CN111310889A - Evaporation waveguide profile estimation method based on deep neural network - Google Patents
Evaporation waveguide profile estimation method based on deep neural network Download PDFInfo
- Publication number
- CN111310889A CN111310889A CN202010045418.XA CN202010045418A CN111310889A CN 111310889 A CN111310889 A CN 111310889A CN 202010045418 A CN202010045418 A CN 202010045418A CN 111310889 A CN111310889 A CN 111310889A
- Authority
- CN
- China
- Prior art keywords
- neural network
- deep neural
- evaporation waveguide
- training
- refractive index
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/29—Geographical information databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Remote Sensing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an evaporation waveguide profile estimation method based on a deep neural network, which can quickly estimate an evaporation waveguide correction refractive index profile and an evaporation waveguide height in a specified ocean area. Aiming at the defect of low iterative computation efficiency of the existing evaporation waveguide model, the method utilizes a climate prediction system of the national prediction center of the United states, re-analyzes meteorological data and the atmospheric correction refractive index obtained by the calculation of the evaporation waveguide calculation model, trains a deep neural network, and utilizes the accuracy of a data verification method of random positions (except training points) in a training area to obtain an evaporation waveguide section estimation method based on the deep neural network.
Description
Technical Field
The invention belongs to the technical fields of offshore surface evaporation waveguide, atmospheric correction refractive index, offshore over-the-horizon communication, evaporation waveguide calculation models, deep learning and the like, and relates to an evaporation waveguide profile estimation method based on a deep neural network. The method is characterized by obtaining the estimation method of the evaporation waveguide section by adopting high-resolution re-analysis meteorological data and combining an atmospheric motion basic equation, an atmospheric boundary layer similarity theory, a modified refractive index theory, an evaporation waveguide calculation model, a standard normalization method, a deep neural network algorithm and the like. The meteorological data of a large-area sea area can be used for real-time monitoring of the offshore evaporation waveguide when being input.
Background
Evaporative waveguide is a more common atmospheric waveguide phenomenon that often occurs in convective layers, the occurrence of which is related to the interaction between meteorological factors in the atmosphere. Due to the evaporation of seawater, the interaction of seawater and vapor occurs on the sea surface, the vapor is continuously diffused, the atmospheric refractive index in the atmospheric environment is continuously reduced along with the increase of the diffusion height, when the atmospheric refractive index reaches a certain height, the phenomenon that the refractive curvature is smaller than the curvature of the earth sea surface occurs, and the electromagnetic waves are trapped in the layer, so that the over-the-horizon transmission is realized.
At present, the method for obtaining the evaporation waveguide modified refractive index profile mainly comprises a direct measurement method, a prediction model method, an inversion method and the like, and the prediction model method is the main method for obtaining the evaporation waveguide height and the modified refractive index thereof at present. The evaporation waveguide calculation model adopts a sea air flux integral algorithm obtained by long-term offshore investigation, air temperature, specific humidity, wind speed, atmospheric pressure and sea surface temperature at a certain height at sea are substituted into the evaporation waveguide calculation model, an atmospheric correction refraction profile is obtained through iterative calculation for a plurality of times by combining an atmospheric correction refraction index formula, and the height value corresponding to the lowest point of the atmospheric correction refraction index is the height of the evaporation waveguide. The evaporation waveguide calculation model is the best model for estimating the refractive index profile and the evaporation waveguide height at present.
However, the existing evaporation waveguide calculation model generally adopts an iterative technique to estimate the evaporation waveguide modified refractive index profile, so that when the evaporation waveguide profile is calculated for a long time or the evaporation waveguide distribution of a large-area sea area is estimated, the calculation speed is slow and the calculation efficiency is low.
Therefore, the invention establishes a brand-new non-linear mapping evaporation waveguide profile estimation model based on the deep neural network algorithm, and can be applied to overcoming the defects generated by the iteration technology in the existing model.
Disclosure of Invention
Technical problem to be solved
In order to avoid the defects of the prior art, the invention provides an evaporation waveguide profile estimation method based on a deep neural network, and solves the problems that a traditional evaporation waveguide profile prediction model is low in calculation efficiency and difficult to estimate an evaporation waveguide profile in a large-area sea area.
