CN110705115A

CN110705115A - Meteorological forecasting method and system based on deep belief network

Info

Publication number: CN110705115A
Application number: CN201910960395.2A
Authority: CN
Inventors: 赵琉涛
Original assignee: Beijing Beike Rongzhi Cloud Computing Technology Co Ltd
Current assignee: Beijing Beike Rongzhi Cloud Computing Technology Co Ltd
Priority date: 2019-10-10
Filing date: 2019-10-10
Publication date: 2020-01-17
Anticipated expiration: 2039-10-10
Also published as: CN110705115B

Abstract

The application discloses a weather forecasting method and system based on a deep belief network, wherein the method comprises the steps of obtaining a plurality of historical observation data according to field information of a target area; constructing an enhanced decision tree to train the historical observation data to obtain first prediction data; inputting the first prediction data into a CFD computational fluid dynamics model, performing steady state numerical simulation to obtain second prediction data, and obtaining flow field distribution under an incoming flow condition according to the second prediction data; interpolating the second prediction data at any position in the target area based on a kriging interpolation method to extract a simulated wind speed value and obtain a training data set; inputting a training sample of a training data set into a DBN deep neural network, training the DBN deep neural network, and performing meteorological prediction on data to be observed by using a trained meteorological forecasting model to generate a prediction result. The method and the system can provide the accuracy of forecasting the weather.

Description

Meteorological forecasting method and system based on deep belief network

Technical Field

The application relates to the field of meteorological service intelligence guarantee, in particular to a meteorological forecasting method and system based on a deep belief network.

Background

Weather forecast guarantees have a significant impact on the safe and smooth hosting of numerous activities by users. The traditional meteorological information guarantee is constructed on a meteorological data service platform of a HPC and cloud platform mixed framework, and by developing information technologies based on numerical prediction, statistical learning, high-performance calculation and the like, the application research of multisource data assimilation short-term prediction in the meteorological prediction aspect is focused, the meteorological prediction speed and accuracy are improved, different target groups are oriented, refined and specialized meteorological services are provided, site refined smart meteorological services are well made, and key site meteorological service guarantees are supported.

However, the weather forecasting method built on the weather data service platform with the HPC and cloud platform hybrid architecture has the defect of low forecasting precision due to the initial error and the mode error.

Aiming at the problem of numerical prediction, objective factors exist in the initial value of the mode, and the initial condition of the mode is only an approximation of the real state of the atmosphere, so that the initial error is persistent and mainly an error generated in the observation process, namely an observation error. When only meteorological satellite observation is used, the scattering of visible light by atmosphere causes certain deviation between the detected ground albedo and the cloud top data, and certain error also exists in the process of inverting the cloud top and ground temperature by the directly observed infrared radiation data. In meteorological radar observation, an error exists in the rainfall amount presumed according to an empirical relation between the rainfall amount and radar echo strength, and particularly, the error generated in long-distance detection is larger.

The mode error is an error caused by inaccurate description of the physical process by the mode. The mechanism of generation and elimination of small scale systems in the model is difficult to describe accurately due to turbulent flow processes in atmospheric motion. Parameterized methods are usually used in the model to take into account the influence of these processes, and the averaging process represents the physical and dynamic processes of the subgrid to the average influence of the above-scale of the grid. Errors exist between this description and the true subgrid process; therefore, the accuracy is low when weather is forecasted.

Disclosure of Invention

It is an object of the present application to overcome the above problems or to at least partially solve or mitigate the above problems.

According to one aspect of the application, a weather forecasting method based on a deep belief network is provided, and comprises the following steps:

obtaining a plurality of historical observation data according to the site information of the target area;

constructing an enhanced decision tree to train the historical observation data to obtain first prediction data;

inputting the first prediction data into a CFD computational fluid dynamics model, performing steady state numerical simulation to obtain second prediction data, and obtaining flow field distribution under an incoming flow condition according to the second prediction data;

interpolating the second prediction data at any position in the target area based on a kriging interpolation method to extract a simulated wind speed value and obtain a training data set;

inputting a training sample of a training data set into a DBN deep neural network, training the DBN deep neural network, and performing meteorological prediction on data to be observed by using a trained meteorological forecasting model to generate a prediction result.

