CN108734357A - Weather prognosis system and method - Google Patents
Weather prognosis system and method Download PDFInfo
- Publication number
- CN108734357A CN108734357A CN201810531260.XA CN201810531260A CN108734357A CN 108734357 A CN108734357 A CN 108734357A CN 201810531260 A CN201810531260 A CN 201810531260A CN 108734357 A CN108734357 A CN 108734357A
- Authority
- CN
- China
- Prior art keywords
- neural network
- network model
- radar
- radar echo
- unit
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
Abstract
The invention discloses a kind of weather prognosis system and methods, wherein weather prognosis system includes：First unit obtains training sample based on history radar echo map；Second unit builds neural network model；Third unit obtains the radar echo map before the moment to be predicted；Wherein, the second unit is trained the neural network model with the training sample；The input for the neural network model that the second unit is completed using the radar echo map before the moment to be predicted obtained as training, so that the neural network model exports the radar echo map at moment to be predicted.Forecast speed can be improved in the weather prognosis system of the embodiment of the present invention.
Description
Technical field
The present invention relates to weather prognosis technical field, more particularly to a kind of weather prognosis system and method.
Background technology
The present invention belongs to the relevant technologies related to the present invention for the description of background technology, be only used for explanation and just
In the invention content for understanding the present invention, it should not be construed as applicant and be specifically identified to or estimate applicant being considered of the invention for the first time
The prior art for the applying date filed an application.
Production activity, social activities, military activity and the daily life of Changes in weather and people suffer from very closely
Relationship.Since ancient times, the people side of always wanting to tries to following Changes in weather of prediction, and using advantageous weather, it is unfavorable to avoid
Weather, to reduce unnecessary loss.Present weather forecast mode mainly passes through meteorological satellite and radar return image,
Then the method forecasting weather for passing through numerical computations with computer.However weather system is sufficiently complex, often chaos system, sharp
Carrying out the methods of numerical computations with computer, time-consuming.
Invention content
In view of this, the embodiment of the present invention provides a kind of weather prognosis system and method, main purpose is to improve forecast speed
Degree.
In order to achieve the above objectives, present invention generally provides following technical solutions：
In a first aspect, an embodiment of the present invention provides a kind of weather prognosis systems, including：
First unit obtains training sample based on history radar echo map；
Second unit builds neural network model；
Third unit obtains the radar echo map before the moment to be predicted；
Wherein, the second unit is trained the neural network model with the training sample；
The nerve that the second unit is completed using the radar echo map before the moment to be predicted obtained as training
The input of network model, so that the neural network model exports the radar echo map at moment to be predicted.
Preferably, further including Unit the 4th, Unit the 4th pre-processes the history radar echo map, institute
It states first unit and obtains the history radar echo map after Unit the 4th pretreatment as training sample, the pre- place
Reason includes by the original echo value in history radar echo map divided by the setting numerical value more than 1, to reduce the sparse journey of solution space
Degree；
Unit the 5th, radar echo map before the moment to be predicted that Unit the 5th obtains the third unit into
Row pretreatment, what the radar echo map before the moment to be predicted after the pretreatment of Unit the 5th was completed as training
The input of the neural network model, the pretreatment includes by the original echo in the radar echo map before the moment to be predicted
Value divided by the setting numerical value more than 1, to reduce the sparse degree of solution space；
Unit the 6th, the thunder at the moment to be predicted that the neural network model that Unit the 6th completes training exports
Original echo value up to reflectogram is multiplied by the setting numerical value.
Preferably, the pretreatment of Unit the 4th and the pretreatment of Unit the 5th further include by the radar echo map
The Echo Rating zero setting in the strong noise area within the scope of middle radar center certain radius.
Preferably, the second unit is using multiple continuous radar echo maps as the defeated of the neural network model
Enter.
Preferably, the neural network model is the multilayer convolutional layer and multilayer warp lamination or described of U-Net structures
Neural network model is the multilayer convolutional layer and multilayer warp lamination of HED structures.
Preferably, the neural network model utilizes net using the general characteristic of network deep layer extraction radar echo map
Network shallow-layer extracts local feature, and the general characteristic is combined with the local feature, fitting cloud layer movement.
Preferably, the neural network model is built based on Google deep learning Open Framework Tensorflow.
Preferably, the neural network model use prediction radar echo map and sample mean square deviation as loss function
Training pattern.
Second aspect, an embodiment of the present invention provides a kind of Predictive meteorological methods, include the following steps：
Training sample is obtained, the training sample is obtained based on history radar echo map；
Neural network model is trained with the training sample；
Obtain the radar echo map before the moment to be predicted；
Using the radar echo map before the moment to be predicted of acquisition as the defeated of the neural network model of training completion
Enter, the radar echo map at the time of neural network model exported before the moment to be predicted.
Preferably, using multiple continuous radar echo maps as the input of the neural network model.
Preferably, being used as training sample after the history radar echo map is preprocessed, the pretreatment includes will
Original echo value in history radar echo map divided by the setting numerical value more than 1, to reduce the sparse degree of solution space；
The nerve that radar echo map before the moment to be predicted obtained is completed after being pre-processed as training
The input of network model, the pretreatment include by original echo value in the radar echo map before the moment to be predicted divided by big
In 1 setting numerical value, to reduce the sparse degree of solution space；
The original echo value of the radar echo map at the moment to be predicted for the neural network model output that training is completed
It is multiplied by the setting numerical value.
