sea state data, and pass the relatively stable information after processing as a whole into the bidirectional GRU, i.e., BiGRU network, for prediction.

[0072] GRU is a variant of the long short-term memory (LSTM) network. It simplifies the internal structure of LSTM. The core structure of the network is the update gate and the reset gate. The update gate combines the new candidate state To determine what information each hidden unit retains or forgets; the reset gate can determine the degree of retention of the hidden layer data at the previous moment. Through this structure, the GRU network can capture the temporal dependency in historical information, and its equation is:

[0073]

[0074] Among them, R t Represents the reset gate, which controls the amount of information of the previous node to be written into the current candidate set Z t represents the update gate, which controls the amount of information saved from the previous time node to the current time node; W r , W z , W h are the weight matrices of the reset gate, update gate and candidate state respectively; b r , b z , b h is the bias vector.

[0075] Compared with the ordinary GRU model, the bidirectional gated recurrent unit (BiGRU) is composed of two independent GRUs, namely the forward GRU and the reverse GRU. The forward GRU processes the input sequence and is responsible for capturing contextual information from the beginning of the sequence. The reverse GRU processes the input data from the end of the sequence and captures contextual information in reverse order. Compared with the unidirectional GRU, the propagation from front to back can better reflect the impact of past and future moments on the current moment.

[0076] S300, model training: performing wave height prediction based on the bidirectional GRU deep learning model, setting bidirectional GRU model parameters, inputting the stable sea state data vector sequence into the bidirectional GRU deep learning model, extracting the global features of the stable sea state data vector sequence through the bidirectional GRU deep learning model, and expanding the forward and backward wave heights for model training, that is:

[0077]

[0078] The ATTENTION mechanism is introduced to evaluate the impact of different global features on the wave height prediction results, and different weights are set for the global features, namely

[0079]

[0080] Where X t Represents the input data, i.e., a vector sequence of steady sea state data;

[0081] represents the forward wave height of the hidden state, represents the backward wave height of the hidden state;

[0082] α t Represents probability distribution, and softmax represents a function that converts a real number vector into a probability distribution;

[0083] W a Represents the weight matrix, which is used to weight the attention scores;

[0084] The above is because in a long time series, it is inevitable that multiple features have different effects on the prediction results. In order to evaluate the impact of each feature on the output result, the ATTENTION mechanism is introduced, which can compare the impact of different features through the input and output results and set different weights for the features. After a series of calculations and weighted processing, different weights are given to different parts of the sea state vector, so that the model can better focus on the relevant parts of the current task.

[0085] The bidirectional GRU model parameters in this embodiment include GRU parameters, ATTENTION parameters, sequence length, batch size and number of iterations; the GRU parameters include the number of hidden layer units and the number of GRU layers; the ATTENTION parameters include the number of attention heads and the attention dimension; in this embodiment, the number of hidden layer units is set to 512 or 1024; the number of GRU layers is 2 to 3; the number of attention heads is a power of 2, such as 2, 4 or 8; the attention dimension matches the number of hidden layer units.

[0086] S400, parameter optimization: optimizing the bidirectional GRU model parameters by using a snow melting algorithm, and searching for an optimal parameter combination, including the following steps:

[0087] S401, determining a search space, initializing individuals in the search space using a good point set, and generating an initial population, wherein a single individual in the initial population represents a set of bidirectional GRU model parameters;

[0088] The control equation of the coordinate value of a single initialized individual of the initial population is:

[0089]

[0090] Among them, a is the lower bound of the search space; b is the upper bound of the search space; x i is the coordinate value of the i-th initialization point, that is, the initial value of the i-th parameter; N is the total number of individuals in the initial population.

[0091] Using the optimal point set to evenly distribute the search for individuals during the initialization process can increase the diversity of the population and help get rid of the attraction of local minimum values.

[0092] S402, introducing a dual population mechanism of the snow melting algorithm, respectively exploring and utilizing the initial population to achieve a balance in the solution space, the total number of individuals in the initial population being N;

[0093] S403. In the exploration phase, the snow melting algorithm uses Brownian motion to simulate the irregular movement of steam, so that the search agent has a highly dispersed feature, and guides the search agent to explore potential areas. The control equation of the center of mass position is:

[0094]

[0095] in, is the centroid position;

[0096] BMi(t) is a vector containing random numbers based on Gaussian distribution, representing Brownian motion; it is used in the exploration phase of the algorithm to increase the dispersion and randomness of the search by simulating the irregular motion of steam (or particles), helping the algorithm to explore the potential solution space. Since BMi(t) is based on Gaussian distribution, it contains randomness, which helps the algorithm to jump out of the local optimal solution and find the global optimal solution.

[0097] is the step-by-step multiplication;

[0098] θ1 is a randomly selected number between [0,1]; as a random factor, it is used to control the randomness and degree of exploration. It can help the algorithm explore more extensively in the search space and avoid falling into local optimality. Or it is used to adjust the balance between randomness and certainty in the exploration process, so that the search agent can move closer to possible better solutions while maintaining diversity;

[0099] G(t) is the current best solution found by the search agent. In the optimization process, G(t) represents the best solution found so far. It is an important reference point in the algorithm search process, used to guide the search agent to approach the known best solution and promote the convergence of the algorithm.

[0100] Elite(t) is an individual randomly selected from several elite groups of the initial population; these elite individuals represent the best performing solutions in the current population. By introducing the information of these elites, they are used to further guide the search process and accelerate the convergence to the optimal solution. The algorithm can explore and utilize the solution space more effectively.

[0101] Z i (t) represents the i-th excellent individual in the initial population after iteration t; Z second (t) and Z third (t) represent the second and third excellent individuals in the initial population after exploration;

[0102] Z i (t+1) represents the i-th excellent individual in the initial population after t+1 iterations;

[0103] Z c (t) represents the centroid position of the individuals whose fitness values ​​are in the top 50%. It is a weighted average value that reflects the central position of the better individuals in the current population. Its control equation is:

[0104]

[0105] Among them, N1 is the number of leaders, that is, the number of the top N1 individuals with the best fitness in the current population;

[0106] S404, in the utilization stage, the snow melting algorithm uses the degree-day method to simulate the snow melting process, and guides the search agent to find a high-quality solution near the current optimal solution. The snow melting rate control equation is:

[0107] M=DDF×(T-T1)

[0108] Among them, M is the snowmelt rate, which is the key parameter for simulating snowmelt behavior in the development stage; DDF is the degree-day factor, which represents the contribution of daily temperature to the snowmelt process; T represents the daily average temperature;

[0109] T1 is the reference temperature, usually set to 0, that is:

[0110]

[0111] Among them, t max represents the termination time, and DDF represents the degree-day factor;

[0112] The degree-day method is used to estimate the amount of snow or ice that would melt due to temperatures above a certain threshold (usually 0°C) within a certain period of time. DDF is the degree-day factor, a proportional constant used to convert temperature differences into snowmelt rates.

[0113] S405, overall position update: combining the strategies of the exploration phase and the utilization phase, as well as the dynamic changes of the population, to optimize the position of each individual in the initial population;

[0114] S406. Use the periodic oscillation mutation strategy to introduce mutations in the search process, escape from local traps and obtain the correct search mode, find the global optimal solution, and model the particle trajectory by solving the particle path equation and applying the second-order differential equation:

[0115]

[0116] Among them, v i,j (t) represents the velocity of the i-th particle in the j-th dimension at the t-th iteration; v i,j (t-1) represents the velocity of the i-th particle in the j-th dimension at the t-1-th iteration; φ1 and φ2 represent constant coefficients used to adjust the movement of particles; p i represents the individual best position of the i-th particle; p g represents the global optimal position; x i,j (t) represents the position of the i-th particle in the j-th dimension at the t-th iteration; x i,j (t-1) represents the position of the i-th particle in the j-th dimension at the t-1-th iteration;

[0117] By updating the position and velocity of the particles, and then optimizing the bidirectional GRU model parameters by simulating natural group behavior, an optimal bidirectional GRU model parameter combination is obtained;

[0118] S500, Evaluation Model: The mean absolute error and root mean square error are used to evaluate the performance of the bidirectional GRU deep learning model on the tide prediction task. The calculation formula is as follows:

[0119]

[0120] Where: n is the number of wave height data test sets; y i is the true value of wave height; is the predicted value of the wave height data test set sample.

[0121] The sea condition data of a certain sea area from September 15 to November 15 for two months were collected as the measured value of wave height. The result of obtaining the predicted value of wave height by using the wave height prediction method of this embodiment is as follows: Figure 2 As shown in the figure, it is obvious that the predicted value is basically consistent with the measured value of wave height;

[0122] like Figure 3 As shown in the figure, the MAE mean absolute error value is calculated by the evaluation model calculation formula to be 0.0509m, and the MSE root mean square error value is 0.0049m 2 , it can be seen that the average difference between the predicted value and the true value is a relatively small error, indicating that the model has a high prediction accuracy; the MSE value is smaller than the MAE value, and there are no large outliers. The evaluation results of the model show that the bidirectional GRU deep learning model can capture the main features of sea state data and make predictions effectively. It has good performance in the tide level prediction task, and the error in predicting wave height is controllable and the error level is low, which can provide valuable reference for decision-making in related fields.

[0123] The above description is a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-strategy snowmelt optimized wave height prediction method, characterized in that: The following steps are involved: S100, sea condition data preprocessing: collecting sea condition data, first processing missing data and outliers, and then normalizing the data; the sea condition data includes ocean wave height, period, wave direction, wind speed and wind direction data; S200, using the CNN model to extract features from the sea state data: setting CNN model parameters, feeding the normalized sea state data into the CNN network structure from the input layer, and obtaining a stable sea state data vector sequence; S300, model training: performing wave height prediction based on the bidirectional GRU deep learning model, setting bidirectional GRU model parameters, inputting the stable sea state data vector sequence into the bidirectional GRU deep learning model, extracting the global features of the stable sea state data vector sequence through the bidirectional GRU deep learning model, and expanding the forward and backward wave heights for model training, that is: The ATTENTION mechanism is introduced to evaluate the impact of different global features on the wave height prediction results, and different weights are set for the global features, namely Where X t Represents the input data, i.e., a vector sequence of steady sea state data; represents the forward wave height of the hidden state, represents the backward wave height of the hidden state; α t Represents probability distribution, and softmax represents a function that converts a real number vector into a probability distribution; W a Represents the weight matrix, which is used to weight the attention scores; S400, parameter optimization: optimizing the bidirectional GRU model parameters by using a snow melting algorithm, and searching for an optimal parameter combination, including the following steps: S401, determining a search space, initializing individuals in the search space using a good point set, and generating an initial population, wherein a single individual in the initial population represents a set of bidirectional GRU model parameters; S402, introducing a dual population mechanism of the snow melting algorithm, respectively exploring and utilizing the initial population, where the total number of individuals in the initial population is N; S403, in the exploration phase, the snow melting algorithm uses Brownian motion to simulate the irregular movement of steam to guide the search agent to explore potential areas, and the control equation of the center of mass position is: in, is the centroid position; BMi(t) is a vector containing random numbers based on Gaussian distribution, representing Brownian motion; is the step-by-step multiplication; θ1 is a randomly selected number between [0,1]; G(t) is the current best solution obtained by the search agent; Elite(t) is an individual randomly selected from several elite groups of the initial population; Z i (t) represents the i-th excellent individual in the initial population after iteration t; Z second (t) and Z third (t) represent the second and third excellent individuals in the initial population after exploration; Z i (t+1) represents the i-th excellent individual in the initial population after t+1 iterations; Z c (t) represents the centroid position of the individuals whose fitness values ​​are in the top 50%; S404, in the utilization stage, the snow melting algorithm uses the degree-day method to simulate the snow melting process, and guides the search agent to find a high-quality solution near the current optimal solution. The snow melting rate control equation is: M=DDF×(T-T1) Among them, M is the snowmelt rate, which is the key parameter for simulating snowmelt behavior in the development stage; DDF is the degree-day factor, which represents the contribution of daily temperature to the snowmelt process; T represents the daily average temperature; T1 is the reference temperature, usually set to 0, that is: Among them, t max represents the termination time, and DDF represents the degree-day factor; S405, overall position update: combining the strategies of the exploration phase and the utilization phase, as well as the dynamic changes of the population, to optimize the position of each individual in the initial population; S406. Use the periodic oscillation mutation strategy to introduce mutations in the search process, escape from local traps and obtain the correct search mode, find the global optimal solution, and model the particle trajectory by solving the particle path equation and applying the second-order differential equation: Among them, v i,j (t) represents the velocity of the i-th particle in the j-th dimension at the t-th iteration; v i,j (t-1) represents the velocity of the i-th particle in the j-th dimension at the t-1-th iteration; φ1 and φ2 represent constant coefficients used to adjust the movement of particles; p i represents the individual best position of the i-th particle; p g represents the global optimal position; x i,j (t) represents the position of the i-th particle in the j-th dimension at the t-th iteration; x i,j (t-1) represents the position of the i-th particle in the j-th dimension at the t-1-th iteration; By updating the position and velocity of the particles, and then optimizing the bidirectional GRU model parameters by simulating natural group behavior, an optimal bidirectional GRU model parameter combination is obtained; S500, Evaluation Model: The mean absolute error and root mean square error are used to evaluate the performance of the bidirectional GRU deep learning model on the tide prediction task. The calculation formula is as follows: Where: n is the number of wave height data test sets; y i is the true value of wave height; is the predicted value of the wave height data test set sample.

2. The multi-strategy snowmelt optimized wave height prediction method according to claim 1 is characterized by: The CNN model parameters include that the number of convolution kernels in the convolution layer is 64 and the size is 2×2; the activation function is the ReLU function; and the size of the pooling layer is 1×1.

3. The multi-strategy snowmelt optimized wave height prediction method according to claim 1 is characterized by: The bidirectional GRU model parameters include GRU parameters, ATTENTION parameters, sequence length, batch size and number of iterations; the GRU parameters include the number of hidden layer units and the number of GRU layers; the ATTENTION parameters include the number of attention heads and attention dimensions.

4. The multi-strategy snowmelt optimized wave height prediction method according to claim 1, characterized in that: The control equation of the coordinate value of a single initialized individual of the initial population is: Among them, a is the lower bound of the search space; b is the upper bound of the search space; x i is the coordinate value of the i-th initialization point, that is, the initial value of the i-th parameter; N is the total number of individuals in the initial population.

5. The multi-strategy snowmelt optimized wave height prediction method according to claim 1 is characterized by: The fitness value is located in the centroid position Z of the top 50% of the individuals c The governing equation for (t) is, Among them, N1 is the number of leaders, that is, the number of the top N1 individuals with the best fitness in the current population.

Images