Optionally, the weather forecasting method compares a prediction result obtained by the DBN deep neural network with a prediction result obtained by a conventional physical model for analysis, so as to optimize parameters of the first prediction data.

Optionally, the adjusting the parameters of the first prediction data of the CFD model by using a surface feature echo recognition algorithm specifically includes the following substeps:

acquiring base data of a radar, and acquiring echo characteristic parameters according to the base data;

analyzing the echo characteristic parameters by using a statistical method to determine the difference between the feature parameters of the ground feature echo and the precipitation echo;

respectively setting membership functions according to the difference of the characteristic parameters of each ground feature echo and precipitation echo, and converting the membership functions into identification results, wherein the identification results are in the range of [0,1 ];

and carrying out weighted average on the identification results to obtain a criterion value of a certain point, and judging the point as a ground object echo when the criterion value of the certain point exceeds a preset threshold value.

Optionally, the traditional physical model is a model equation set model for solving the atmosphere evolution model, a plurality of dynamic cores, and a physical option solution model.

Optionally, the enhanced decision tree is a CART decision tree model.

Optionally, constructing the CART decision tree model includes the following sub-steps:

acquiring meteorological features based on the Gini index by using a feature selection method;

dividing according to the meteorological features and regression variances to obtain a regression tree;

pruning the regression tree, correcting the regression tree by adopting a post-pruning method, and constructing the CART decision tree model according to the corrected regression tree.

Optionally, the obtaining of the flow field distribution under the incoming flow condition according to the second prediction data includes the following sub-steps:

performing CFD pre-calculation on the flow field according to the site inflow conditions of the first prediction data to obtain key flow field data, storing the key flow field data into a database, and establishing a scene flow field characteristic database;

predicting the wind speed by utilizing the scene flow field characteristic database to obtain second prediction data;

and taking the second prediction data as an input boundary condition of the CFD model, inquiring the incoming flow conditions similar to the second prediction data, and calling flow field distribution data of the scene flow field characteristic database to obtain the flow field distribution under the incoming flow conditions.

Optionally, the historical observation data, the first prediction data, or the second prediction data are optimized to obtain normalized prediction data, and the training data set is constructed based on the normalized prediction data.

According to another aspect of the application, a weather forecasting system based on a deep belief network is provided, and the system comprises an acquisition module, a first training module, a simulation module, an interpolation module and a second training module;

the acquisition module is used for acquiring a plurality of historical observation data according to the site information of the target area;

the first training module is used for constructing an enhanced decision tree to train the historical observation data to obtain first prediction data;

the simulation module is used for inputting the first prediction data into a CFD computational fluid dynamics model, carrying out steady state numerical simulation to obtain second prediction data, and obtaining flow field distribution under an incoming flow condition according to the second prediction data;

the interpolation module interpolates the second prediction data at any position in a target area based on a Krigin interpolation method to extract a simulated wind speed value and obtain a training data set;

the second training module is used for inputting training samples of the training data set into the DBN deep neural network, training the DBN deep neural network, and performing meteorological prediction on data to be observed by using the trained meteorological forecasting model to generate a prediction result.

Optionally, the simulation module performs the following operations:

According to another aspect of the application, there is provided a computer electronic device comprising a memory, a processor and a computer program stored in said memory and executable by said processor, the computer program being stored in a space in the memory for program code, the computer program, when executed by the processor, implementing the method steps for performing any of the weather forecasting methods according to the invention.

According to another aspect of the present application, a computer-readable storage medium is provided, comprising a storage unit for program code, the storage unit being provided with a program for performing the steps of the weather forecasting method according to the invention, the program being executed by a processor.

According to another aspect of the present application, a computer program product comprising instructions for causing a computer to perform the steps of the weather forecast method according to the present invention, when the computer program product is run on the computer is provided.

In the embodiment, the CFD computational fluid dynamics model is used for predicting historical observation data and real-time observation data, so that refined forecast can be realized, the CFD computational fluid dynamics model is used for mining observation data in a time dimension, and identifying and learning historical repeated weather patterns, so that the atmospheric motion process of weather phenomena can be predicted;

in addition, a Krigin interpolation method is utilized to expand the second prediction data from the observation point position to any point in the target area range; meanwhile, the correlation among all meteorological elements is considered, and the DBN deep neural network is utilized to train the simulated wind speed value, so that the data to be observed is forecasted, and the physical correlation among different elements of all observation data is reflected;

the weather forecasting method of the embodiment realizes multi-source data assimilation, mesoscale numerical forecasting and refined weather forecasting based on deep learning, and multi-dimensional display analysis, information monitoring and management of original data, intermediate data, forecast data and evaluation analysis data are included.

The above and other objects, advantages and features of the present application will become more apparent to those skilled in the art from the following detailed description of specific embodiments thereof, taken in conjunction with the accompanying drawings.

Drawings

Some specific embodiments of the present application will be described in detail hereinafter by way of illustration and not limitation with reference to the accompanying drawings. The same reference numbers in the drawings identify the same or similar elements or components. Those skilled in the art will appreciate that the drawings are not necessarily drawn to scale. In the drawings:

FIG. 1 is a schematic flow diagram of a weather forecasting method based on a deep belief network according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a subscription message model provided in accordance with an embodiment of the present application;

FIG. 3 is a schematic structural diagram of a DBN deep neural network provided in accordance with an embodiment of the present application;

FIG. 4 is a schematic diagram of optimizing data according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a weather forecasting system based on a deep belief network according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a computing device according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a computer-readable storage medium according to an embodiment of the application.

Detailed Description

FIG. 1 is a schematic flow diagram of a weather forecasting method based on a deep belief network according to an embodiment of the present application; as can be seen from fig. 1, a weather forecasting method provided by an embodiment of the present application may include the following steps:

s100, obtaining a plurality of historical observation data based on the site information of the target area;

s200, constructing an enhanced decision tree and training historical observation data to obtain first prediction data;

s300, inputting the first prediction data into a CFD computational fluid dynamics model, performing steady state numerical simulation to obtain second prediction data, and obtaining flow field distribution under an incoming flow condition according to the second prediction data;

s400, interpolating the second prediction data at any position in the target area based on a Krigin interpolation method to extract a simulated wind speed value and obtain a training data set;

and S500, inputting the training samples of the training data set into the DBN deep neural network, training the DBN deep neural network, and performing meteorological prediction on data to be observed by using the trained meteorological forecasting model to generate a prediction result.

Optionally, the historical observation data in this embodiment includes surface roughness data and the like. The conventional observation data can obtain fixed-point meteorological information more accurate than the remote sensing data in a direct measurement mode. According to actual requirements, a ground observation station is established at key points of a specific scene, conventional observation data such as wind speed, wind direction, air temperature, air pressure and relative humidity are collected to be used for a multi-source data assimilation model and building a training data set of a small area of the specific scene, and meanwhile quality control is carried out on data stored in an application center database to ensure that data quality is over-critical and improve usability and reliability of the data. The quality control scheme mainly performs initial inspection and marking on the data products from the aspects of climate extreme value range, time consistency, space consistency and consistency among elements.

Optionally, the enhanced decision tree of S200 in this embodiment is a CART decision tree model. Generally, when the dependent variable of the data set is a continuity value, a regression tree is generated, and the mean value observed by the leaf nodes can be used as a prediction value.

Wherein the constructing of the CART decision tree model comprises the following substeps:

s210: acquiring meteorological features based on the Gini index by using a feature selection method; wherein, the meteorological feature selection can adopt methods such as meteorological information gain, gain rate, a Kernel coefficient, chi-square test and the like;

s220: dividing according to the meteorological features and regression variances to obtain a regression tree; the larger the regression variance is, the more dispersed (impure) the node data is, so the minimization method is adopted for partitioning in the embodiment;

s230: pruning the regression tree, correcting the regression tree by adopting a post-pruning method, and constructing the CART decision tree model according to the corrected regression tree;

specifically, after the regression tree is fully grown, pruning is performed based on a loss matrix or a complexity method, and the regression tree is corrected by adopting a post-pruning method.

Specifically, in the prior art, a numerical prediction result of observation data is generally used as a boundary condition, and a Computational Fluid Dynamics (CFD) model is used to simulate observation data such as terrain and surface roughness, so as to obtain finer spatial flow field distribution in a field under a certain incoming flow condition; however, in the CFD model in the prior art, the Navier-Stokes equation (N-S equation) needs to be solved to obtain the spatial flow field distribution during each prediction, so that the calculation amount is huge, and the requirement on prediction timeliness is difficult to meet.

In this embodiment, assuming that the historical observation data is a steady-state inflow condition, it is simulated that the spatial flow field distribution corresponding to the inflow condition in the target region is unique; therefore, in the present embodiment, the CFD model is used to perform the steady-state numerical simulation on the spatial flow field, so as to obtain the spatial flow field distribution under the incoming flow condition.

In this embodiment, if several predicted time points have the same or similar incoming flow conditions, the flow fields in the sites at these time points will also exhibit the same or similar distribution. Therefore, when wind speed prediction is performed at a certain moment, the flow field data of the same or similar inflow conditions at the previous moment (for example, the flow field data of the inflow conditions at intervals of 30S in 1 year) can be referred to, that is, calculation is performed from several elements such as the size of a numerical simulation calculation domain, the order of flow term interpolation, a calculation grid, judgment of a numerical simulation result and the like, and the Navier-Stokes equation does not need to be solved again at each prediction point to obtain spatial flow field distribution, so that the embodiment can quickly predict the adjacent wind speed.

Wherein, step S300 includes the following substeps;

s310: performing CFD pre-calculation on the flow field according to the site inflow conditions of the first prediction data to obtain key flow field data, storing the key flow field data into a database, and establishing a scene flow field characteristic database;

s320: predicting the wind speed by utilizing the scene flow field characteristic database to obtain second prediction data;

s330: and taking the second prediction data as an input boundary condition of the CFD model, inquiring the incoming flow conditions similar to the second prediction data, and calling flow field distribution data of the scene flow field characteristic database to obtain the flow field distribution under the incoming flow conditions.

The prediction method completes the numerical simulation of the complex CFD flow field before the wind speed prediction, and reasonably solves the timeliness problem of the CFD model.

In the embodiment, the CFD is used for calculating the fluid dynamics model, the mesoscale forecast data (namely the first forecast data) is calculated, the small-scale forecast data (namely the second forecast data) is finally output, the small-scale forecast data and the mesoscale data are analyzed and compared, and the small-scale forecast data are continuously optimized, so that the weather forecast capability in a small-scale range is improved.

In the embodiment, a DBN (digital-to-digital converter) deep neural network is constructed, namely current observation data, mesoscale forecast data, an assimilation method, small-scale forecast data and the like are added into the DBN, and a deep learning method and various input data are utilized for training; referring to fig. 3, by setting a contrast test, the optimal configuration of parameters such as the number of hidden layers of the neural network, the number of neurons in each layer, the training learning rate, the iteration number, and the like can be determined. The matching performance of the small-scale forecast result output by the deep learning, the actual observation data collected by the weather data sensor with the fixed context and the laser radar is ensured, and the timeliness and the accuracy of the small-scale forecast result are improved.

Optionally, referring to fig. 4, in S500, the historical observation data, the first prediction data, or the second prediction data are optimized to obtain normalized prediction data, and the training data set is constructed based on the normalized prediction data.

Specifically, the data such as the forecasting data of the scale value, the flow field distribution characteristics simulated by the fluid mechanics model, the historical observation data and the like are optimized (namely, a format conversion method and the like are adopted), the standardized data obtained after optimization is used as a true value to construct a training data set, and a training sample in the training data set is input into a DBN deep neural network for training.

In this example, meteorological data files such as historical observation data, first prediction data or second prediction data that do not meet the file name naming specification may be converted into standard file names, and the following examples will illustrate that various types of data are stored and classified in a unified format to obtain the training data set;

the text file can adopt a standard longitude and latitude format, namely, the text file is classified and stored according to the sequence of station numbers, years, months, days, hours, latitudes, longitudes, station heights, station types, variables and quality control codes, and specific default value information is given;

the binary file can adopt a standard number file format, corresponding year, month, day, time and variable information is given to first row description information of the binary file, and the main body part of the number file is stored according to row latitude and column longitude;

the HDF file can adopt a standard HDF4 format, and relevant information such as time, variables, attributes, longitude and latitude information, missing measurement and the like is given in the HDF4 format file;

the NC file can adopt a standard Netcdf file format, and relevant information such as time, variables, attributes, longitude and latitude information, missing measurement and the like is given in the Netcdf file.