Preferably, the pretreatment further includes in the radar echo map for input the neural network model in radar
The Echo Rating zero setting in the strong noise area in certain (10km) radius of the heart.
Preferably, the neural network model is the multilayer convolutional layer and multilayer warp lamination or described of U-Net structures
Neural network model is the multilayer convolutional layer and multilayer warp lamination of HED structures.
Preferably, the neural network model utilizes net using the general characteristic of network deep layer extraction radar echo map
Network shallow-layer extracts local feature, and the general characteristic is combined with the local feature, fitting cloud layer movement.
Preferably, the neural network model is built based on Google deep learning Open Framework Tensorflow.
Preferably, the neural network model use prediction radar echo map and sample mean square deviation as loss function
Training pattern.
The third aspect, an embodiment of the present invention provides a kind of computer readable storage mediums, are stored thereon with computer journey
The step of sequence, which realizes above-mentioned method when being executed by processor.
Fourth aspect an embodiment of the present invention provides a kind of computer equipment, including memory, processor and is stored in
On reservoir and the computer program that can run on a processor, the processor realize above-mentioned method when executing described program
Step.
Compared with prior art, the advantageous effect of the embodiment of the present invention is：
Weather prognosis system and method provided in an embodiment of the present invention is based on deep learning, in conjunction with true radar echo map
Picture establishes model and has successfully trained prediction model, can complete to predict in a few seconds, improve forecast speed.Greatly
The uncertainty artificially analyzed and brought is reduced, while can help to reduce prediction error.
Scheme attached explanation
Fig. 1 shows the schematic diagram of an embodiment of weather prognosis system of the present invention.
Fig. 2 shows the schematic diagrames of another embodiment of weather prognosis system of the present invention.
Fig. 3 shows the flow chart of an embodiment of Predictive meteorological methods of the present invention.
Fig. 4 shows the flow chart of another embodiment of Predictive meteorological methods of the present invention.
Specific implementation mode
With reference to specific embodiment, present invention is further described in detail, but not as a limitation of the invention.?
In following the description, what different " embodiment " or " embodiment " referred to is not necessarily the same embodiment.In addition, one or more are implemented
Special characteristic, structure or feature in example can be combined by any suitable form.
Fig. 1 shows the schematic diagram of an embodiment of weather prognosis system of the present invention.Fig. 2 shows weather prognosis of the present invention
The schematic diagram of another embodiment of system.Referring to Fig. 1 and Fig. 2, the weather prognosis system of the embodiment of the present invention, including：
First unit 10 obtains training sample based on history radar echo map；
Second unit 20, builds neural network model；
Third unit 30 obtains the radar echo map before the moment to be predicted；
Wherein, second unit 20 is trained neural network model with training sample；
The nerve net that second unit 20 is completed using the radar echo map before the moment to be predicted of acquisition as training
The input of network model, so that the neural network model exports the radar echo map at moment to be predicted.
Weather prognosis system provided in an embodiment of the present invention is based on deep learning, in conjunction with true radar return image, establishes
Model has simultaneously successfully trained prediction model, and prediction model can complete the moment to be predicted (the radar return of input in a few seconds
The subsequent time of figure) radar echo map prediction, weather prognosis can be carried out on this basis, improve forecast speed.Greatly
Reduce the uncertainty artificially analyzed and brought, while can help to reduce prediction error.
Image, semantic cutting techniques of the weather prognosis system of the embodiment of the present invention based on deep learning, the nerve net built
Network model takes out movement locus and the variation of radar return using the ability of convolutional neural networks extraction feature and Function Fitting
Trend can realize weather forecast to obtain the radar return image of subsequent time based on this.Such as precipitation radar echo strength
With in air drop size and content be positively correlated, this directly implies the possible rainfall of current and future.Therefore, pass through
Deep learning, the wherein complicated relationship of study, predicts next radar return image, can be used for weather forecast.
Radar return image needs the regular hour, therefore two adjacent radar return images can also have certain reality
Border interval.For example, the every 6 minutes width of radar return data, differed 6 minutes between adjacent two radar return images.Therefore,
Second unit using multiple continuous radar echo maps (include as the radar echo map of training sample and before the moment to be predicted
It is radar echo map, that is, identical when the number of radar echo map inputted when predicting is with training) it (including is instructed as neural network model
Before white silk and training complete neural network model) input can solve radar image interruption, the discontinuous problem of data.Even
The quantity of continuous radar return image can be specifically 2 width, 3 width, 4 width, 5 width, 6 width or more.In one embodiment, it adopts
Use continuous 5 width radar return image as mode input, with the radar echo map of the subsequent time of the radar return image of input
As being trained as sample, neural network model is established, instruction is inputted with the continuous 5 width radar echo map before the moment to be predicted
Practice the neural network model completed, prediction exports moment radar echo map to be predicted, and (sequential is last in the radar echo map of input
A radar echo map correspond to the subsequent time at moment) in every bit output valve.
In the embodiment of the present invention, the radar echo map for inputting neural network model can be original graph, can also be by
Pretreated figure.For example, training sample can directly use history radar return image.In view of original radar data unit
For dbz.Value is 0,5,10,15 ....The value of original image is relatively sparse, if directly instructed original image for network
Practice, corresponding solution space also can be sparse, is unfavorable for model convergence.Therefore, single the embodiment of the invention also includes the 4th referring to Fig. 2
First 40, the 5th unit 50 and the 6th unit 60, wherein the 4th unit 40 pre-processes history radar echo map, and first is single
Member 10 obtains the history radar echo map after the pretreatment of the 4th unit 40 as training sample, and pretreatment includes by history radar
Original echo value in reflectogram divided by the setting numerical value more than 1, to reduce the sparse degree of solution space.Equally, Unit the 5th
Radar echo map before the moment to be predicted that 50 pairs of third units 30 obtain pre-processes, after the pretreatment of the 5th unit 50
Moment to be predicted before radar echo map as training complete neural network model input, pretreatment with as training
The pretreatment of the radar echo map of sample is identical, including：By the original echo value in the radar echo map before the moment to be predicted
Divided by the setting numerical value more than 1, to reduce the sparse degree of solution space；The neural network model that 6th unit 60 completes training
The original echo value of radar echo map at the moment to be predicted of output is multiplied by above-mentioned setting numerical value, to be restored, in order to
The original radar echo map directly acquired is consistent.Specific setting numerical value can determine according to the value of original image, can be with
It is the numerical value that can divide exactly.In one embodiment, set numerical value as 5, i.e., it, will be pre- by the data of original image divided by 5
The data of the image of survey are multiplied by 5.In the present embodiment, can all be normalized to avoid by each image, can cause radar echo map it
Between numerical values recited relationship lose the problem of.And use and division is done to initial data, the sparse degree of solution space is reduced, together
When remain Echo Rating relative size between image, reduce the sparse degree of solution space, be conducive to the convergence of model.
In one embodiment of the present of invention, the pretreatment of Unit the 4th further includes by radar center in history radar echo map
The Echo Rating zero setting in the strong noise area within the scope of certain radius.Equally, the pretreatment of Unit the 5th further includes nerve to be inputted
The echo in the strong noise area in the radar echo map before the moment to be predicted of network model within the scope of radar center certain radius
It is worth zero setting.By the way that radar echo value zero setting at radar center strong noise to remove noise, may be implemented pre- in the present embodiment
Survey the reduction of error, acceleration model convergence.The range in specific strong noise area can be 8km radius at radar center,
10km radius or 11km radius etc..Such as the Echo Rating of the border circular areas in radar center 10km radius is set
Zero, reduce influence of noise, acceleration model convergence.
Image, semantic cutting techniques of the weather prognosis system of the embodiment of the present invention based on deep learning, specifically, neural
Network model can be the multilayer convolutional layer of U-Net structures and multilayer warp lamination or neural network model can be HED structures
Multilayer convolutional layer and multilayer warp lamination.In the embodiment of the present invention, neural network model is returned using network deep layer extraction radar
The general characteristic of wave figure extracts local feature using network shallow-layer, general characteristic is combined with the local feature, is fitted cloud
Layer movement.The present embodiment can fit the compound movement of cloud layer, including the direction of motion and size variation trend.In modelling
On, with reference to the relevant knowledge in meteorology, high-altitude wind speed is first set, adjacent two width radar return image is can be obtained by this way and obtains
Take the displacement distance in interval that can obtain single-point in radar return image according to the number of the radar return image of input
Maximum moving distance obtains the pixel of single point movement in radar return image according to the resolution ratio of radar return image.For example,
High-altitude wind speed is set as 50m/s, in embodiment as above, the time interval of adjacent two width radar return image is 6 minutes, therefore,
According to above-mentioned high-altitude wind speed, 18km can be moved within 6 minutes, by taking input picture uses 5 width consecutive images as an example, single-point in radar map
Maximum removable 90km in the input image, by taking the resolution ratio of radar return image is 1km as an example, 90km corresponds to about 90 pixels.
So as to obtain enough information, the movement of cloud layer complexity is fitted.
In the embodiment of the present invention, neural network model can be taken based on Google deep learning Open Framework Tensorflow
It builds.Certainly, the neural network model in the embodiment of the present invention can also be built using other frames, no longer be gone to live in the household of one's in-laws on getting married one by one herein
It states.
In one embodiment of the invention, neural network model use prediction radar echo map and sample mean square deviation as lose
Function training pattern.For the MSE errors of finally obtained model predication value and actual value in 3dbz or so, performance is very outstanding.
In the embodiment of the present invention, what third unit 30 obtained is used to input the model after training, to predict the moment to be predicted
Radar return image moment to be predicted before radar return image include model output prediction radar return image,
So as to realize rolling forecast.In the present embodiment, in the model prediction stage, after can continuously being predicted by rolling forecast
Radar return image is used in short-time weather forecasting.
Second aspect, an embodiment of the present invention provides a kind of Predictive meteorological methods, Fig. 3 shows weather prognosis side of the present invention
The flow chart of one embodiment of method.Fig. 4 shows the flow chart of another embodiment of Predictive meteorological methods of the present invention.Referring to Fig. 3
And Fig. 4, the Predictive meteorological methods include the following steps：
Training sample is obtained, training sample is obtained based on history radar echo map；
Neural network model is trained with training sample；
Obtain the radar echo map before the moment to be predicted；
The input for the neural network model completed using the radar echo map before the moment to be predicted of acquisition as training, god
The radar echo map at moment to be predicted is exported through network model.
The Predictive meteorological methods of the embodiment of the present invention can realize by the forecasting system of any of the above-described embodiment, therefore above-mentioned pre-
The scheme that examining system embodiment is related to can be used in the embodiment of prediction technique of the present invention.
Predictive meteorological methods provided in an embodiment of the present invention are based on deep learning, in conjunction with true radar return image, establish
Model has simultaneously successfully trained prediction model, and prediction model can complete the moment to be predicted (the radar return of input in a few seconds
The subsequent time of figure) radar echo map prediction, weather prognosis can be carried out on this basis, improve forecast speed.Greatly
Reduce the uncertainty artificially analyzed and brought, while can help to reduce prediction error.
Image, semantic cutting techniques of the Predictive meteorological methods of the embodiment of the present invention based on deep learning, the nerve net built
Network model takes out movement locus and the variation of radar return using the ability of convolutional neural networks extraction feature and Function Fitting
Trend can realize weather forecast to obtain the radar return image of subsequent time based on this.Such as precipitation radar echo strength
With in air drop size and content be positively correlated, this directly implies the possible rainfall of current and future.Therefore, pass through
Deep learning, the wherein complicated relationship of study, predicts next radar return image, can be used for weather forecast.
Radar return image needs the regular hour, therefore two adjacent radar return images can also have certain reality
Border interval.For example, the every 6 minutes width of radar return data, differed 6 minutes between adjacent two radar return images.Therefore,
In the embodiment of prediction technique of the present invention using multiple continuous radar echo maps (include as training sample radar echo map and
Radar echo map before moment to be predicted, that is, identical when the number of the radar echo map inputted when predicting is with training) as god
Input through network model (including the neural network model completed with training before training) can solve radar image and interrupt, number
According to discontinuous problem.The quantity of continuous radar return image can be specifically 2 width, 3 width, 4 width, 5 width, 6 width or more.Its
In middle one embodiment, using continuous 5 width radar return image as mode input, with the next of the radar return image of input
The radar return image at moment is trained as sample, establishes neural network model, with continuous 5 width before the moment to be predicted
The neural network model that radar echo map input training is completed, prediction export the moment to be predicted (when in the radar echo map of input
A last radar echo map of sequence corresponds to the subsequent time at moment) output valve of every bit in radar echo map.
In the embodiment of prediction technique of the present invention, the radar echo map for inputting neural network model can be original graph,
Can be by pretreated figure.Referring to Fig. 4, for example, training sample can directly use history radar return image.It considers
Original radar data unit is dbz.Value is 0,5,10,15 ....The value of original image is relatively sparse, if directly will be former
Beginning image is used for network training, and corresponding solution space also can be sparse, is unfavorable for model convergence.Therefore, the embodiment of the present invention is to defeated
The radar echo map for entering neural network model includes the radar echo map as training sample and the radar before the moment to be predicted
Reflectogram) it is pre-processed, to improve accuracy rate and efficiency.For example, using the history radar echo map after pretreatment as instruction
Practice sample, pretreatment includes by the original echo value in history radar echo map divided by the setting numerical value more than 1, to reduce solution sky
Between sparse degree.Radar echo map before the moment to be predicted obtained is completed described as training after being pre-processed
The input of neural network model, pretreatment include more than removing the original echo value in the radar echo map before the moment to be predicted
The setting numerical value more than 1 is stated, to reduce the sparse degree of solution space；The neural network model that training is completed exports to be predicted
The original echo value of the radar echo map at moment is multiplied by above-mentioned setting numerical value, to be restored, in order to the original that directly acquires
Beginning radar echo map is consistent.Specific setting numerical value can be determined according to the value of original image, such as can be can be whole
Except a value of radar numerical value.In one embodiment, numerical value is set as 5, and nerve will be inputted behind the data of original image divided by 5
The data of the image of prediction are multiplied by 5 by network model.It in the present embodiment, uses and division is done to initial data, reduce solution space
Sparse degree, while remaining Echo Rating relative size between image, reduce the sparse degree of solution space, be conducive to model
Convergence.And can all be normalized to avoid by each image, the numerical values recited relationship between radar echo map can be caused to lose
Problem.
In one embodiment of the method for the present invention, pretreatment further includes the radar echo map of neural network model to be inputted
Radar center certain radius in (including the radar echo map as training sample and the radar echo map before the moment to be predicted)
The Echo Rating zero setting in the strong noise area in range.For example, will be in history radar echo map within the scope of radar center certain radius
The Echo Rating zero setting in strong noise area.Equally, in the radar echo map before the moment to be predicted that neural network model will be inputted
The Echo Rating zero setting in the strong noise area within the scope of radar center certain radius.By will be at radar center strong noise in the present embodiment
Radar echo value zero setting to remove noise, may be implemented prediction error reduction, acceleration model convergence.Specific strong noise area
Range can be 8km radius at radar center, 10km radius or 11km radius etc..Such as by radar center
The Echo Rating zero setting of border circular areas in 10km radius reduces influence of noise, acceleration model convergence.
Image, semantic cutting techniques of the Predictive meteorological methods of the embodiment of the present invention based on deep learning, specifically, neural
Network model can be the multilayer convolutional layer of U-Net structures and multilayer warp lamination or neural network model can be HED structures
Multilayer convolutional layer and multilayer warp lamination.In the embodiment of the present invention, neural network model is returned using network deep layer extraction radar
The general characteristic of wave figure extracts local feature using network shallow-layer, general characteristic is combined with the local feature, is fitted cloud
Layer movement.The present embodiment can fit the compound movement of cloud layer, including the direction of motion and size variation trend.In modelling
On, with reference to the relevant knowledge in meteorology, high-altitude wind speed is first set, adjacent two width radar return image is can be obtained by this way and obtains
Take the displacement distance in interval that can obtain single-point in radar return image according to the number of the radar return image of input
Maximum moving distance obtains the pixel of single point movement in radar return image according to the resolution ratio of radar return image.For example,
High-altitude wind speed is set as 50m/s, in embodiment as above, the time interval of adjacent two width radar return image is 6 minutes, therefore,
According to above-mentioned high-altitude wind speed, 18km can be moved within 6 minutes, by taking input picture uses 5 width consecutive images as an example, single-point in radar map
Maximum removable 90km in the input image, by taking the resolution ratio of radar return image is 1km as an example, 90km corresponds to about 90 pixels.
So as to obtain enough information, the movement of cloud layer complexity is fitted.
In the embodiment of the present invention, neural network model can be taken based on Google deep learning Open Framework Tensorflow
It builds.Certainly, the neural network model in the embodiment of the present invention can also be built using other frames, no longer be gone to live in the household of one's in-laws on getting married one by one herein
It states.
In one embodiment of the invention, neural network model use prediction radar echo map and sample mean square deviation as lose
Function training pattern.For the MSE errors of finally obtained model predication value and actual value in 3dbz or so, performance is very outstanding.
In the embodiment of the present invention, the radar return image before moment to be predicted for inputting the model after training includes
The radar return image of the prediction of model output, so as to realize rolling forecast.In the present embodiment, in the model prediction stage,
Radar return image after can continuously being predicted by rolling forecast is in short-time weather forecasting.
The third aspect, an embodiment of the present invention provides a kind of computer readable storage mediums, are stored thereon with computer journey
The step of sequence, which realizes above-mentioned method when being executed by processor.
Fourth aspect an embodiment of the present invention provides a kind of computer equipment, including memory, processor and is stored in
On reservoir and the computer program that can run on a processor, the processor realize above-mentioned method when executing described program
Step.
Those skilled in the art can be understood that the embodiment of the present invention technical solution can by software and/or
Hardware is realized." unit " or " unit " in this specification is to refer to complete independently or completed with other component cooperation specific
The software and/or hardware of function, wherein hardware for example can be that (Field-Programmable Gate Array show FPGA
Field programmable gate array), IC (Integrated Circuit, integrated circuit) etc..
The embodiment of the present invention additionally provides a kind of computer readable storage medium, is stored thereon with computer program, the journey
The step of any of the above-described embodiment method is realized when sequence is executed by processor.Wherein, computer readable storage medium may include
But be not limited to any kind of disk, including floppy disk, CD, DVD, CD-ROM, mini drive and magneto-optic disk, ROM, RAM,
EPROM, EEPROM, DRAM, VRAM, flash memory device, magnetic or optical card, nanosystems (including molecular memory IC),
Or it is suitable for any kind of medium or equipment of store instruction and/or data.
The embodiment of the present invention additionally provides a kind of computer equipment, including memory, processor and storage are on a memory
And the computer program that can be run on a processor, the step of realizing any of the above-described embodiment method when processor executes program.
In embodiments of the present invention, processor is the control centre of computer system, can be the processor of physical machine, can also be void
The processor of quasi- machine.
In the present invention, term " first ", " second " etc. are only used for the purpose of description, are not understood to indicate or imply
Relative importance or sequence；Term " multiple " then refers to two or more, unless otherwise restricted clearly.Term " installation ",
The terms such as " connected ", " connection ", " fixation " shall be understood in a broad sense, for example, " connection " may be a fixed connection, can also be can
Dismantling connection, or be integrally connected；" connected " can be directly connected, can also be indirectly connected through an intermediary.For this
For the those of ordinary skill in field, the specific meanings of the above terms in the present invention can be understood according to specific conditions.
In description of the invention, it is to be understood that the orientation or positional relationship of the instructions such as term "upper", "lower" be based on
Orientation or positional relationship shown in the drawings, is merely for convenience of description of the present invention and simplification of the description, rather than indicates or imply institute
The device or unit of finger must have specific direction, with specific azimuth configuration and operation, it is thus impossible to be interpreted as to this hair
Bright limitation.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, protection scope of the present invention should be based on the protection scope of the described claims.
Claims (18)
1. weather prognosis system, including：
First unit obtains training sample based on history radar echo map；
Second unit builds neural network model；
Third unit obtains the radar echo map before the moment to be predicted；
Wherein, the second unit is trained the neural network model with the training sample；
The neural network that the second unit is completed using the radar echo map before the moment to be predicted obtained as training
The input of model, so that the neural network model exports the radar echo map at moment to be predicted.
2. system according to claim 1, which is characterized in that further include：
Unit the 4th, Unit the 4th pre-process the history radar echo map, described in the first unit acquisition
For the history radar echo map after the pretreatment of Unit the 4th as training sample, the pretreatment includes returning history radar
Original echo value in wave figure divided by the setting numerical value more than 1, to reduce the sparse degree of solution space；
Unit the 5th, the radar echo map before the moment to be predicted that Unit the 5th obtains the third unit carry out pre-
Processing, the radar echo map before the moment to be predicted after the pretreatment of Unit the 5th are completed described as training
The input of neural network model, the pretreatment includes removing the original echo value in the radar echo map before the moment to be predicted
With the setting numerical value more than 1, to reduce the sparse degree of solution space；
The radar of Unit the 6th, the moment to be predicted that the neural network model that Unit the 6th completes training exports returns
The original echo value of wave figure is multiplied by the setting numerical value.
3. system according to claim 2, which is characterized in that preferably, the pretreatment and the 5th of Unit the 4th
The pretreatment of unit further includes by the Echo Rating in the strong noise area in the radar echo map within the scope of radar center certain radius
Zero setting.
4. system according to claim 1 or 2, which is characterized in that the second unit is with multiple continuous radar returns
Scheme the input as the neural network model.
5. system according to claim 1 or 2, which is characterized in that the neural network model is the multilayer of U-Net structures
Convolutional layer and multilayer convolutional layer and multilayer warp lamination that multilayer warp lamination or the neural network model are HED structures.
6. according to claim 1 or the system, which is characterized in that the neural network model extracts thunder using network deep layer
Up to the general characteristic of reflectogram, local feature is extracted using network shallow-layer, the general characteristic and the local feature are mutually tied
It closes, fitting cloud layer movement.
7. according to claim 1 or the system, which is characterized in that the neural network model is opened based on Google's deep learning
Source frame Tensorflow is built.
8. according to claim 1 or the system, which is characterized in that the radar echo map of the neural network model prediction
Mean square deviation with sample is as loss function training pattern.
9. Predictive meteorological methods include the following steps：
Training sample is obtained, the training sample is obtained based on history radar echo map；
Neural network model is trained with the training sample；
Obtain the radar echo map before the moment to be predicted；
The input for the neural network model completed using the radar echo map before the moment to be predicted of acquisition as training, institute
State the radar echo map of neural network model output subsequent time.
10. system according to claim 9, which is characterized in that
Training sample is used as after the history radar echo map is preprocessed, the pretreatment includes by history radar echo map
In original echo value divided by setting numerical value more than 1, to reduce the sparse degree of solution space；
The neural network that radar echo map before the moment to be predicted obtained is completed after being pre-processed as training
The input of model, it is described pretreatment include by the radar echo map before the moment to be predicted original echo value divided by be more than 1
Setting numerical value, to reduce the sparse degree of solution space；
The original echo value of the radar echo map at the moment to be predicted for the neural network model output that training is completed is multiplied by
The setting numerical value.
11. system according to claim 10, which is characterized in that the pretreatment further includes inputting the nerve net
The Echo Rating zero setting in the strong noise area in the radar echo map of network model within the scope of radar center certain radius.
12. system according to claim 9 or 10, which is characterized in that using multiple continuous radar echo maps described in
The input of neural network model.
13. system according to claim 9 or 10, which is characterized in that the neural network model is the more of U-Net structures
Layer convolutional layer and multilayer convolutional layer and multilayer warp lamination that multilayer warp lamination or the neural network model are HED structures.
14. system according to claim 9 or 10, which is characterized in that the neural network model is carried using network deep layer
The general characteristic for taking radar echo map extracts local feature, by the general characteristic and the local feature using network shallow-layer
It is combined, fitting cloud layer movement.
15. system according to claim 9 or 10, which is characterized in that the neural network model is based on Google's depth
Open Framework Tensorflow is practised to build.
16. system according to claim 9 or 10, which is characterized in that the radar of the neural network model prediction returns
The mean square deviation of wave figure and sample is as loss function training pattern.
17. a kind of computer readable storage medium, is stored thereon with computer program, power is realized when which is executed by processor
Profit requires the step of any one of 9-17 the methods.
18. a kind of computer equipment, including memory, processor and storage are on a memory and the meter that can run on a processor
The step of calculation machine program, the processor realizes any one of claim 9-17 the methods when executing described program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810531260.XA CN108734357A (en) | 2018-05-29 | 2018-05-29 | Weather prognosis system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810531260.XA CN108734357A (en) | 2018-05-29 | 2018-05-29 | Weather prognosis system and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108734357A true CN108734357A (en) | 2018-11-02 |
Family
ID=63935805
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810531260.XA Pending CN108734357A (en) | 2018-05-29 | 2018-05-29 | Weather prognosis system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108734357A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109919862A (en) * | 2019-02-01 | 2019-06-21 | 北京佳格天地科技有限公司 | Radar image denoising system, method and computer equipment |
| CN110135654A (en) * | 2019-05-24 | 2019-08-16 | 北京百度网讯科技有限公司 | Method and apparatus for predicting strong convective weather |
| CN110363327A (en) * | 2019-06-04 | 2019-10-22 | 东南大学 | Short based on ConvLSTM and 3D-CNN faces Prediction of Precipitation method |
| CN110705115A (en) * | 2019-10-10 | 2020-01-17 | 北京北科融智云计算科技有限公司 | Meteorological forecasting method and system based on deep belief network |
| CN110717546A (en) * | 2019-10-15 | 2020-01-21 | 上海眼控科技股份有限公司 | Image distribution similarity calculation method and device, computer equipment and storage medium |
| CN111028260A (en) * | 2019-12-17 | 2020-04-17 | 上海眼控科技股份有限公司 | Image prediction method, image prediction device, computer equipment and storage medium |
| CN111105093A (en) * | 2019-12-23 | 2020-05-05 | 浙江师范大学 | Near precipitation prediction method based on CNN and BilSTM |
| CN111242372A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| CN111239739A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| CN111445460A (en) * | 2020-03-30 | 2020-07-24 | 上海眼控科技股份有限公司 | Radar cloud picture prediction method and device, computer equipment and storage medium |
| CN111507929A (en) * | 2020-04-15 | 2020-08-07 | 上海眼控科技股份有限公司 | Meteorological cloud picture prediction method and device, computer equipment and storage medium |
| CN112180375A (en) * | 2020-09-14 | 2021-01-05 | 成都信息工程大学 | Meteorological radar echo extrapolation method based on improved TrajGRU network |
| CN112819199A (en) * | 2020-12-31 | 2021-05-18 | 上海眼控科技股份有限公司 | Precipitation prediction method, device, equipment and storage medium |
| CN113239614A (en) * | 2021-04-22 | 2021-08-10 | 西北工业大学 | Atmospheric turbulence phase space-time prediction algorithm |
| CN113298846A (en) * | 2020-11-18 | 2021-08-24 | 西北工业大学 | Intelligent interference detection method based on time-frequency semantic perception |
| US11131790B2 (en) | 2019-03-25 | 2021-09-28 | Yandex Europe Ag | Method of and system for generating weather forecast |
| CN114365153A (en) * | 2019-10-04 | 2022-04-15 | 国际商业机器公司 | Weather-forecast radar image |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100245165A1 (en) * | 2009-03-31 | 2010-09-30 | Honeywell International Inc. | Systems and methods for assessing weather in proximity to an airborne aircraft |
| CN105425319A (en) * | 2015-09-16 | 2016-03-23 | 河海大学 | Rainfall satellite rainstorm assimilation method based on ground measuring data correction |
| CN106651886A (en) * | 2017-01-03 | 2017-05-10 | 北京工业大学 | Cloud image segmentation method based on superpixel clustering optimization CNN |
| CN106886023A (en) * | 2017-02-27 | 2017-06-23 | 中国人民解放军理工大学 | A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks |
| CN107121679A (en) * | 2017-06-08 | 2017-09-01 | 湖南师范大学 | Recognition with Recurrent Neural Network predicted method and memory unit structure for Radar Echo Extrapolation |
| CN107632295A (en) * | 2017-09-15 | 2018-01-26 | 广东工业大学 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
-
2018
- 2018-05-29 CN CN201810531260.XA patent/CN108734357A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100245165A1 (en) * | 2009-03-31 | 2010-09-30 | Honeywell International Inc. | Systems and methods for assessing weather in proximity to an airborne aircraft |
| CN105425319A (en) * | 2015-09-16 | 2016-03-23 | 河海大学 | Rainfall satellite rainstorm assimilation method based on ground measuring data correction |
| CN106651886A (en) * | 2017-01-03 | 2017-05-10 | 北京工业大学 | Cloud image segmentation method based on superpixel clustering optimization CNN |
| CN106886023A (en) * | 2017-02-27 | 2017-06-23 | 中国人民解放军理工大学 | A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks |
| CN107121679A (en) * | 2017-06-08 | 2017-09-01 | 湖南师范大学 | Recognition with Recurrent Neural Network predicted method and memory unit structure for Radar Echo Extrapolation |
| CN107632295A (en) * | 2017-09-15 | 2018-01-26 | 广东工业大学 | A kind of Radar Echo Extrapolation method based on sequential convolutional neural networks |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109919862B (en) * | 2019-02-01 | 2021-08-20 | 北京佳格天地科技有限公司 | Radar image denoising system and method and computer equipment |
| CN109919862A (en) * | 2019-02-01 | 2019-06-21 | 北京佳格天地科技有限公司 | Radar image denoising system, method and computer equipment |
| RU2757591C1 (en) * | 2019-03-25 | 2021-10-19 | Общество С Ограниченной Ответственностью «Яндекс» | Method and system for generating weather forecast |
| US11131790B2 (en) | 2019-03-25 | 2021-09-28 | Yandex Europe Ag | Method of and system for generating weather forecast |
| KR102337776B1 (en) | 2019-05-24 | 2021-12-09 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Method and apparatus for predicting severe convection weather |
| CN110135654A (en) * | 2019-05-24 | 2019-08-16 | 北京百度网讯科技有限公司 | Method and apparatus for predicting strong convective weather |
| US11555921B2 (en) | 2019-05-24 | 2023-01-17 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for predicting severe convection weather |
| KR20200135125A (en) * | 2019-05-24 | 2020-12-02 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Method and apparatus for predicting severe convection weather |
| CN110135654B (en) * | 2019-05-24 | 2021-08-27 | 北京百度网讯科技有限公司 | Method and apparatus for predicting strong convection weather |
| CN110363327B (en) * | 2019-06-04 | 2021-09-03 | 东南大学 | ConvLSTM and 3D-CNN-based short rainfall prediction method |
| CN110363327A (en) * | 2019-06-04 | 2019-10-22 | 东南大学 | Short based on ConvLSTM and 3D-CNN faces Prediction of Precipitation method |
| CN114365153B (en) * | 2019-10-04 | 2023-04-18 | 国际商业机器公司 | Weather-predicted radar images |
| CN114365153A (en) * | 2019-10-04 | 2022-04-15 | 国际商业机器公司 | Weather-forecast radar image |
| CN110705115B (en) * | 2019-10-10 | 2023-05-05 | 北京北科融智云计算科技有限公司 | Weather forecast method and system based on deep belief network |
| CN110705115A (en) * | 2019-10-10 | 2020-01-17 | 北京北科融智云计算科技有限公司 | Meteorological forecasting method and system based on deep belief network |
| CN110717546A (en) * | 2019-10-15 | 2020-01-21 | 上海眼控科技股份有限公司 | Image distribution similarity calculation method and device, computer equipment and storage medium |
| CN111028260A (en) * | 2019-12-17 | 2020-04-17 | 上海眼控科技股份有限公司 | Image prediction method, image prediction device, computer equipment and storage medium |
| CN111105093B (en) * | 2019-12-23 | 2023-04-18 | 浙江师范大学 | Near precipitation prediction method based on CNN and BilSTM |
| CN111105093A (en) * | 2019-12-23 | 2020-05-05 | 浙江师范大学 | Near precipitation prediction method based on CNN and BilSTM |
| CN111242372A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| CN111239739A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| CN111445460A (en) * | 2020-03-30 | 2020-07-24 | 上海眼控科技股份有限公司 | Radar cloud picture prediction method and device, computer equipment and storage medium |
| CN111507929A (en) * | 2020-04-15 | 2020-08-07 | 上海眼控科技股份有限公司 | Meteorological cloud picture prediction method and device, computer equipment and storage medium |
| CN112180375B (en) * | 2020-09-14 | 2022-12-20 | 成都信息工程大学 | Weather radar echo extrapolation method based on improved TrajGRU network |
| CN112180375A (en) * | 2020-09-14 | 2021-01-05 | 成都信息工程大学 | Meteorological radar echo extrapolation method based on improved TrajGRU network |
| CN113298846A (en) * | 2020-11-18 | 2021-08-24 | 西北工业大学 | Intelligent interference detection method based on time-frequency semantic perception |
| CN113298846B (en) * | 2020-11-18 | 2024-02-09 | 西北工业大学 | Interference intelligent detection method based on time-frequency semantic perception |
| CN112819199A (en) * | 2020-12-31 | 2021-05-18 | 上海眼控科技股份有限公司 | Precipitation prediction method, device, equipment and storage medium |
| CN113239614A (en) * | 2021-04-22 | 2021-08-10 | 西北工业大学 | Atmospheric turbulence phase space-time prediction algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108734357A (en) | Weather prognosis system and method | |
| CN107703564B (en) | Rainfall prediction method and system and electronic equipment | |
| CN111160525B (en) | Task unloading intelligent decision-making method based on unmanned aerial vehicle group in edge computing environment | |
| CN109493599A (en) | A kind of Short-time Traffic Flow Forecasting Methods based on production confrontation network | |
| CN109492334A (en) | Delayed method for establishing model, prediction technique and device | |
| CN111461325B (en) | Multi-target layered reinforcement learning algorithm for sparse rewarding environmental problem | |
| CN109919862B (en) | Radar image denoising system and method and computer equipment | |
| CN103578119A (en) | Target detection method in Codebook dynamic scene based on superpixels | |
| CN115879535A (en) | Training method, device, equipment and medium for automatic driving perception model | |
| CN107731011B (en) | Port berthing monitoring method and system and electronic equipment | |
| CN109462821A (en) | Method, apparatus, storage medium and the electronic equipment of predicted position | |
| CN113112791A (en) | Traffic flow prediction method based on sliding window long-and-short term memory network | |
| CN113343858B (en) | Road network geographic position identification method and device, electronic equipment and storage medium | |
| CN104850677A (en) | Deep learning cellular automaton simulation analysis method based on geographic information system | |
| CN115204478A (en) | Public traffic flow prediction method combining urban interest points and space-time causal relationship | |
| US9806528B2 (en) | Systems and methods for estimating net solar energy production for airborne photovoltaic systems | |
| CN109460874A (en) | A kind of ariyoshi wave height prediction technique based on deep learning | |
| CN107330538B (en) | Method for compiling reservoir adaptive scheduling rules under climate change condition | |
| CN114715145A (en) | Trajectory prediction method, device and equipment and automatic driving vehicle | |
| CN109299825A (en) | A kind of prediction technique and forecasting system based on the real-time passenger flow of rail traffic | |
| Moraguez et al. | Convolutional neural network for detection of residential photovoltalc systems in satellite imagery | |
| Hong et al. | A study on rainfall prediction based on meteorological time series | |
| CN115423810B (en) | Blade icing form analysis method for wind generating set | |
| CN112884831B (en) | Method for extracting long-term static characteristics of indoor parking lot based on probability mask | |
| CN113344277A (en) | Prediction model training method, state updating method, device, equipment and medium |
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 |