Technical scheme
An evaporation waveguide profile estimation method based on a deep neural network is characterized by comprising the following steps:
step 1: downloading necessary re-analysis meteorological data of a certain year from the NCEP CFS, selecting a training area and selecting a training point in the area; extracting five items of data of sea surface pressure, sea surface temperature, air temperature at a position of 2m, specific humidity at a position of 2m and wind speed at a position of 10m at a training point in the NCEP CFS meteorological data; obtaining an atmospheric correction refractive index value of 0-50 m through iterative calculation for a plurality of times by utilizing an evaporation waveguide calculation model and a calculation formula of the atmospheric correction refractive index; obtaining an atmospheric correction refractive index profile by utilizing the atmospheric refractive index, wherein the height corresponding to the lowest point of the atmospheric correction refractive index profile is the height of the evaporation waveguide; the atmospheric refractive index value is 0.1 m at the point taking interval of 0-50 m;
step 2: for any input or output data XiTo find the mathematical expectation EiAnd standard deviation SiAccording to a standardized formula
Zi＝(Xi-Ei)/Si
Carrying out normalization processing on the training data;
and step 3: constructing Deep Neural Network (Deep Neural Network, DNN),
the number of the neurons of the input layer of the deep neural network is 5;
hidden layer neurons of the deep neural network are initially set to be 3 layers, and the number of each layer of neurons is 6, 8 or 15;
the number of hidden layers of the evaporation waveguide profiles in the low latitude sea area and the high latitude sea area is set to be 3-5 layers;
the number of hidden layers of the medium latitude sea area evaporation waveguide section is set to be 3-10 layers;
determining the number of neurons in a hidden layer of the deep neural network according to the output form, wherein the number of the neurons increases layer by layer, and the number of the neurons in the last layer does not exceed the number of the neurons in an output layer;
step 4, training the deep neural network, namely training the deep neural network according to 1-12 months, inputting every 1000 random data serving as a group of sampling samples into the deep neural network, activating an activation function, transmitting the activation function layer by layer through the deep neural network, continuously updating a link weight of the deep neural network, finally returning training time, sample errors and the root mean square error of the trained network, ending the cycle, accumulating α times of weight updating after α times of cycles, converging the root mean square error of the final sample to β to obtain 12 deep neural network models in total, and storing the trained deep neural network models;
and 5: randomly selecting one point in the training area as a test point, obtaining meteorological data and atmospheric correction refractive index of the test point for 12 months by adopting the method in the step 2, and verifying the accuracy of the deep neural network method;
let the parameter obtained before denormalization be MiThen the real parameters obtained after the normalization contain dimension XiComprises the following steps:
Xi＝Si*Mi+Ei
inputting meteorological data of the test points by using the trained deep neural network, and calculating to obtain an evaporation waveguide profile and an evaporation waveguide height value; if the evaporation waveguide section diagram at the test point has a low fitting degree due to the fact that the data features are not obvious because of large data difference of the training points, returning to the step 4 after removing abnormal values;
step 6: and estimating the average evaporation waveguide height in 1-12 months by using a deep neural network model to obtain a training sea area evaporation waveguide distribution diagram.
The activation function used by the regional deep neural network adopts a nonlinear function such as an S-shaped growing Sigmoid function, a hyperbolic tangent function or a linear rectification function.
The learning rate of the deep neural network is 1, and the deep neural network structures adopt a full-link form.
The number of cycles α of the deep neural network is satisfied according to the training data volume θ hour:
the root mean square error convergence value β satisfies 0< β <2, M units.
The storage content of the deep neural network comprises: deep neural network structure, activation function type, learning rate, weight of neuron between layers and bias.
The number of the neurons of the output layer of the deep neural network is set according to engineering requirements, and is not more than 501.
Advantageous effects
The evaporation waveguide profile estimation method based on the deep neural network can quickly estimate the evaporation waveguide correction refractive index profile and the evaporation waveguide height in a specified ocean area. Aiming at The defect of low iterative computation efficiency of The existing evaporation waveguide model, The method analyzes The gas data and The atmospheric correction refractive index calculated by The evaporation waveguide computation model again by using The national environment Prediction center Climate Prediction System (NCEP CFS), trains a deep neural network, and verifies The accuracy of The method by using The data of random positions in a training area (except training points) to obtain an evaporation waveguide section estimation method based on The deep neural network.
The method of the invention can efficiently and rapidly estimate the evaporation waveguide section and the evaporation waveguide height condition through the deep neural network model, and has the following beneficial effects:
1. the method utilizes global meteorological reanalysis data and atmospheric correction refractive index calculated by an evaporation waveguide calculation model, combines the change rule of an evaporation waveguide section, trains a deep neural network, establishes a set of large-area sea evaporation waveguide section estimation method based on the deep neural network, utilizes data at random positions in an area (except training points) to verify the method, calculates meteorological data average time of 73.2594 seconds in 743 hours by using a traditional evaporation waveguide calculation model, calculates the average time of 1.9543 seconds by using the method, and improves the calculation speed by more than 37 times. Compared with the prior model for estimating the section of the evaporation waveguide by adopting an iteration method, the method has the advantages that: the method has the advantages of high estimation speed, high calculation efficiency, good accuracy and convenience in use.
2. In the global sea area, the method has strong practicability. The deep learning model can be directly put into use after training is completed, complex calculation steps are not needed, real-time monitoring of the offshore evaporation waveguide and prediction and forecast of the offshore evaporation waveguide in a large sea area are facilitated, a communication strategy can be rapidly changed according to changes of the evaporation waveguide condition, and the offshore over-the-horizon communication quality is guaranteed.