In another embodiment, the weather forecasting method based on the deep belief network further includes step S600, comparing and analyzing the prediction result obtained by the DBN deep neural network with the forecasting result obtained by the traditional physical model, so as to further optimize the parameters of the first prediction data of the statistical model (i.e. CFD model);

wherein the traditional physical model is, for example, a model for solving the atmosphere evolution model or a plurality of dynamic core and physical option solution models for a mode equation set.

Ground object echoes generated by the super-refraction of radar pulses in the atmosphere belong to non-meteorological factors, so that radar detection results are polluted, and the quantitative application of radar data is seriously influenced; the parameters of the first prediction data of the CFD model can thus be optimized using a clutter recognition algorithm.

The feature echo identification algorithm of the embodiment is based on the fuzzy logic method. The fuzzy logic method is to analyze various radar echo characteristics by using a statistical method, find out characteristics of the ground object echo different from the precipitation echo, give the same weight to the characteristics, obtain a quantitative value indicating the possibility of being influenced by the ground object echo in each distance library, and finally identify the ground object echo information exceeding a certain threshold value.

Because the ground target is generally static, the echo intensities are distributed irregularly and often appear on a low elevation layer, and the echo intensity of a higher layer is far less than that of the layer; the layered cloud precipitation has wider horizontal range and uniform strength; the convection cloud precipitation structure is compact, the height is high, and the echo intensity is strong. Therefore, the tuning of the parameters of the first prediction data of the CFD model by the feature echo recognition algorithm in this example comprises the following sub-steps:

s610: acquiring base data of a radar, and acquiring echo characteristic parameters according to the base data;

specifically, the base data of the radar includes: basic data such as echo intensity, radial velocity, velocity spectrum width and the like; the echo characteristic parameters comprise: horizontal texture of echo intensity, vertical variation, sign change in the radial direction, regional average of radial velocity, regional variance of radial velocity, regional average of velocity spectrum width, and the like.

S620: analyzing the echo characteristic parameters by using a statistical method to determine the difference between the feature parameters of the ground feature echo and the precipitation echo;

s630: respectively setting membership functions according to the difference of the characteristic parameters of each ground feature echo and precipitation echo, and converting the membership functions into identification results, wherein the identification results are in the range of [0,1 ];

and S640, carrying out weighted average on the identification result to obtain a criterion value of a certain point, and judging that the point is a ground object echo when the criterion value of the certain point exceeds a preset threshold value.

The weather forecasting method in this embodiment may continuously display real-time weather data and the generated prediction result to the user through a display device (e.g., a web page), and may send early warning information to the user in real time for the prediction result that the predicted weather, rain, snow, and the like may affect normal handling activities.

In one embodiment, referring to fig. 2, a display method of a message queue background is used to display real-time data and a prediction result of weather so as to meet the requirement of a user for synchronously pushing service requests of customized business data.

The message queue background comprises a point-to-point message queue model, wherein the point-to-point message queue model comprises a message Producer (Producer) and a message Consumer (Consumer); the message Producer sends a message to the message consumer via a point-to-point message queue model (i.e., the message Producer sends a message to a particular queue from which the message consumer receives messages; the message consumer signs for an acknowledgement (acknowledgement) of successful processing).

Wherein the message Producer is configured to send a message to a message queue; the message Consumer Consumer is used for receiving the message from the message queue; it should be noted that the producer and consumer of the message may not be running at the same time.

Optionally, the message queue background further includes a publish-subscribe message model, which includes publishers and subscribers, and supports publishers to publish messages to a specific Topic (Topic), and 0 or more subscribers receive messages from the specific Topic (Topic); referring to FIG. 2, under this model, the publisher and subscriber are unaware of each other. In the actual operation process, the subscription is required first, then the message is sent, and then the subscribed message is received.

The subscription message model in this embodiment supports persistence of messages, and is very important for some key core services, and after the persistence of messages is enabled, and after a message queue is down and restarted, the messages can be recovered from persistent storage, and the messages are not lost and can be continuously consumed and processed. The message receiving and sending also support an asynchronous mode: the message is sent asynchronously without waiting for the receiving confirmation of the message queue; and asynchronously receiving the message, and triggering the message consumer to receive the message in a Push mode.

In addition, the real-time weather data and the generated prediction result are uninterruptedly displayed through the display device, so that a data display platform is constructed, the platform supports the functions of image-text combination, a geographic information system and the like, and the current observation data, the medium-scale forecast data, the small-scale forecast data and other data are displayed to the user in an image-text manner, so that data support is provided for the short-term weather forecast with small range, high frequency and accuracy.

In the embodiment, the CFD computational fluid dynamics model is used for predicting historical observation data and real-time observation data, so that refined forecast can be realized, the CFD computational fluid dynamics model is used for mining observation data in a time dimension, and identifying and learning historical repeated weather patterns, so that the atmospheric motion process of weather phenomena can be predicted; in addition, a Krigin interpolation method is utilized to expand the second prediction data from the observation point position to any point in the target area range; meanwhile, the correlation among all meteorological elements is considered, and the DBN deep neural network is utilized to train the simulated wind speed value, so that the data to be observed is forecasted, and the physical correlation among different elements of all observation data is reflected;

According to the embodiment, the data collected by the meteorological data sensor and the laser radar for constructing the fixed scene is taken as historical observation data, the forecast data is output from the mesoscale forecast, and the data is assimilated, so that the meteorological forecast capability in the mesoscale range is improved.

Based on the same inventive concept, as shown in fig. 5, an embodiment of the present application further provides a weather forecasting system based on a depth confidence network, where the weather forecasting system includes an obtaining module, a first training module, a simulation module, an interpolation module, and a second training module;

Optionally, the simulation module performs the following operations:

The identification system provided in this embodiment may execute the method provided in any one of the weather forecasting methods based on the deep belief network, and the detailed process is described in the method embodiment and is not described herein again.

An embodiment of the present application further provides a computing device, referring to fig. 6, comprising a memory 520, a processor 510 and a computer program stored in said memory 520 and executable by said processor 510, the computer program being stored in a space 530 for program code in the memory 520, the computer program, when executed by the processor 510, implementing the steps 531 for performing any of the identification methods according to the present invention.

The embodiment of the application also provides a computer readable storage medium. Referring to fig. 7, the computer readable storage medium includes a storage unit for program code provided with a program 531' for performing the steps of the identification method according to the present invention, the program being executed by a processor.

The embodiment of the application also provides a computer program product containing instructions. Which, when run on a computer, causes the computer to carry out the steps of the method according to the invention.

In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed by a computer, cause the computer to perform, in whole or in part, the procedures or functions described in accordance with the embodiments of the application. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website site, computer, server, or data center to another website site, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

Those of skill would further appreciate that the various illustrative components and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

It will be understood by those skilled in the art that all or part of the steps in the method for implementing the above embodiments may be implemented by a program, and the program may be stored in a computer-readable storage medium, where the storage medium is a non-transitory medium, such as a random access memory, a read only memory, a flash memory, a hard disk, a solid state disk, a magnetic tape (magnetic tape), a floppy disk (floppy disk), an optical disk (optical disk), and any combination thereof.

The above description is only for the preferred embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present application should be covered within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A weather forecasting method based on a deep belief network comprises the following steps:

2. The weather forecasting method according to claim 1, wherein the weather forecasting method compares the prediction result obtained by the DBN deep neural network with the prediction result obtained by the conventional physical model for analyzing so as to optimize the parameter of the first prediction data.

3. The weather forecasting method according to claim 2, wherein the step of optimizing the parameters of the first prediction data of the CFD model by using a ground feature echo recognition algorithm comprises the following sub-steps:

4. The weather forecasting method of claim 2, wherein the conventional physical model is a model of solving the atmospheric evolution model, the plurality of dynamic kernels and the physical options for the system of mode equations.

5. The weather forecasting method of claim 1, wherein the enhanced decision tree is a CART decision tree model.

6. The weather forecasting method of claim 5, wherein constructing the CART decision tree model comprises the sub-steps of:

7. The weather forecasting method according to claim 1, wherein the deriving of the flow field distribution under the incoming flow condition from the second prediction data comprises the sub-steps of:

8. The weather forecasting method of any one of claims 1 to 7, wherein the historical observation data, the first prediction data or the second prediction data are optimized to obtain normalized prediction data, and the training data set is constructed based on the normalized prediction data.

9. A weather forecasting system based on a deep belief network comprises an acquisition module, a first training module, a simulation module, an interpolation module and a second training module;

10. The weather forecasting system of claim 9, wherein the simulation module performs the following operations: