CN118395106A - Green tide coverage area forecasting method based on deep learning - Google Patents

Abstract

The invention discloses a green tide coverage area forecasting method based on deep learning, and belongs to the field of ocean forecasting. The method comprises the following steps: a. determining a forecasting area and forecasting time effect; b. collecting green tide coverage area historical multisource fusion data and atmospheric and ocean multi-element historical observation data; c. determining a training set and a testing set based on the long-short memory neural network, and designing factor combinations and parameter setting experimental schemes; d. training a model to obtain a forecast model of different factor combinations; e. performing forecasting effect inspection, and selecting an optimal forecasting factor combination; f. further adopting a parameter setting experimental scheme to optimize a training model; g. performing forecasting effect inspection, and selecting optimal parameter setting to obtain an optimal forecasting model of the green tide coverage area; h, inputting an actual measurement forecasting factor to obtain a green tide coverage area forecasting result. The method can forecast the green tide coverage area for 3 days, and provides effective auxiliary decision information for green tide disaster prevention and control work.

Description

A method for predicting green tide coverage area based on deep learning

Technical Field

The present invention relates to the field of ocean forecasting, and in particular to a method for forecasting green tide coverage area based on deep learning.

Background technique

Green tide refers to a phenomenon in which some large algae (such as Enteromorpha) in the ocean float, proliferate or aggregate to a certain level under certain environmental conditions, resulting in abnormal marine ecological environment. Green tide outbreaks are becoming more and more frequent in coastal waters around the world and have become a global marine disaster. Due to different atmospheric and marine environments, the development scale of green tides in different years varies significantly. Especially in recent years, the increase in extreme weather in the offshore marine environment has increased the uncertainty of green tide development. Studying the influencing factors of green tide development and making timely and accurate forecasts of the green tide coverage area in the Yellow Sea can provide effective auxiliary decision-making information for relevant departments in green tide prevention and control and pre-salvage work, which is of great significance to reducing the impact and losses caused by green tide disasters.

At present, some scholars have conducted some preliminary research on the influencing factors and prediction and forecasting technologies of green tide. "Numerical simulation of emergency drift of green tide in the Yellow Sea", Huang Juan et al., Ocean Forecast, published in 2011, records the emergency prediction of the drift trajectory of green tide based on the three-dimensional full-dynamic POM ocean model and the multi-source measured and monitored data of the Yellow Sea green tide from 2008 to 2009, using the Lagrangian particle tracking method. The drift trajectory of the green tide in the Yellow Sea obtained is closely related to the numerical simulation results. The Chinese invention patent with application publication number CN116467565A and publication date of July 21, 2023 discloses a method for predicting the optimal search area of green tide patches of Enteromorpha. This method takes into account the uncertainty of drift, aggregation and splitting of Enteromorpha patches under the influence of environmental factors such as wind and waves, constructs a Monte Carlo probability drift model of Enteromorpha green tide patches, and then predicts the optimal search area of Enteromorpha green tide patches. However, the generation and development of green tides are affected by multiple factors. Currently, its outbreak mechanism and the coupling between various factors are still unclear. Physical or ecological models alone are not sufficient to describe its occurrence and development laws.

Summary of the invention

Based on the above technical problems, the present invention proposes a green tide coverage area forecasting method based on deep learning.

The technical solution adopted by the present invention is:

A method for predicting green tide coverage area based on deep learning comprises the following steps:

a. Determine the forecast area and forecast time limit;

b. Collect historical multi-source fusion data on green tide coverage in the forecast area and historical observation data on multiple elements of the atmosphere and ocean;

c. Use the long short-term memory neural network model, determine the training set and test set, and design the factor combination and parameter setting experimental plan of the long short-term memory neural network model;

d. According to the factor combination experimental scheme designed in step c, the long short-term memory neural network models are trained one by one to obtain the neural network prediction models of green tide coverage area with different factor combinations;

e. Testing the forecasting effect of the neural network forecasting model for green tide coverage area obtained by training in step d, and selecting the optimal forecasting factor combination of the neural network forecasting model for green tide coverage area according to the test results;

f. Based on the optimal prediction factor combination obtained in step e, the parameter setting experimental scheme in step c is adopted to optimize the trained neural network prediction model of green tide coverage area;

g. Testing the prediction effect of the neural network prediction model for green tide coverage area obtained by training in step f, and selecting the optimal parameter setting of the neural network prediction model for green tide coverage area according to the test results to obtain the optimal neural network prediction model for green tide coverage area;

h. According to the optimal neural network prediction model for green tide coverage area obtained in step g, input the measured prediction factors to obtain the green tide coverage area prediction results.

The beneficial technical effects of the present invention are:

The present invention proposes a method for forecasting green tide coverage area based on deep learning. The method is based on the deep learning technology of long short-term memory neural network (LSTM), and uses the measured data of green tide coverage area and the measured data of multiple factors of atmosphere and ocean environment to construct an optimal forecast model for green tide coverage area, and forecast the green tide coverage area for the next three days on a daily basis. The present invention has the advantages of accurate forecast results and simple forecast process. The present invention can predict the impact area and degree of green tide in advance, and then provide effective auxiliary decision-making information for the prevention and control of green tide disasters such as scientifically demarcating green tide nearshore salvage areas and formulating efficient salvage plans.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

FIG1 is a flow chart of an implementation method of a green tide coverage area forecasting method based on deep learning according to the present invention;

FIG2 shows an error test diagram of the Yellow Sea green tide coverage area forecast model in the first round of experiments in a specific application example of the present invention (in each group of experimental bar graphs, starting from the left, the first three are MAE; the last three are RMSE, and from left to right they represent the error values of 1d, 2d, and 3d respectively);

Fig. 3 shows the error test diagram of the prediction model of the green tide coverage area in the Yellow Sea in the second round of experiments in a specific application example of the present invention (in each group of experimental bar graphs, starting from the left, the first three are MAE; the last three are RMSE, and from left to right they represent the error values of 1d, 2d, and 3d respectively);

Figure 4 is the root mean square error result of model training;

Figure 5 is a graph showing the model training loss function results.

Detailed ways

The development process of green tide is affected by multiple factors. At present, its outbreak mechanism and the coupling between various factors are still unclear. Physical or ecological models alone are not enough to describe its occurrence and development laws. Therefore, deep learning methods are considered to solve the problem of green tide forecasting. Based on this, the present invention creatively uses the LSTM method to construct a forecast model for the evolution of green tide coverage area to achieve business forecasting.

As shown in FIG1 , a method for predicting green tide coverage area based on deep learning includes the following steps:

a. Determine the forecast area and forecast period.

The forecast area is selected as 32-37°N, 119-124°E, and the forecast period is selected as 3 days.

b. Collect historical multi-source fusion data on green tide coverage in the forecast area and historical observation data on multiple elements of the atmosphere and ocean.

The data collected above are from 2010 to 2023. The elements of the historical atmospheric observation data include air temperature, precipitation, shortwave radiation and 10m height wind vector, and the elements of the historical ocean observation data include water temperature, salinity and surface current vector.

c. Use the long short-term memory neural network model, determine the training set and test set, and design the factor combination and parameter setting experimental plan of the long short-term memory neural network model.

The specific steps include:

c1. Determine the long short-term memory neural network model (LSTM) as the training model.

c2. Use the historical multi-source fusion data of green tide coverage area in the forecast area and the historical observation data of multiple elements of atmosphere and ocean collected in step b as raw data, and process the raw data into a format that can be read by the LSTM training model.

c3. Divide the processed data in step c2 into two time periods, determine the training set and the test set, which are used for model training and model testing respectively.

The data from 2010 to 2021 are selected as the training set, and the data from 2022 to 2023 are selected as the test set.

c4. Use different factor combinations and neural network parameter combinations to design experimental plans.

Step c2 includes the following steps:

c21. Read and process historical multi-source fusion data of green tide coverage area in the forecast area.

Extract the green tide coverage area in the forecast area from the multi-source fusion data of green tide; read in the source data; and standardize the data.

;

Among them, _yi represents the green tide coverage area on the ith day, is the average green tide coverage area, is the standard deviation of green tide coverage area.

c22. Reading and processing of historical observation data of multiple elements of atmosphere and ocean.

Read the original observation data of multiple elements of the atmosphere and ocean in the forecast area; extract the element values and process the data into a matrix form of m×n; where m is the number of forecast factors and n is the number of samples; standardize each element.

The standardization formula is as follows:

;

Among them, _xij represents the i-th sample of the j-th predictor, represents the average value of the j-th predictor, represents the standard deviation of the j-th predictor.

c23. Other data reading and processing;

Convert the date corresponding to the observation data into numerical form. For example, the numerical form of January 1 is 1, the numerical form of February 1 is 32, and so on. Standardize the date values:

;

Among them, z _i represents the numerical form of the date of the i-th day, is the average value of the date, is the standard deviation of the date values.

Step c4 includes the following steps:

c41. Design experimental plans based on different factor combinations.

The prediction factors are divided into three categories: the first category is mandatory factors, including the measured value of green tide coverage area; the second category is optional factors, including forecast date; the third category is environmental factors, including temperature, precipitation, shortwave radiation, 10m height wind vector, water temperature, salinity and surface flow vector; the experimental design method is: mandatory factors + environmental factor combination (+ optional factors); that is, mandatory factors + environmental factor combination, or mandatory factors + environmental factor + optional factor combination.

c42. Design experimental plan based on parameter adjustment.

The parameter settings of the long short-term memory neural network include the input factor size, response factor size, number of hidden units, maximum number of cycles and original learning rate. The number of hidden units, maximum number of cycles and original learning rate are adjusted and the experimental plan is designed.

d. According to the factor combination experimental scheme designed in step c, the long short-term memory neural network models are trained one by one to obtain the neural network prediction models of green tide coverage area with different factor combinations. Each factor combination is trained multiple times, and the long short-term memory neural network model obtained by training is saved.

e. Test the prediction effect of the green tide coverage area neural network prediction model trained in step d, and select the optimal prediction factor combination of the green tide coverage area neural network prediction model according to the test results. Specifically, calculate the error of the long short-term memory neural network model, and select the factor combination with the smallest error for the next step of model optimization.

f. Based on the optimal prediction factor combination obtained in step e, the parameter setting experimental scheme in step c is adopted to optimize the trained green tide coverage area neural network prediction model. Each parameter setting scheme is trained multiple times, and the trained long short-term memory neural network model is saved.

g. Test the prediction effect of the neural network prediction model for green tide coverage area obtained by training in step f, and select the optimal parameter setting of the neural network prediction model for green tide coverage area according to the test results to obtain the optimal neural network prediction model for green tide coverage area. Specifically, calculate the error of the obtained long short-term memory neural network model, and select the one with the smallest error as the optimal neural network prediction model for green tide coverage area.

h. According to the optimal neural network prediction model for green tide coverage area obtained in step g, input the measured prediction factors to obtain the green tide coverage area prediction results.

In the above method, for the default values of the historical multi-source fusion data of green tide coverage area in the forecast area and the historical observation data of multiple elements of atmosphere and ocean, the interpolation method can be used to fill in the data.

In the above method, the principle of designing the experiment is to first conduct the experiment with a single factor, and then select the factor combination with better predictive effect to conduct a multi-factor experiment.

During the test process, the test factors are input, and the long short-term memory neural network model is used to predict the green tide coverage area to obtain the forecast results, which are then inversely standardized.

;

Among them, _yi represents the green tide coverage area, _yi ' represents the result of the model output green tide coverage area forecast, is the standard deviation of green tide coverage area, is the average value of green tide coverage area; calculate the forecast error, including the mean absolute error and root mean square error.

The present invention will be further described below in conjunction with specific application examples:

The Yellow Sea green tide usually occurs from April to August each year. It is scattered in mid-April, and erupts on a large scale in May. From mid-June to mid-July, the green tide reaches its peak area and begins to land on the southern coast of Shandong Peninsula. After that, the area begins to decrease and gradually disappears around August.

Collect multi-source fusion monitoring data of the Yellow Sea green tide in the forecast area for a total of 14 years from 2010 to 2023, and extract daily monitoring data of the green tide coverage area. At the same time, collect daily reanalysis data from 2010 to 2023 in the forecast area, refer to the research results of relevant literature, and collect elements that may be related to the occurrence and disappearance of green tides. Various atmospheric and oceanic elements, including temperature, precipitation, shortwave radiation, 10m height wind vector, water temperature, salinity and surface current vector, are collected.

The data were quality controlled and format converted. For the default values of green tide coverage area and atmospheric and oceanic element data, the data were filled by interpolation. The observation and forecast dates were digitized and converted into a matrix form m×n with the green tide coverage area and atmospheric and oceanic element data. Among them, m is the number of forecast factors and n is the number of samples.

Normalize the data in matrix form:

;

Among them, _xij represents the i-th sample of the j-th predictor, represents the average value of the j-th predictor, represents the standard deviation of the j-th predictor.

After standardization, the data are divided into two periods: 12 years of data from 2010 to 2021 are selected for training and model establishment. Taking into account the birth and disappearance time of the green tide in the Yellow Sea, May 1st to August 31st of each year are selected as valid data, totaling 1,476 days; 2 years of data from 2022 to 2023 are selected for model testing, and May 1st to August 31st of each year are selected as valid data, totaling 246 days.

The prediction factors used to construct the hourly prediction model are shown in Table 1.

Table 1

In Table 1, t is the model running time, and t+1 is the first time to start forecasting.

When training the green tide coverage area forecast model, one factor (combination) is selected for experiment each time, each factor (combination) is trained 10 times, and the trained neural network is saved.

The principle of designing experiments is to first conduct experiments with a single factor, and then select factor combinations with better forecasting effects to conduct multi-factor experiments. Table 2 shows the factor combinations of the first round of design experiments. Experiment 1-1 is a training with only the green tide coverage area, and experiments 1-2 to 1-12 are training with the green tide coverage area combined with atmospheric and marine environmental factors. Among them, wind and current are synthesized by latitude and longitude, and can also be combined to be regarded as one factor. In summary, a total of 12 groups of experiments were designed in the first round, each group was trained 10 times, and a total of 120 times.

Table 2

The training process is as follows:

1. Model parameter setting;

Set the training period. As mentioned above, set the first 1476 days of observation data as the training set and set the number of repeated training (10 times). Set the training plan according to Table 2 (Experiment 1-1 to Experiment 1-12).

LSTM neural network parameter settings, including input factor size, response factor size, number of hidden units, maximum number of cycles, original learning rate, etc.

2. Extraction and processing of prediction factors;

According to the training scheme, the predictors are extracted from the standardized data matrix and written into a matrix of the form pq× _k1 .

p represents the number of prediction factors; q represents the number of days of observation data, and the preferred embodiment of the present invention is q=3, which are the observation data of (t-2), (t-1), and t days respectively; _k1 represents the number of training samples, and as mentioned above, _k1 =1476.

From the standardized data matrix, extract the prediction target (label) and write it into a matrix of r× _k1 form.

r represents the forecast time, and the preferred embodiment of the present invention is r=3 (r represents the forecast days), which are the green tide coverage area data of (t+1), (t+2), and (t+3) days respectively; _k1 represents the number of training samples, as mentioned above, _k1 =1476.

3. Training and storage;

Read data; start neural network training, the model training results are shown in Figures 4 and 5; store the trained neural network and parameters.

The trained LSTM neural network is used for prediction test, and the test variables are mean absolute error (MAE) and root mean square error (RMSE).

The inspection process is as follows:

1. Model parameter setting;

Set the test period. In this experiment, the measured data from 2022 to 2023 are used for testing (a total of 246 days).

Set the LSTM neural network parameters consistent with those during training, including input factor size, response factor size, number of hidden units, maximum number of cycles, original learning rate, etc.

2. Data extraction and processing;

According to the factor combination of the test model, the predictive factors are extracted from the standardized data matrix and written into a matrix in the form of pq× _k2 .

p represents the number of prediction factors; q represents the number of days of observation data, and the preferred embodiment of the present invention is that q=3, which are the observation data of (t-2), (t-1), and t days respectively; k ₂ represents the number of test samples, and as mentioned above, k ₂ =246.

From the standardized data matrix, extract the prediction target (label) and write it into a matrix of r× _k2 form.

r represents the forecast time, and the preferred embodiment of the present invention is r=3, which are the green tide coverage area data of (t+1), (t+2), and (t+3) days respectively; k ₂ represents the number of test samples, and as mentioned above, k ₂ =246.

3. Inspection and storage;

Read the trained neural network model and parameters.

Input the test factor and use the read neural network to predict the green tide coverage area; get the forecast result and perform inverse standardization on the result.

;

Where _yi represents the coverage area of the Yellow Sea green tide, is the standard deviation of green tide coverage area, is the average green tide coverage area.

After obtaining the forecast value of the green tide coverage area, calculate the forecast error (mean absolute error MAE and root mean square error RMSE); store the experimental plan and its error.

table 3

Table 3 and Figure 2 show the forecast test results of each group of experiments, among which Table 3 shows the daily forecast errors of green tide coverage area in each group of experiments. From the forecast test results (MAE and RMSE), in Experiment 1-1, when only the observed value of green tide coverage area was used as the forecast factor, the average absolute errors of the three days were 18.01 km ² , 26.13 km ² , and 34.03 km ² , respectively, and the root mean square errors were 38.68 km ² , 55.66 km ² , and 70.5 km ² , respectively. After adding atmospheric and marine environmental factors, only the errors of Experiments 1-2, 1-3, and 1-8 were reduced, indicating that adding air temperature, sea temperature, and precipitation factors is beneficial to improving the forecast effect, while adding other factors is prone to generate redundant information.

On this basis, a second round of experiments was designed to arrange and combine the three factors of air temperature, sea temperature and precipitation, and compare their forecasting effects, as shown in Table 4.

Table 4

The same method as the first round of experiments was used to train and test the model, and the model errors are shown in Table 5 and Figure 3. From the test results, the error of the combination of sea temperature and precipitation is the smallest, and the error of the combination of temperature and precipitation is basically the same as that of sea temperature and precipitation. Considering the temperature, sea temperature and precipitation factors at the same time, the error does not increase significantly. Therefore, the optimal combination of atmospheric and marine environmental factors is sea temperature and precipitation.

table 5

On the basis of the optimal factor combination, the optional factor of date is added, and the model is trained again. The average absolute errors of the three days are 15.83 km ² , 19.57 km ² , and 23.20 km ² , and the root mean square errors are 30.76 km ² , 38.91 km ² , and 45.79 km ² , respectively, which are further reduced than those in Experiment 2-3. It can be seen that adding the date factor can improve the prediction accuracy of the model.

Based on the above, the optimal factor combination of the green tide coverage area prediction model is green tide coverage area + forecast area sea temperature + forecast area precipitation + forecast date.

Next, we conducted parameter adjustment experiments on LSTM. The parameter setting experiments are shown in Table 6. The parameter combination with the smallest error was selected as the optimal prediction model. The test results show that when the number of hidden layer units is 100, the initial learning rate is 0.01, and the maximum number of iterations is 50, the prediction error is the smallest. The average absolute errors for three days are 15.25 km ² , 19.28 km ² , and 23.34 km ² , respectively, and the root mean square errors are 28.82 km ² , 36.60 km ² , and 44.19 km ² , respectively.

Table 6

In summary, the optimal forecast model for the Yellow Sea coverage area is obtained. The optimal factor combination is: the observed values of the Yellow Sea green tide coverage area, surface sea temperature, precipitable water, and forecast date. The optimal parameter settings are: the number of hidden layer units is 100, the initial learning rate is 0.01, and the maximum number of iterations is 50. The average absolute errors of the optimal model for three days are 15.25 km ² , 19.28 km ² , and 23.34 km ² , respectively, and the root mean square errors are 28.82 km ² , 36.60 km ² , and 44.19 km ² , respectively.

Furthermore, the optimal forecast neural network model for the coverage area of the Yellow Sea green tide can be set to run automatically on a daily basis on the server, automatically reading the measured data of the coverage area of the green tide, as well as the measured data of sea temperature and precipitation, and calculating the forecast results of the coverage area of the Yellow Sea green tide in the next three days, thus realizing the automatic business operation of the whole process. The running time of the entire process model is within 1 minute, which can greatly improve the work efficiency during the outbreak of green tide disasters and provide auxiliary decision-making information for green tide disaster prevention and mitigation work.

The parts not mentioned in the above methods can be realized by adopting or drawing on existing technologies.

It should be noted that any equivalent substitution or obvious modification made by those skilled in the art under the guidance of this specification should be within the protection scope of the present invention.

Claims (10)

1. A method for predicting green tide coverage area based on deep learning, characterized by comprising the following steps: a. Determine the forecast area and forecast time limit; b. Collect historical multi-source fusion data on green tide coverage in the forecast area and historical observation data on multiple elements of the atmosphere and ocean; c. Use the long short-term memory neural network model, determine the training set and test set, and design the factor combination and parameter setting experimental plan of the long short-term memory neural network model; d. According to the factor combination experimental scheme designed in step c, the long short-term memory neural network models are trained one by one to obtain the neural network prediction models of green tide coverage area with different factor combinations; e. Testing the forecasting effect of the neural network forecasting model for green tide coverage area obtained by training in step d, and selecting the optimal forecasting factor combination of the neural network forecasting model for green tide coverage area according to the test results; f. Based on the optimal prediction factor combination obtained in step e, the parameter setting experimental scheme in step c is adopted to optimize the trained neural network prediction model of green tide coverage area; g. Testing the prediction effect of the neural network prediction model for green tide coverage area obtained by training in step f, and selecting the optimal parameter setting of the neural network prediction model for green tide coverage area according to the test results to obtain the optimal neural network prediction model for green tide coverage area; h. According to the optimal neural network prediction model for green tide coverage area obtained in step g, input the measured prediction factors to obtain the green tide coverage area prediction results. 2. According to a method for forecasting green tide coverage area based on deep learning according to claim 1, it is characterized in that in step a: the forecast area is selected as 32-37°N, 119-124°E, and the forecast time is selected as 3 days. 3. According to a green tide coverage area forecasting method based on deep learning according to claim 1, it is characterized in that in step b: the time period for collecting historical multi-source fusion data of green tide coverage area in the forecast area and historical observation data of multiple elements of atmosphere and ocean is 2010-2023; the elements of historical atmospheric observation data include temperature, precipitation, shortwave radiation and 10m height wind vector, and the elements of historical ocean observation data include water temperature, salinity and surface flow vector. 4. The method for predicting green tide coverage area based on deep learning according to claim 1, characterized in that step c comprises the following steps: c1. Determine the long short-term memory neural network model as the training model; c2. The historical multi-source fusion data of the green tide coverage area in the forecast area and the historical observation data of multiple elements of the atmosphere and the ocean collected in step b are used as raw data, and the raw data are processed into a format that can be read by the training model; c3, dividing the data processed in step c2 into two time periods, determining a training set and a test set, which are used for model training and model testing respectively; c4. Use different factor combinations and parameters to design experimental plans. 5. The method for predicting green tide coverage area based on deep learning according to claim 4, characterized in that step c2 comprises the following steps: c21. Reading and processing of historical multi-source fusion data of green tide coverage area in the forecast area; Extract the green tide coverage area in the forecast area from the multi-source fusion data of green tide; read in the source data; standardize the data; ; Among them, _yi represents the green tide coverage area on the ith day, is the average green tide coverage area, is the standard deviation of green tide coverage area; c22. Reading and processing of historical observation data of multiple elements of atmosphere and ocean; Read the original observation data of multiple elements of the atmosphere and ocean in the forecast area; extract the element values and process the data into a matrix form of m×n; where m is the number of forecast factors and n is the number of samples; standardize each element; The standardization formula is as follows: ; Among them, _xij represents the i-th sample of the j-th predictor, represents the average value of the j-th predictor, represents the standard deviation of the j-th predictor; c23. Other data reading and processing; Convert the dates corresponding to the observed data into numerical form and standardize the date values: ; Among them, z _i represents the numerical form of the date of the i-th day, is the average value of the date, is the standard deviation of the date values. 6. A method for forecasting green tide coverage area based on deep learning according to claim 4, characterized in that, in step c3: data from 2010 to 2021 are selected as a training set, and data from 2022 to 2023 are selected as a test set. 7. The method for predicting green tide coverage area based on deep learning according to claim 4, characterized in that step c4 comprises the following steps: c41. Design experimental schemes based on different factor combinations; The prediction factors are divided into three categories: the first category is mandatory factors, including the measured value of green tide coverage area; the second category is optional factors, including forecast date; the third category is environmental factors, including temperature, precipitation, shortwave radiation, 10m height wind vector, water temperature, salinity and surface flow vector; the experimental design method is: mandatory factors + environmental factors, or mandatory factors + environmental factors + optional factors. c42. Design experimental plan based on parameter adjustment; The parameter settings of the long short-term memory neural network include the input factor size, response factor size, number of hidden units, maximum number of cycles and original learning rate. The number of hidden units, maximum number of cycles and original learning rate are adjusted and the experimental plan is designed. 8. According to a method for predicting green tide coverage area based on deep learning as described in claim 1, it is characterized in that in step d and step e: each factor combination is repeatedly trained multiple times, and the long short-term memory neural network model obtained by training is saved; then the error of the long short-term memory neural network model is calculated, and the factor combination with the smallest error is selected for the next step of model optimization. 9. A method for predicting green tide coverage area based on deep learning according to claim 8, characterized in that, in step f and step g: each parameter setting scheme is repeatedly trained multiple times, and the long short-term memory neural network model obtained by training is saved; then the error of the obtained long short-term memory neural network model is calculated, and the one with the smallest error is selected as the optimal neural network prediction model for green tide coverage area. 10. The method for predicting green tide coverage area based on deep learning according to claim 1, characterized in that: For the default values in the historical multi-source fusion data of green tide coverage area in the forecast area and the historical observation data of multiple elements of atmosphere and ocean, the interpolation method is used to fill in the data; During the test process, the test factors are input, and the green tide coverage area is predicted using the long short-term memory neural network model to obtain the prediction results, which are then subjected to inverse standardization. ; Among them, _yi represents the green tide coverage area, _yi ' represents the result of the model output green tide coverage area forecast, is the standard deviation of green tide coverage area, is the average value of green tide coverage area; calculate the forecast error, including the mean absolute error and root mean square error.