Drawings
FIG. 1 is a block diagram of regional deep neural network training based on meteorological data
FIG. 2 shows the specific positions of the training points
FIG. 3 is a schematic diagram of a regional deep neural network structure
FIG. 4 is a root mean square error plot of the evaporative waveguide intensity of training data
FIG. 5 is a random sample comparison graph of an evaporative waveguide computational model and a deep neural network method
FIGS. 6 to 8 are distribution diagrams of monthly average evaporation waveguide height calculated by an evaporation waveguide calculation model in a region and a depth neural network method
Detailed Description
The invention will now be further described with reference to the following examples and drawings:
the method is characterized in that the re-analysis gas image data of a climate forecasting system of the national environmental forecasting center of America is used as the training input of a deep neural network, the atmosphere correction refractive index calculated by an evaporation waveguide calculation model is used as the training output of the network, and the deep neural network is trained to obtain a deep neural network model capable of estimating the evaporation waveguide section in the region and calculating the evaporation waveguide height by combining the correction refractive index theory, the evaporation waveguide characteristic, the deep neural network algorithm, the standard normalization principle and the like. And constructing an evaporation waveguide profile calculation formula by using the deep neural network obtained by training, and inputting meteorological data to obtain an evaporation waveguide profile and an evaporation waveguide height value.
The method comprises the following steps:
step 1: firstly, weather data of 2018 years are downloaded from the NCEP CFS, a training area is selected, and training points are selected in the area. Five items of data of sea surface pressure, sea surface temperature, air temperature at 2m, specific humidity at 2m and wind speed at 10m at a training point in the NCEP CFS re-analysis meteorological data are extracted. And (3) obtaining an atmospheric correction refractive index value of 0-50 m (with an interval of 0.1 m) by utilizing an evaporation waveguide calculation model and combining an atmospheric correction refractive index formula through a plurality of iterative calculations, and obtaining an atmospheric correction refractive index profile by utilizing the atmospheric refractive index, wherein the height corresponding to the lowest point of the atmospheric correction refractive index profile is the evaporation waveguide height. The meteorological data is used as the input of the deep neural network, and the atmospheric modified refractive index is used as the output of the deep neural network.
And the NCEP CFS re-analyzes the meteorological data, extracts and stores the meteorological data in the selected area according to the positions of the training points or the test points respectively, and groups the meteorological data according to the monthly. The selection of the training points should avoid the landlike areas such as islands, reefs and the like, and avoid the generation of random errors.
Step 2: normalizing the training data by adopting a standard normalization method to process the input data and the output data obtained in the step 1, wherein the normalized data is dimensionless and has the size of [ -1, 1]Within the interval. For any input or output data XiTo find the mathematical expectation EiAnd standard deviation SiAccording to the standard, the formula (1) is subjected to standard normalization processing.
Zi＝(Xi-Ei)/Si(1)
In the formula, ZiIs the value of each input parameter after the normalization processing.
And step 3: determining (or adjusting) a Deep Neural Network (DNN) hidden neuron structure. Hidden layer neurons are initially set to be 3 layers, and the number of each layer of neurons is 6, 8 and 15 respectively. The hidden layer quantity is adjusted according to the characteristic value of the training data, generally speaking, the change rule of the evaporation waveguide section in the low latitude sea area and the high latitude sea area in one month is easy to obtain, and the hidden layer quantity is set to be 3-5 layers; the change rule of the evaporation waveguide section in one month is not easy to obtain in the middle latitude sea area under the influence of factors such as climate and the like, and the number of hidden layers is set to be 3-10; and determining the number of neurons in the hidden layer of the deep neural network according to the output form, wherein the number of neurons increases layer by layer, and the number of neurons in the last layer does not exceed the number of neurons in the output layer.
The number of the neurons at the output end of the deep neural network is determined according to the change rule of the evaporation waveguide section. Generally speaking, within 0-20 m, the change situation of the atmospheric correction refractive index is complex, the influence on the height of the evaporation waveguide is large, and the value interval of the neuron can be selected from 0.1 m and 0.2 m … 1.0.0 m; within 20-50 m, the atmospheric correction refractive index change condition is relatively simple, the influence on the height of the evaporation waveguide is small, and the value interval of the neuron can be 1.0 m and 2.0 m … 10.0.0 m.
Step 4, training the deep neural network according to 1-12 months, training the deep neural network by month, inputting every 1000 random data serving as a group of sampling samples into the deep neural network, activating a function, transmitting the activated function through the deep neural network layer by layer, continuously updating a link weight of the deep neural network, finally returning training time, sample errors and the root mean square error of the trained network, ending the cycle till the end of the cycle, accumulating α times of weight updating after α times of cycles, converging the root mean square error of the final sample to β to obtain 12 deep neural network models in total, and storing the trained deep neural network models.
And 5: and (3) randomly selecting one point in the training area as a test point, and obtaining meteorological data and atmospheric correction refractive index of the test point for 12 months by using the method in the step (2) to verify the accuracy of the deep neural network. Assume that the parameter obtained before denormalization is MiThen the real parameters (including dimension) X obtained after the denormalizationiComprises the following steps:
Xi＝Si*Mi+Ei(2)
in the formula (2), SiAnd EiThe standard deviation and the mathematical expectation of each physical quantity obtained in the formula (1) are shown. And constructing an evaporation waveguide profile calculation formula by using the trained deep neural network, inputting meteorological data of the test points, and calculating to obtain an evaporation waveguide profile and an evaporation waveguide height value. And if the evaporation waveguide section diagram at the test point has a low fitting degree due to the fact that the data features are not obvious because of large data difference of the training points, returning to the step 4 after removing the abnormal value.
Step 6: and estimating the average evaporation waveguide height in 1-12 months by using a deep neural network model to obtain a training sea area evaporation waveguide distribution diagram.
The activation function used by the regional deep neural network can be a nonlinear function such as an S-shaped growth (Sigmoid) function, a hyperbolic tangent function, a linear rectification function and the like, the learning rate of the network is 2, and the deep neural network structures adopt a full-link form. The Sigmoid function expression is:
the cycle times α of the deep neural network should satisfy the following conditions according to the training data quantity theta (hour):
the root mean square error convergence value β satisfies 0< β <2(M units).
The contents stored by the deep neural network model by using a deep neural network algorithm comprise: the method comprises the following steps of deep neural network structure, activation function, learning rate, weight of neurons among layers, bias, hidden layer number and total deep neural network layer number.
The evaporation waveguide calculation formula assumes that w is weight and neuron a1Neuron a2And bias b1Input layer, neuron d, forming a neural network11Neuron d12And bias b21Constituting the hidden layer, neuron c1And neuron c2Forming an output layer. E.g. d11Can be expressed as the sum of the weights and biases of the neuron outputs of the previous layer, i.e.:
d11＝a1*w1+a2*w2+b1*1 (5)
wherein the layers are linked by a weight w and a bias b. Input d of any neuron in hidden layer of deep neural networknn(n is an integer greater than 1) can be expressed as a weighted sum of neurons and bias in the previous layer,
dnn＝d(n-1)1*w(n-1)1+…+a(n-1)n*w(n-1)n+b2(n-1)*1 (6）
the input data of the neuron is activated by an activation function, transferred between hidden layers and finally reaches an output layer. The output value is the atmospheric correction refractive index estimated by the deep neural network, and the height corresponding to the lowest point of the atmospheric refractive index profile is the height value of the evaporation waveguide.
Fig. 1 is a block diagram of a regional deep neural network training based on meteorological data, and a sea area in south china sea is selected as a training area in this embodiment. The method is realized by the following steps:
step 1: first, the re-analysis meteorological data of 2018 years is downloaded from the climate forecasting system of the national environmental forecasting center of the United states, a training area (8 degrees N-11 degrees N, 112 degrees E-116 degrees E) is selected, 12 training points are selected in the area, and fig. 2 shows the specific positions of all coordinate points on the Google Earth. Extracting five items of data including sea surface pressure, sea surface temperature, air temperature at 2m, specific humidity at 2m and wind speed at 10m at a training point in the reanalyzed meteorological data, wherein the data volume of 1 month, 3 months, 5 months, 7 months, 8 months and 12 months is 8916 hours; the data volume of months 4, 6, 9 and 11 is 8628 hours; the data volume for month 2 is 8052 hours; since there are many abnormal values in the data of the two-coordinate positions of month 10 (9.7131 ° N, 115.6365 ° E), (10.6389, 115.6365), the effective data amount of month 10 was 7430 hours after the abnormal values were removed by the quality control. And (3) obtaining an atmospheric correction refractive index value of 0-50 m (with an interval of 0.1 m) by utilizing an evaporation waveguide calculation model and combining an atmospheric correction refractive index formula through a plurality of iterative calculations, and obtaining an atmospheric correction refractive index profile by utilizing the atmospheric refractive index, wherein the height corresponding to the lowest point of the atmospheric correction refractive index profile is the evaporation waveguide height. The meteorological data is used as the input of the deep neural network, and the atmospheric modified refractive index is used as the output of the deep neural network.
Step 2: and (3) normalizing the training data to obtain input data and output data obtained in the step (1), wherein the normalized data is dimensionless and has the size within the range of [ -1, 1 ].
Surface pressure: p is a radical off＝(pf-101000)/104；
Temperature: t is (T-293)/100;
specific humidity: q ═ q × 10;
wind speed: v ═ 10)/50.
Modified refractive index: m ═ M-330)/500.
And step 3: determining (or adjusting) a Deep Neural Network (DNN) hidden neuron structure. Hidden layer neurons are initially set to be 3 layers, and the number of each layer of neurons is 6, 8 and 15 respectively. The deep neural network model structure used in this embodiment is: 5-6-8-15-24, the output is the modified refractive index of 0-20 meters (at a spacing of 1.0 meter) and the modified refractive index of 20-50 meters (at a spacing of 10.0 meters). Fig. 3 is a schematic diagram of a deep neural network structure used in this embodiment, I denotes input layer neurons, h denotes hidden layer neurons, O denotes output layer neurons, W and V are weights between neurons, and a and b are weights of bias and neurons.
And 4, step 4: and training the deep neural network. Training the deep neural network according to month division of 1-12 months, inputting every 1000 random data serving as a group of sampling samples into the deep neural network, after activation of an activation function, transmitting the sampling samples layer by layer through the deep neural network, continuously updating a link weight of the deep neural network, finally returning training time, sample errors and root-mean-square errors of the trained network, and ending one cycle. Fig. 4 is a root mean square error graph of the strength of the evaporation waveguide in the training set, after 100 cycles, 100000 times of accumulated weight updates are performed, and the root mean square error convergence value of the final sample is between 0 and 1(M units), so as to obtain 12 deep neural network models in total, and the trained deep neural network models are stored corresponding to 12 months.
And 5: and (3) randomly selecting a point (9.6738 degrees N,113.9968 degrees E) in the training area as a test point, and obtaining meteorological data and atmospheric correction refractive index of the test point for 12 months by using the method in the step (2) to verify the accuracy of the deep neural network. Modified refractive index after denormalization: m × 500+ 330. And constructing an evaporation waveguide profile calculation formula by using the trained deep neural network, inputting meteorological data of the test points, and calculating to obtain the evaporation waveguide profile and the evaporation waveguide height value. Fig. 5 is a comparison diagram of an evaporation waveguide section of a random sample estimated by respectively adopting an evaporation waveguide calculation model and a deep neural network method in the present embodiment, where NPS represents the evaporation waveguide calculation model, and DNN represents the deep neural network method.
Step 6: and estimating the average evaporation waveguide height in 1-12 months by using a deep neural network model to obtain a training sea area evaporation waveguide distribution diagram. Fig. 6 is a 2-month-average evaporation waveguide height distribution graph calculated by an evaporation waveguide calculation model and a depth neural network method, and fig. 7 is a 2-month-average difference distribution graph calculated by the evaporation waveguide calculation model and the depth neural network method. The unit is meters.
Claims (6)
1. An evaporation waveguide profile estimation method based on a deep neural network is characterized by comprising the following steps:
step 1: downloading necessary re-analysis meteorological data of a certain year from the NCEP CFS, selecting a training area and selecting a training point in the area; extracting five items of data of sea surface pressure, sea surface temperature, air temperature at a position of 2m, specific humidity at a position of 2m and wind speed at a position of 10m at a training point in the NCEP CFS meteorological data; obtaining an atmospheric correction refractive index value of 0-50 m through iterative calculation for a plurality of times by utilizing an evaporation waveguide calculation model and a calculation formula of the atmospheric correction refractive index; obtaining an atmospheric correction refractive index profile by utilizing the atmospheric refractive index, wherein the height corresponding to the lowest point of the atmospheric correction refractive index profile is the height of the evaporation waveguide; the atmospheric refractive index value is 0.1 m at the point taking interval of 0-50 m;
step 2: for any input or output data XiTo find the mathematical expectation EiAnd standard deviation SiAccording to the standardized formula Zi＝(Xi-Ei)/Si
Carrying out normalization processing on the training data;
and step 3: constructing Deep Neural Network (Deep Neural Network, DNN),
the number of the neurons of the input layer of the deep neural network is 5;
hidden layer neurons of the deep neural network are initially set to be 3 layers, and the number of each layer of neurons is 6, 8 or 15;
the number of hidden layers of the evaporation waveguide profiles in the low latitude sea area and the high latitude sea area is set to be 3-5 layers;
the number of hidden layers of the medium latitude sea area evaporation waveguide section is set to be 3-10 layers;
determining the number of neurons in a hidden layer of the deep neural network according to the output form, wherein the number of the neurons increases layer by layer, and the number of the neurons in the last layer does not exceed the number of the neurons in an output layer;
step 4, training the deep neural network, namely training the deep neural network according to 1-12 months, inputting every 1000 random data serving as a group of sampling samples into the deep neural network, activating an activation function, transmitting the activation function layer by layer through the deep neural network, continuously updating a link weight of the deep neural network, finally returning training time, sample errors and the root mean square error of the trained network, ending the cycle, accumulating α times of weight updating after α times of cycles, converging the root mean square error of the final sample to β to obtain 12 deep neural network models in total, and storing the trained deep neural network models;
and 5: randomly selecting one point in the training area as a test point, obtaining meteorological data and atmospheric correction refractive index of the test point for 12 months by adopting the method in the step 2, and verifying the accuracy of the deep neural network method;
let the parameter obtained before denormalization be MiThen the real parameters obtained after the normalization contain dimension XiComprises the following steps:
Xi＝Si*Mi+Ei
inputting meteorological data of the test points by using the trained deep neural network, and calculating to obtain an evaporation waveguide profile and an evaporation waveguide height value; if the evaporation waveguide section diagram at the test point has a low fitting degree due to the fact that the data features are not obvious because of large data difference of the training points, returning to the step 4 after removing abnormal values;
step 6: and estimating the average evaporation waveguide height in 1-12 months by using a deep neural network model to obtain a training sea area evaporation waveguide distribution diagram.
2. The evaporation waveguide profile estimation method based on the deep neural network as claimed in claim 1, wherein: the activation function used by the regional deep neural network adopts a nonlinear function such as an S-shaped growing Sigmoid function, a hyperbolic tangent function or a linear rectification function.
3. The evaporation waveguide profile estimation method based on the deep neural network as claimed in claim 1, wherein: the learning rate of the deep neural network is 1, and the deep neural network structures adopt a full-link form.
4. The evaporation waveguide profile estimation method based on the deep neural network as claimed in claim 1, wherein the cycle times α of the deep neural network are satisfied according to the training data volume θ h:
the root mean square error convergence value β satisfies 0< β <2, M units.
5. The evaporation waveguide profile estimation method based on the deep neural network as claimed in claim 1 or 2, wherein: the storage content of the deep neural network comprises: deep neural network structure, activation function type, learning rate, weight of neuron between layers and bias.
6. The evaporation waveguide profile estimation method based on the deep neural network as claimed in claim 1, wherein: the number of the neurons of the output layer of the deep neural network is set according to engineering requirements, and is not more than 501.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010045418.XA CN111310889B (en) | 2020-01-16 | 2020-01-16 | Evaporation waveguide profile estimation method based on deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010045418.XA CN111310889B (en) | 2020-01-16 | 2020-01-16 | Evaporation waveguide profile estimation method based on deep neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111310889A true CN111310889A (en) | 2020-06-19 |
| CN111310889B CN111310889B (en) | 2022-04-12 |
Family
ID=71148815
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010045418.XA Active CN111310889B (en) | 2020-01-16 | 2020-01-16 | Evaporation waveguide profile estimation method based on deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111310889B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111835443A (en) * | 2020-07-23 | 2020-10-27 | 中国联合网络通信集团有限公司 | Method and device for inhibiting atmospheric waveguide interference |
| CN112215299A (en) * | 2020-10-26 | 2021-01-12 | 中山大学 | Block bootstrap method for mean value estimation of hydrological meteorological space data |
| CN112560342A (en) * | 2020-12-15 | 2021-03-26 | 哈尔滨工程大学 | DNN-based atmospheric waveguide parameter estimation method |
| CN112711899A (en) * | 2020-11-27 | 2021-04-27 | 山东省科学院海洋仪器仪表研究所 | Fusion prediction method for height of evaporation waveguide |
| CN112801407A (en) * | 2021-02-24 | 2021-05-14 | 中国人民解放军国防科技大学 | Evaporation waveguide height prediction method based on long-term and short-term memory network |
| CN113014279A (en) * | 2021-01-12 | 2021-06-22 | 西安交通大学 | Frame detection method with anti-interference capability in line-of-sight communication |
| CN113011600A (en) * | 2021-03-25 | 2021-06-22 | 天津大学 | Method for correcting evaporation waveguide prediction model based on machine learning |
| CN113670378A (en) * | 2021-08-09 | 2021-11-19 | 西北工业大学 | Long-term real-time evaporation waveguide profile measurement and channel monitoring system and monitoring method |
| CN114065607A (en) * | 2021-10-19 | 2022-02-18 | 中国电波传播研究所（中国电子科技集团公司第二十二研究所） | Offshore non-uniform evaporation waveguide profile inversion method based on deep convolutional network |
| CN114611415A (en) * | 2022-05-11 | 2022-06-10 | 中国海洋大学 | Beyond-visual-range propagation loss prediction method based on SL-TrellisNets network |
| CN114629577A (en) * | 2020-12-10 | 2022-06-14 | 中国科学院上海高等研究院 | Channel prediction method and system under atmospheric waveguide, storage medium and terminal |
| CN114895378A (en) * | 2022-05-06 | 2022-08-12 | 青岛智慧蓝色海洋工程研究院有限公司 | Method for acquiring near-sea surface atmospheric waveguide state data by multiple nodes |
| CN115356702A (en) * | 2022-08-18 | 2022-11-18 | 中国电波传播研究所（中国电子科技集团公司第二十二研究所） | Evaporation waveguide deep learning joint inversion method based on marine multi-source radar echo |
| CN116822567A (en) * | 2023-08-28 | 2023-09-29 | 山东省科学院海洋仪器仪表研究所 | Optimization method for evaporation waveguide prediction model parameters |
| CN117421601A (en) * | 2023-12-19 | 2024-01-19 | 山东省科学院海洋仪器仪表研究所 | Sea surface evaporation waveguide near-future rapid forecasting method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160116582A1 (en) * | 2011-04-29 | 2016-04-28 | Spatial Digital Systems, Inc. | Radar imaging via spatial spectrum measurement and MIMO waveforms |
| CN106354979A (en) * | 2016-10-08 | 2017-01-25 | 西安电子科技大学 | Method for inverting evaporation waveguide of radar sea clutters based on quantum genetic algorithm |
| CN108663727A (en) * | 2017-03-31 | 2018-10-16 | 西北工业大学 | The method for estimating height of evaporation duct within the scope of world marine site using evaporation rate |
| CN108828547A (en) * | 2018-06-22 | 2018-11-16 | 西安电子科技大学 | The high method of the low Elevation of metre wave radar based on deep neural network |
| CN108981957A (en) * | 2018-05-31 | 2018-12-11 | 西北工业大学 | Submarine temperatures field reconstructing method based on self organizing neural network and Empirical Orthogonal Function |
| CN109858523A (en) * | 2019-01-03 | 2019-06-07 | 武汉大学 | A kind of shallow sea velocity of sound profile inversion method of neural network and ray theory |
| CN110458342A (en) * | 2019-07-26 | 2019-11-15 | 国网江苏省电力有限公司金湖县供电分公司 | One kind monitoring system and method based on improved NARX neural network microclimate |
-
2020
- 2020-01-16 CN CN202010045418.XA patent/CN111310889B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160116582A1 (en) * | 2011-04-29 | 2016-04-28 | Spatial Digital Systems, Inc. | Radar imaging via spatial spectrum measurement and MIMO waveforms |
| CN106354979A (en) * | 2016-10-08 | 2017-01-25 | 西安电子科技大学 | Method for inverting evaporation waveguide of radar sea clutters based on quantum genetic algorithm |
| CN108663727A (en) * | 2017-03-31 | 2018-10-16 | 西北工业大学 | The method for estimating height of evaporation duct within the scope of world marine site using evaporation rate |
| CN108981957A (en) * | 2018-05-31 | 2018-12-11 | 西北工业大学 | Submarine temperatures field reconstructing method based on self organizing neural network and Empirical Orthogonal Function |
| CN108828547A (en) * | 2018-06-22 | 2018-11-16 | 西安电子科技大学 | The high method of the low Elevation of metre wave radar based on deep neural network |
| CN109858523A (en) * | 2019-01-03 | 2019-06-07 | 武汉大学 | A kind of shallow sea velocity of sound profile inversion method of neural network and ray theory |
| CN110458342A (en) * | 2019-07-26 | 2019-11-15 | 国网江苏省电力有限公司金湖县供电分公司 | One kind monitoring system and method based on improved NARX neural network microclimate |
Non-Patent Citations (4)
| Title |
|---|
| D G SEREBRIAKOV 等: "Electric field intensity in cross-section of evaporation duct", 《2018 IEEE RADIO AND ANTENNA DAYS OF THE INDIAN OCEAN》 * |
| 张金鹏 等: "基于雷达海杂波的区域性非均匀蒸发波导反演方法", 《物理学报》 * |
| 杨少波 等: "基于NPS蒸发波导预测模型的适应性研究", 《电子测量与仪器学报》 * |
| 郭晓乐 等: "浅海声速剖面与移动声源的跟踪定位", 《声学学报》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111835443A (en) * | 2020-07-23 | 2020-10-27 | 中国联合网络通信集团有限公司 | Method and device for inhibiting atmospheric waveguide interference |
| CN111835443B (en) * | 2020-07-23 | 2023-03-24 | 中国联合网络通信集团有限公司 | Method and device for inhibiting atmospheric waveguide interference |
| CN112215299A (en) * | 2020-10-26 | 2021-01-12 | 中山大学 | Block bootstrap method for mean value estimation of hydrological meteorological space data |
| CN112215299B (en) * | 2020-10-26 | 2023-08-15 | 中山大学 | Block bootstrap method for hydrological space data mean value estimation |
| CN112711899B (en) * | 2020-11-27 | 2022-04-26 | 山东省科学院海洋仪器仪表研究所 | Fusion prediction method for height of evaporation waveguide |
| CN112711899A (en) * | 2020-11-27 | 2021-04-27 | 山东省科学院海洋仪器仪表研究所 | Fusion prediction method for height of evaporation waveguide |
| CN114629577A (en) * | 2020-12-10 | 2022-06-14 | 中国科学院上海高等研究院 | Channel prediction method and system under atmospheric waveguide, storage medium and terminal |
| CN112560342A (en) * | 2020-12-15 | 2021-03-26 | 哈尔滨工程大学 | DNN-based atmospheric waveguide parameter estimation method |
| CN113014279A (en) * | 2021-01-12 | 2021-06-22 | 西安交通大学 | Frame detection method with anti-interference capability in line-of-sight communication |
| CN112801407A (en) * | 2021-02-24 | 2021-05-14 | 中国人民解放军国防科技大学 | Evaporation waveguide height prediction method based on long-term and short-term memory network |
| CN113011600A (en) * | 2021-03-25 | 2021-06-22 | 天津大学 | Method for correcting evaporation waveguide prediction model based on machine learning |
| CN113670378B (en) * | 2021-08-09 | 2023-06-30 | 西北工业大学 | Long-term real-time evaporation waveguide profile measurement and channel monitoring system and monitoring method |
| CN113670378A (en) * | 2021-08-09 | 2021-11-19 | 西北工业大学 | Long-term real-time evaporation waveguide profile measurement and channel monitoring system and monitoring method |
| CN114065607A (en) * | 2021-10-19 | 2022-02-18 | 中国电波传播研究所（中国电子科技集团公司第二十二研究所） | Offshore non-uniform evaporation waveguide profile inversion method based on deep convolutional network |
| CN114895378B (en) * | 2022-05-06 | 2024-01-26 | 青岛智慧蓝色海洋工程研究院有限公司 | Method for collecting near sea surface atmosphere waveguide state data through multiple nodes |
| CN114895378A (en) * | 2022-05-06 | 2022-08-12 | 青岛智慧蓝色海洋工程研究院有限公司 | Method for acquiring near-sea surface atmospheric waveguide state data by multiple nodes |
| CN114611415B (en) * | 2022-05-11 | 2022-08-12 | 中国海洋大学 | Beyond-visual-range propagation loss prediction method based on SL-TrellisNets network |
| CN114611415A (en) * | 2022-05-11 | 2022-06-10 | 中国海洋大学 | Beyond-visual-range propagation loss prediction method based on SL-TrellisNets network |
| CN115356702A (en) * | 2022-08-18 | 2022-11-18 | 中国电波传播研究所（中国电子科技集团公司第二十二研究所） | Evaporation waveguide deep learning joint inversion method based on marine multi-source radar echo |
| CN116822567A (en) * | 2023-08-28 | 2023-09-29 | 山东省科学院海洋仪器仪表研究所 | Optimization method for evaporation waveguide prediction model parameters |
| CN116822567B (en) * | 2023-08-28 | 2023-11-21 | 山东省科学院海洋仪器仪表研究所 | Optimization method for evaporation waveguide prediction model parameters |
| CN117421601A (en) * | 2023-12-19 | 2024-01-19 | 山东省科学院海洋仪器仪表研究所 | Sea surface evaporation waveguide near-future rapid forecasting method |
| CN117421601B (en) * | 2023-12-19 | 2024-03-01 | 山东省科学院海洋仪器仪表研究所 | Sea surface evaporation waveguide near-future rapid forecasting method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111310889B (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111310889B (en) | Evaporation waveguide profile estimation method based on deep neural network | |
| Zhang et al. | Short-term rainfall forecasting using multi-layer perceptron | |
| CN111783987A (en) | Farmland reference crop evapotranspiration prediction method based on improved BP neural network | |
| CN112287294B (en) | Space-time bidirectional soil water content interpolation method based on deep learning | |
| CN111833202B (en) | Farmland evapotranspiration short-term prediction method considering crop coefficient dynamic change and rainfall | |
| CN113255972B (en) | Short-term rainfall prediction method based on Attention mechanism | |
| CN112288164A (en) | Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast | |
| CN114445634A (en) | Sea wave height prediction method and system based on deep learning model | |
| CN112163375A (en) | Long-time sequence near-surface ozone inversion method based on neural network | |
| CN113935249B (en) | Upper-layer ocean thermal structure inversion method based on compression and excitation network | |
| CN112180369B (en) | Depth learning-based sea surface wind speed inversion method for one-dimensional synthetic aperture radiometer | |
| CN106570592A (en) | Artificial-neural-network-based intelligent numerical value forecasting correction system | |
| CN112396152A (en) | Flood forecasting method based on CS-LSTM | |
| CN116644848A (en) | Multi-source data-based medium-long term area wind power prediction method | |
| CN116029419A (en) | Deep learning-based long-term new energy daily average generation power prediction method and system | |
| CN115859789A (en) | Method for improving inversion accuracy of polar atmosphere temperature profile | |
| CN116559977A (en) | NPS-GRYANIK20 evaporation waveguide height prediction method based on BP neural network | |
| CN113095547B (en) | Short-term wind power prediction method based on GRA-LSTM-ICE model | |
| CN117172149A (en) | Evaporation waveguide prediction method based on data feature classification and neural network model | |
| CN116822381B (en) | Marine temperature salt structure inversion method based on artificial intelligence | |
| CN111488974B (en) | Ocean wind energy downscaling method based on deep learning neural network | |
| CN112085272A (en) | Wind power prediction method | |
| CN111428421A (en) | Rainfall runoff simulation method for deep learning guided by physical mechanism | |
| Chang et al. | Neural network with multi-trend simulating transfer function for forecasting typhoon wave | |
| Dinzi et al. | The use of meteorology data in short-term prediction of wind speed for wind turbine using elman recurrent neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |