CN116451881A - Short-time precipitation prediction method based on MSF-Net network model - Google Patents
Short-time precipitation prediction method based on MSF-Net network model Download PDFInfo
- Publication number
- CN116451881A CN116451881A CN202310715521.4A CN202310715521A CN116451881A CN 116451881 A CN116451881 A CN 116451881A CN 202310715521 A CN202310715521 A CN 202310715521A CN 116451881 A CN116451881 A CN 116451881A
- Authority
- CN
- China
- Prior art keywords
- data
- precipitation
- prediction
- time
- module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001556 precipitation Methods 0.000 title claims abstract description 118
- 238000000034 method Methods 0.000 title claims abstract description 45
- 230000004927 fusion Effects 0.000 claims abstract description 51
- 238000000605 extraction Methods 0.000 claims abstract description 36
- 238000002310 reflectometry Methods 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 16
- 238000011160 research Methods 0.000 claims abstract description 12
- 238000006243 chemical reaction Methods 0.000 claims description 16
- 238000010586 diagram Methods 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 4
- 239000000284 extract Substances 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 4
- 238000001704 evaporation Methods 0.000 claims description 3
- 230000008020 evaporation Effects 0.000 claims description 3
- 230000004907 flux Effects 0.000 claims description 3
- 230000003993 interaction Effects 0.000 claims description 3
- 239000007788 liquid Substances 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000005855 radiation Effects 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000005068 transpiration Effects 0.000 claims description 3
- 239000013598 vector Substances 0.000 claims description 3
- 230000003213 activating effect Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 5
- 238000013135 deep learning Methods 0.000 description 7
- 239000011159 matrix material Substances 0.000 description 4
- 238000002679 ablation Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000013277 forecasting method Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012876 topography Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002592 echocardiography Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000000053 physical method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Human Resources & Organizations (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Strategic Management (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a short-time precipitation prediction method based on an MSF-Net network model, which comprises the following steps: generating a precipitation prediction data set; an original prediction model is built based on an MSF-Net network, and the original prediction model comprises an input module, a meteorological feature extraction module, an attention fusion prediction module and an output module; training an original prediction model by using a rainfall prediction data set; calculating the loss of the original prediction model by using the loss function, and optimizing the training network parameters to obtain a short-time rainfall prediction model after training; GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map data and DEM elevation data in a research area acquired in real time are input into a short-time precipitation prediction model, and a corresponding short-time precipitation prediction result is output. According to the invention, live precipitation and multisource data are effectively fused, and the short-time precipitation prediction effect is improved.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence intelligent meteorology, and particularly relates to a short-time precipitation prediction method based on an MSF-Net network model.
Background
Precipitation is an important source of earth water resources, has accurate and timely short-time precipitation forecasting capability, can effectively avoid social and economic life risks, and can provide better decision basis for production planning. The traditional precipitation prediction method utilizes the characteristics of frequency distribution, spatial distribution and the like of historical precipitation to carry out statistical analysis, and predicts the future precipitation based on a physical model; the numerical model simulates the weather gradual change process by various observation data and other characteristics, and is the basis for manufacturing weather forecast by the weather stations in all places of the world at present. However, the conventional method consumes a lot of resources and is easily interfered by various factors, and it has been difficult to meet the current demands. Therefore, research into short-term precipitation prediction is becoming more important. The rise of artificial intelligence technology provides a new opportunity for the development of objective forecasting methods, and more researches are applied to the meteorological field due to the fact that deep learning has strong data processing capability. Along with the continuous enhancement of the observation means of the weather system, the high-resolution and high-frequency meteorological observation data are continuously increased, and the potential of short-time precipitation prediction by utilizing multi-source data is also continuously increased. The accuracy of the current prediction method for precipitation through fusion of various meteorological data is still low.
Deep learning is capable of automatically learning an effective feature representation from data, and is generally used to solve the problem of insufficient fitting capability of traditional machine learning in a high-dimensional feature space. The data driving prediction model based on deep learning does not depend on the traditional physical framework, and can simulate nonlinear precipitation which is difficult to predict by the traditional method well through data training models such as radars, satellites and the like and predicting the motion process of echoes. The deep learning has wide application prospect in short-time precipitation prediction, so that the realization of more efficient and accurate short-time precipitation prediction has important significance. Thus, there is a need for an efficient method that can apply deep learning in predicting short-term precipitation from multi-source data.
Disclosure of Invention
The technical problems to be solved are as follows: the invention provides a short-time precipitation prediction method based on an MSF-Net network model, which integrates the precipitation characteristics of GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data, considers the influence of DEM elevation data on precipitation distribution, and effectively improves the short-time precipitation prediction effect.
The technical scheme is as follows:
a short-time precipitation prediction method based on an MSF-Net network model comprises the following steps:
s1, GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map data and DEM elevation data in a research area are obtained, all the obtained data are in the same scale after bilinear interpolation preprocessing, a precipitation prediction data set is generated, and the precipitation prediction data set is divided into a training set and a verification set according to years;
s2, constructing an original prediction model based on an MSF-Net network, wherein the original prediction model comprises an input module, a meteorological feature extraction module, an attention fusion prediction module and an output module;
the input module is used for inputting the received GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data into the meteorological feature extraction module in a time sequence mode to serve as a first input, convolving DEM elevation data, and inputting the convolved feature map into the attention fusion prediction module to serve as a second input;
the meteorological feature extraction module performs feature extraction on space-time information affecting precipitation in all input data, and outputs corresponding meteorological features to the attention fusion prediction module;
the attention fusion prediction module is used for fusing the characteristic diagram after the DEM elevation data are convolved with the meteorological characteristics learned by the meteorological characteristic extraction module to obtain a corresponding rainfall prediction result, and the rainfall prediction result is output through the output module; the attention fusion prediction module comprises an attention fusion module and a fusion prediction module;
s3, GPM precipitation grid point data of 12 time dimensions from t-11 to t, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data are used as first input in a time sequence mode, DEM elevation data are used as second input in a single time mode, GPM precipitation grid point data of 6 time dimensions from t+1 to t+6 are used as labels to be output, and an original prediction model is trained by using a precipitation prediction data set;
s4, calculating the loss of the original prediction model by using the loss function, and optimizing the training network parameters to obtain a short-time rainfall prediction model after training;
s5, GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map data and DEM elevation data in the research area obtained in real time are input into a short-time precipitation prediction model, and a corresponding short-time precipitation prediction result is output.
Further, in step S1, the spatial resolution of the GPM precipitation grid point data isThe time resolution is 0.5h, which is precipitation live data calibrated by using a rain gauge and ground observation data;
the spatial resolution of the ERA5 meteorological factor data is thatThe system comprises 13 meteorological factors including u and v components of 100m wind, 2m temperature and 2m dew point temperature, cloud bottom height, relative humidity, total column cloud liquid water, total column cloud vapor, atmospheric pressure, net heat radiation, latent heat flux, vegetation leaf area and vegetation transpiration evaporation;
the Doppler Lei Dazhen color basic reflectivity map data has the resolution ofThe system is used for reflecting rainfall areas and trend information;
the spatial resolution of the DEM elevation data isIncluding grade, slope direction, terrain height and altitude.
Further, in step S2, the meteorological feature extraction module includes a compiling layer, a converting layer and an interpretation layer;
the compiling layer convolves the input GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data in the channel dimension, extracts spatial features and increases the channel, and the operation steps of the compiling layer are expressed as follows by formulas:
wherein the method comprises the steps ofGPM precipitation grid point data, doppler Lei Dazhen color basic reflectivity map data and ERA5 meteorological factor data representing times t-11 to t,the output of the compilation layer is represented,indicating that the non-linear activation is to be performed,the normalization is indicated by the fact that,the method is characterized in that the downsampling is performed after 2-dimensional convolution, and the convolution kernel size is 1 multiplied by 1;
the conversion layer stacks three conversion modules, local features and global features of images are extracted by means of depth separable expansion convolution with different sizes, changes of meteorological elements in time and space are learned, the number of parameters of a model is reduced, and interaction of information among channels is obtained by means of channel convolution, wherein the operation steps of the conversion layer are expressed as follows:
wherein the method comprises the steps ofRepresenting the first of the conversion layersThe feature map, i.e. the preceding feature mapThe output after the operation is performed,；a channel-by-channel convolution is represented,the layer of the representation is standardized and,the depth separable convolution results Concat which represent the convolution kernel sizes of 3 and 7 are respectively fused, and the characteristics are output through a conversion layer；
The interpretation layer reconstructs the size of the input feature map by utilizing deconvolution, compresses the channel to 1 dimension and then outputs the feature map, and the operation steps of the interpretation layer are expressed as follows by formulas:
wherein the method comprises the steps ofA compile layer feature map corresponding to the path connection between interpretation layer images is represented,，representing up-sampling after a 2-dimensional convolution, the convolution kernel size is 1 x 1,representing the output characteristics of the interpretation layer.
Further, in step S2, the attention fusion module is composed of a channel attention unit focusing on deep features of the weather and a spatial attention unit focusing on spatial and temporal information and variation trend of the weather elements, and the fusion prediction module is used for fusing the elevation data of the DEM with the weather features output by the weather feature extraction module;
the operation steps of the attention fusion prediction module are expressed as follows:
wherein the method comprises the steps ofRepresenting the output of the attention fusion prediction module,andrespectively represent the characteristic diagrams of the meteorological characteristics learned by the meteorological characteristic extraction module and the convolution of the DEM elevation data,representing the multiplication by element,representing the operational functions of the attention fusion module,representing image stitching.
Further, the operation steps of the spatial attention unit include:
the method comprises the steps of splicing input feature images after global average pooling and global maximum pooling are used on channels, extracting complex features of the images on different receptive fields by two depth separable expansion convolutions with the size of 3 and the expansion factors of 2 and 3 respectively, and outputting feature images with the same size by filling operation during convolution; and splicing the two learned features, activating by sigmoid to obtain a single-channel feature map containing weather information with different scales, and multiplying the single-channel feature map with the input original map element by element to obtain the feature map with the same input size.
Further, the operation steps of the channel attention unit include:
fusing two 3 multiplied by 3 convolutions with the step length of 1 and 3 on the input feature map, and then adopting global average pooling on the fused feature map to compress space information to obtain a vector only containing channel information; and adopting a point-by-point convolution learning characteristic, compressing the channel, and obtaining a channel attention characteristic diagram of Cx1×1 through a Sigmoid activation function, and multiplying the channel attention characteristic diagram with an input original diagram element by element to obtain the output of a channel attention part.
Further, the fusion prediction module multiplies the characteristic image output by the attention fusion module with the characteristic image output by the meteorological characteristic extraction module and the characteristic image after the DEM elevation data convolution by elements, and finally fuses the two characteristic images and outputs a short-time rainfall prediction result.
Further, the input module inputs GPM precipitation grid point data, ERA5 weather factor data and Doppler Lei Dazhen color basic reflectivity map data of 12 time dimensions from t-11 time to t time into the weather feature extraction module in an hour-by-hour mode on the same spatial scale to perform feature extraction, and the output module outputs a 6-channel feature map, wherein the 6-channel feature map comprises hour-by-hour short-time precipitation predictions from t+1 time to t+6 time.
The beneficial effects are that:
firstly, the short-time precipitation prediction method based on the MSF-Net model effectively integrates the GPM precipitation data, ERA5 meteorological factors and the space-time characteristics of the information related to precipitation in the radar chart, fully considers the influence of the DEM elevation data on precipitation amount distribution, effectively improves the short-time precipitation prediction effect,
secondly, according to the short-time precipitation prediction method based on the MSF-Net model, the designed meteorological feature extraction module can fully extract local information and global information of meteorological features in multi-source data, and loss of detail information is reduced.
Thirdly, according to the short-time precipitation prediction method based on the deep learning MSF-Net model, the designed attention fusion prediction model can strengthen deep information of multi-source data meteorological features in two aspects of channels and spaces, and is fused with DEM elevation data, and the model fully considers precipitation distribution by combining with terrain factors so as to improve short-time precipitation prediction precision.
Drawings
FIG. 1 is a flow chart of a short-time precipitation prediction method based on an MSF-Net model in an embodiment of the invention.
Fig. 2 is a network model overall structure diagram of a short-time precipitation prediction method based on an MSF-Net model according to an embodiment of the present invention.
FIG. 3 is a block diagram of a weather feature extraction module ME-Net according to an embodiment of the present invention.
Fig. 4 is a network configuration diagram of a conversion layer according to an embodiment of the present invention.
Fig. 5 is a network structure diagram of an attention fusion prediction module AFPM according to an embodiment of the present invention.
Fig. 6 is a network structure diagram of an attention fusion module AFM according to an embodiment of the present invention.
Detailed Description
The following examples will provide those skilled in the art with a more complete understanding of the invention, but are not intended to limit the invention in any way.
The invention discloses a short-time precipitation prediction method based on an MSF-Net model, which is constructed based on an MSF-Net network, and comprises an input module, a meteorological feature extraction module ME-Net (Meteorological Extraction Network), an attention fusion prediction module AFPM (Attention Fuse Prediction Module) and an output module as shown in figure 2.
Referring to fig. 1 and 2, the short-time precipitation prediction method includes the steps of:
s1, GPM precipitation grid point data, ERA5 meteorological factor data, doppler radar map data and DEM elevation data in a research area are obtained, all the data are in the same scale after bilinear interpolation preprocessing is carried out on the data, a precipitation prediction data set is generated, the data set is divided into a training set and a verification set according to years, and step S2 is carried out.
In practical application, the GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map and DEM elevation data have the following characteristics: the spatial resolution of the GPM precipitation grid point data isThe time resolution is 0.5h, which is precipitation live data calibrated by using a rain gauge and ground observation data; the spatial resolution of ERA5 meteorological factor data isThe system comprises 13 meteorological factors including u and v components of 100m wind, 2m temperature and 2m dew point temperature, cloud bottom height, relative humidity, total column cloud liquid water, total column cloud vapor, atmospheric pressure, net heat radiation, latent heat flux, vegetation leaf area and vegetation transpiration evaporation; the resolution of the Doppler Lei Dazhen color primary reflectivity map isThe information such as precipitation areas and trends can be reflected; the spatial resolution of the DEM elevation data isIncluding grade, slope direction, terrain height and altitude.
S2, splicing GPM precipitation grid point data, ERA5 weather factor data and Doppler Lei Dazhen color basic reflectivity map data Concat in the step S1, inputting the spliced GPM precipitation grid point data, ERA5 weather factor data and Doppler Lei Dazhen color basic reflectivity map data Concat into a weather feature extraction module ME-Net in a time sequence mode, extracting features of space-time information affecting precipitation in all data, and then entering the step S3.
Referring to fig. 3, the weather feature extraction module ME-Net is composed of a compiling layer, a converting layer and an interpreting layer. The meteorological feature extraction module ME-Net is used for extracting meteorological features in GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data.
Specifically, the compiling layer convolves the input three image data in the channel dimension, extracts the spatial features and enlarges the channel to make the converting layer better learn the features, and the operation steps of the compiling layer can be expressed as follows:
wherein the method comprises the steps ofGPM precipitation grid point data, doppler Lei Dazhen color basic reflectivity map data and ERA5 meteorological factor data representing times t-11 to t,the output of the compilation layer is represented,indicating that the non-linear activation is to be performed,the normalization is indicated by the fact that,representing downsampling after a 2-dimensional convolution, the convolution kernel size is 1 x 1.
Referring to fig. 4, the conversion layer stacks three conversion modules, firstly, extracting local features and global features of an image by using depth separable expansion convolution with different sizes, learning the change of meteorological elements in time and space, reducing the parameter quantity of a model to improve the training speed, and then obtaining the interaction of information among channels by using channel convolution, wherein the operation steps of the conversion layer can be expressed as follows:
wherein the method comprises the steps ofRepresenting the first of the conversion layersThe feature map, i.e. the preceding feature mapThe output after the operation is performed,，is a convolution from channel to channel,the layer of the representation is standardized and,the depth separable convolution results Concat which represent the convolution kernel sizes of 3 and 7 are respectively fused, and the characteristics are output through a conversion layer。
The interpretation layer reconstructs the size of the input feature map by utilizing deconvolution, compresses the channel to 1 dimension and then outputs the feature map, and the operation steps of the interpretation layer can be expressed as follows:
wherein the method comprises the steps ofA compile layer feature map corresponding to the path connection between interpretation layer images is represented,，representing up-sampling after a 2-dimensional convolution, the convolution kernel size is 1 x 1,an output feature map of the interpretation layer is shown.
And S3, respectively inputting the feature map obtained by convolving the DEM elevation data and the meteorological features learned in the step S2 into an attention fusion prediction module AFPM, focusing the model on the region with rich rainfall features and the space-time information of meteorological elements contained in the region by two different attentions, taking the influence of the topography on the precipitation distribution into consideration, fusing the extracted feature map Add, outputting a precipitation prediction result, and then entering the step S4.
Referring to fig. 5, the attention fusion prediction module AFPM includes an attention fusion module AFM (Attention Fuse Module) and a fusion prediction module, where the attention fusion module AFM is composed of channel attention focusing on deep features of weather and spatial attention focusing on space-time information and variation trend of weather elements, and the fusion prediction module is used to fuse a feature map after the elevation data convolution of DEM with output features of the weather feature extraction module ME-Net, and finally output a short-time precipitation prediction result. The attention fusion prediction module AFPM enables the short-time precipitation prediction model to pay attention to the region with obvious precipitation characteristics and the change trend thereof, and fully considers the influence of the terrain on precipitation distribution. The operational step characteristics of the attention fusion prediction module AFPM may be formulated as:
wherein the method comprises the steps ofRepresenting the output of the attention fusion prediction module AFPM,andrespectively representing the weather feature map learned by the weather feature extraction module ME-Net and the feature map obtained by the elevation data convolution of the DEM, respectively inputting the two data images into the fusion prediction module,representing the multiplication by element,representing the operating functions of the attention fusion module AFM,representing image stitching, inputting the feature map after Concat into the attention comprising two parts of spatial attention and channel attentionFusion module AFM.
Referring to fig. 6, spatial attention is first spliced on a channel by using global average pooling and global maximum pooling on an input feature map, then complex features of images are extracted on different receptive fields by depth separable expansion convolution with the size of 3 and the expansion factors of 2 and 3 respectively, a filling operation is used during convolution, feature maps with the same size are output, two learned feature maps are activated by sigmoid to obtain a single-channel feature map containing weather information with different scales, and the single-channel feature map is multiplied element by element with an input original map to obtain the feature map with the same input size.
The channel attention is first fused by two 3X 3 convolutions with step sizes of 1 and 3, then global average pooling is adopted for the fused feature images, space information is compressed to obtain vectors only containing channel information, then the point-by-point convolution is used for learning features, channel is compressed, channel attention feature images of Cx1×1 are obtained through a Sigmoid activation function, and the channel attention feature images are multiplied by the input original images element by element to obtain the output of a channel attention part. And finally, fusing the features Add learned by the spatial attention and the channel attention and outputting the fused features Add to a fusion prediction module.
And in the fusion prediction module, the characteristic image output by the attention fusion module AFM is multiplied by the weather characteristic image learned by the weather characteristic extraction module ME-Net and the characteristic image obtained by the elevation data convolution of the DEM element by element, and finally, the two characteristic images Add are fused and then a short-time rainfall prediction result is output.
S4, the historical precipitation data of the GPM, the weather factor data of ERA5 and the Doppler radar data of 12 time dimensions from the t-11 moment to the t moment are input in a time sequence, the elevation data of the DEM is input in a single moment mode, the GPM precipitation grid point data of 6 h at the t+1 moment to the t+6 moment are used as labels, a network model is built for training, and then step S5 is carried out.
And S5, calculating the loss of the original prediction model by using a loss function according to the training model constructed in the step S4, optimizing training network parameters to obtain a short-time precipitation prediction model, and then entering the step S6.
S6, GPM precipitation grid point data, ERA5 meteorological factor data, doppler radar map data and DEM elevation data in the research area are acquired in real time and are input into a short-time precipitation prediction model, and a corresponding short-time precipitation prediction result is output.
Experimental protocol and results
We choose (97.2-110E, 20.4-30N) as the research area, mainly including regions such as Yunnan, guizhou, guangxi and Sichuan, and the mountain area is more plateau and complex in topography, so that our precipitation prediction is more challenging. And taking data of 6-8 months in 2017-2019 as a training set and data of 6-8 months in 2020 as a verification set, and carrying out experiments on different forecasting methods.
In order to verify that the weather feature extraction module ME-Net scheme and the attention fusion prediction module AFPM scheme effectively promote the short-time precipitation prediction method, the performance of the MSF-Net model is evaluated by using GPM precipitation live as a comparison standard. ERA5 data (ERA 5 scheme), dopplerf radar data (ERA 5-Dopplerf scheme), DEM elevation data (ERA 5-Dopplerf-DEM scheme), a meteorological feature extraction module ME-Net (ME-ERA 5-Dopplerf-DEM scheme) and an attention fusion prediction module AFPM (ME-ERA 5-Dopplerf-DEM-AFPM scheme) are sequentially added into a prediction network to complete an ablation experiment, and a precipitation confusion matrix is constructed according to a prediction result and divided into positive and negative categories, wherein the confusion matrix is shown in table 1:
TABLE 1 precipitation confusion matrix
Calculating two evaluation indexes of a risk score (TS) and a Bias score (Bias) of the prediction results of all grid points in the research area according to the constructed confusion matrix, wherein the TS formula and the Bias formula are as follows:
we divide the precipitation into 6 grades, usingRepresentation of whereinIndicating that there is rain in the predicted area and that the rain level forecast is correct,indicating the number of times that there is rain but the forecast is no rain or the rain level forecast is wrong,and (3) indicating the times of no rain but forecasting to be rainy, adding the rainfall forecasting results of each level, and calculating to obtain TS and Bias as evaluation indexes. The result shows that after ERA5 data, doppler radar data and DEM elevation data are sequentially added on the basis of single GPM precipitation data, TS of the scheme is respectively improved by 0.02, 0.04 and 0.02, bias is increased by 0.03, 0.06 and 0.04; after the weather feature extraction module ME-Net and the attention fusion prediction module AFPM are added, TS is respectively improved by 0.03, bias is increased by 0.04 and 0.03, and specific ablation experimental results are shown in Table 2.
Table 2 ablation experimental results
In order to verify the effect of the MSF-Net network on short-time precipitation prediction, 6-hour GPM precipitation live data in the same time period in a research area is taken as a comparison standard, the prediction result of the MSF-Net is compared with the WRF physical prediction result, the U-Net prediction result, the RNN prediction result and the ConvLSTM prediction result, the result shows that the effect of the MSF-Net based on deep learning on short-time precipitation prediction is better than that of the other four prediction methods, the WRF precipitation prediction based on the physical method is greatly advanced, and the specific comparison experimental result is shown in a table 3.
Table 3 results of comparative experiments
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the invention without departing from the principles thereof are intended to be within the scope of the invention as set forth in the following claims.
Claims (8)
1. The short-time precipitation prediction method based on the MSF-Net network model is characterized by comprising the following steps of:
s1, GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map data and DEM elevation data in a research area are obtained, all the obtained data are in the same scale after bilinear interpolation preprocessing, a precipitation prediction data set is generated, and the precipitation prediction data set is divided into a training set and a verification set according to years;
s2, constructing an original prediction model based on an MSF-Net network, wherein the original prediction model comprises an input module, a meteorological feature extraction module, an attention fusion prediction module and an output module;
the input module is used for inputting the received GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data into the meteorological feature extraction module in a time sequence mode to serve as a first input, convolving DEM elevation data, and inputting the convolved feature map into the attention fusion prediction module to serve as a second input;
the meteorological feature extraction module performs feature extraction on space-time information affecting precipitation in all input data, and outputs corresponding meteorological features to the attention fusion prediction module;
the attention fusion prediction module is used for fusing the characteristic diagram after the DEM elevation data are convolved with the meteorological characteristics learned by the meteorological characteristic extraction module to obtain a corresponding rainfall prediction result, and the rainfall prediction result is output through the output module; the attention fusion prediction module comprises an attention fusion module and a fusion prediction module;
s3, GPM precipitation grid point data of 12 time dimensions from t-11 to t, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data are used as first input in a time sequence mode, DEM elevation data are used as second input in a single time mode, GPM precipitation grid point data of 6 time dimensions from t+1 to t+6 are used as labels to be output, and an original prediction model is trained by using a precipitation prediction data set;
s4, calculating the loss of the original prediction model by using the loss function, and optimizing the training network parameters to obtain a short-time rainfall prediction model after training;
s5, GPM precipitation grid point data, ERA5 meteorological factor data, doppler Lei Dazhen color basic reflectivity map data and DEM elevation data in the research area obtained in real time are input into a short-time precipitation prediction model, and a corresponding short-time precipitation prediction result is output.
2. The short-time precipitation prediction method based on the MSF-Net network model according to claim 1, wherein in step S1, the spatial resolution of the GPM precipitation grid point data isThe time resolution is 0.5h, which is precipitation live data calibrated by using a rain gauge and ground observation data;
the spatial resolution of the ERA5 meteorological factor data is thatThe system comprises 13 meteorological factors including u and v components of 100m wind, 2m temperature and 2m dew point temperature, cloud bottom height, relative humidity, total column cloud liquid water, total column cloud vapor, atmospheric pressure, net heat radiation, latent heat flux, vegetation leaf area and vegetation transpiration evaporation;
the Doppler Lei Dazhen color basic reflectivity map data has the resolution ofThe system is used for reflecting rainfall areas and trend information;
the spatial resolution of the DEM elevation data isIncluding grade, slope direction, terrain height and altitude.
3. The short-time precipitation prediction method based on the MSF-Net network model according to claim 1, wherein in step S2, the meteorological feature extraction module comprises a compiling layer, a conversion layer and an interpretation layer;
the compiling layer convolves the input GPM precipitation grid point data, ERA5 meteorological factor data and Doppler Lei Dazhen color basic reflectivity map data in the channel dimension, extracts spatial features and increases the channel, and the operation steps of the compiling layer are expressed as follows by formulas:
；
wherein the method comprises the steps ofGPM precipitation grid data, doppler Lei Dazhen color basic reflectivity map data and ERA5 meteorological factor data representing time t-11 to time t +.>Output representing compilation layer, +.>Indicating non-linear activation, ++>The normalization is indicated by the fact that,the method is characterized in that the downsampling is performed after 2-dimensional convolution, and the convolution kernel size is 1 multiplied by 1;
the conversion layer stacks three conversion modules, local features and global features of images are extracted by means of depth separable expansion convolution with different sizes, changes of meteorological elements in time and space are learned, the number of parameters of a model is reduced, and interaction of information among channels is obtained by means of channel convolution, wherein the operation steps of the conversion layer are expressed as follows:
；
wherein the method comprises the steps ofRepresenting the%>A feature map, i.e. for its previous feature map +.>Output after operation, ++>；Representing a channel-by-channel convolution, ">Representation layer normalization->The depth separable convolution results Concat which indicate the convolution kernel sizes of 3 and 7 are respectively fused, and the characteristic is output through a conversion layer>；
The interpretation layer reconstructs the size of the input feature map by utilizing deconvolution, compresses the channel to 1 dimension and then outputs the feature map, and the operation steps of the interpretation layer are expressed as follows by formulas:
；
wherein the method comprises the steps ofRepresenting a compiled layer feature map corresponding to a channel connection between interpretation layer images, < >>，Representing up-sampling after a 2-dimensional convolution, the convolution kernel size is 1 x 1,/for>Representing the output characteristics of the interpretation layer.
4. The short-time precipitation prediction method based on the MSF-Net network model according to claim 1, wherein in the step S2, the attention fusion module consists of a channel attention unit focusing on the deep characteristics of the weather and a space attention unit focusing on the space-time information and the change trend of the weather elements, and the fusion prediction module is used for fusing the elevation data of the DEM with the weather characteristics output by the weather characteristic extraction module;
the operation steps of the attention fusion prediction module are expressed as follows:
；
wherein the method comprises the steps ofRepresenting the output of the attention fusion prediction module, +.>And->Respectively representing the characteristic diagram after the convolution of the meteorological characteristics learned by the meteorological characteristic extraction module and the DEM elevation data,/and%>Representing element-by-element multiplication>Representing the operating function of the attention fusion module, +.>Representing image stitching.
5. The method for short-term precipitation prediction based on an MSF-Net network model according to claim 4, wherein the step of operating the spatial attention unit comprises:
the method comprises the steps of splicing input feature images after global average pooling and global maximum pooling are used on channels, extracting complex features of the images on different receptive fields by two depth separable expansion convolutions with the size of 3 and the expansion factors of 2 and 3 respectively, and outputting feature images with the same size by filling operation during convolution; and splicing the two learned features, activating by sigmoid to obtain a single-channel feature map containing weather information with different scales, and multiplying the single-channel feature map with the input original map element by element to obtain the feature map with the same input size.
6. The method for short-term precipitation prediction based on an MSF-Net network model according to claim 4, wherein the operation step of the channel attention unit comprises:
fusing two 3 multiplied by 3 convolutions with the step length of 1 and 3 on the input feature map, and then adopting global average pooling on the fused feature map to compress space information to obtain a vector only containing channel information; and adopting a point-by-point convolution learning characteristic, compressing the channel, and obtaining a channel attention characteristic diagram of Cx1×1 through a Sigmoid activation function, and multiplying the channel attention characteristic diagram with an input original diagram element by element to obtain the output of a channel attention part.
7. The short-term precipitation prediction method based on the MSF-Net network model according to claim 4, wherein the fusion prediction module multiplies the feature images output by the attention fusion module with the feature images output by the meteorological feature extraction module and the feature images after the DEM elevation data convolution by elements respectively, and finally fuses the two feature images to output a short-term precipitation prediction result.
8. The short-time precipitation prediction method based on the MSF-Net network model according to claim 1, wherein the input module inputs GPM precipitation grid point data, ERA5 weather factor data and Doppler Lei Dazhen color basic reflectivity map data of 12 time dimensions from t-11 time to t time into the weather feature extraction module on the same spatial scale hour by hour for feature extraction, and the output module outputs a 6-channel feature map, wherein the feature map comprises hour-by-hour short-time precipitation predictions from t+1 time to t+6 time.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310715521.4A CN116451881B (en) | 2023-06-16 | 2023-06-16 | Short-time precipitation prediction method based on MSF-Net network model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310715521.4A CN116451881B (en) | 2023-06-16 | 2023-06-16 | Short-time precipitation prediction method based on MSF-Net network model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116451881A true CN116451881A (en) | 2023-07-18 |
| CN116451881B CN116451881B (en) | 2023-08-22 |
Family
ID=87132468
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310715521.4A Active CN116451881B (en) | 2023-06-16 | 2023-06-16 | Short-time precipitation prediction method based on MSF-Net network model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116451881B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117368881A (en) * | 2023-12-08 | 2024-01-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
| CN117808650A (en) * | 2024-02-29 | 2024-04-02 | 南京信息工程大学 | Precipitation prediction method based on Transform-Flown and R-FPN |
| CN117909927A (en) * | 2024-03-19 | 2024-04-19 | 中国气象局公共气象服务中心(国家预警信息发布中心) | Precipitation quantitative estimation method and device based on multisource data fusion model |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111489525A (en) * | 2020-03-30 | 2020-08-04 | 南京信息工程大学 | Multi-data fusion meteorological prediction early warning method |
| CN111428676B (en) * | 2020-04-01 | 2023-04-07 | 南京信息工程大学 | Short-term rainfall prediction method based on sparse correspondence and deep neural network |
| CN111475950B (en) * | 2020-04-09 | 2022-11-29 | 首都师范大学 | Method for simulating rainfall flood of concave overpass |
| CN111624682B (en) * | 2020-06-24 | 2021-11-30 | 海南省气象科学研究所 | Quantitative precipitation estimation method based on multi-source data fusion |
| CN114139690A (en) * | 2021-12-09 | 2022-03-04 | 南京邮电大学 | Short-term rainfall prediction method and device |
| CN114881286A (en) * | 2022-04-02 | 2022-08-09 | 大连海事大学 | Short-time rainfall prediction method based on deep learning |
| CN115544889A (en) * | 2022-10-18 | 2022-12-30 | 南京信息工程大学 | Numerical mode precipitation deviation correction method based on deep learning |
-
2023
- 2023-06-16 CN CN202310715521.4A patent/CN116451881B/en active Active
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117368881A (en) * | 2023-12-08 | 2024-01-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
| CN117368881B (en) * | 2023-12-08 | 2024-03-26 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
| CN117808650A (en) * | 2024-02-29 | 2024-04-02 | 南京信息工程大学 | Precipitation prediction method based on Transform-Flown and R-FPN |
| CN117808650B (en) * | 2024-02-29 | 2024-05-14 | 南京信息工程大学 | Precipitation prediction method based on Transform-Flownet and R-FPN |
| CN117909927A (en) * | 2024-03-19 | 2024-04-19 | 中国气象局公共气象服务中心(国家预警信息发布中心) | Precipitation quantitative estimation method and device based on multisource data fusion model |
| CN117909927B (en) * | 2024-03-19 | 2024-06-04 | 中国气象局公共气象服务中心(国家预警信息发布中心) | Precipitation quantitative estimation method and device based on multisource data fusion model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116451881B (en) | 2023-08-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116451881B (en) | Short-time precipitation prediction method based on MSF-Net network model | |
| US11966670B2 (en) | Method and system for predicting wildfire hazard and spread at multiple time scales | |
| US11205028B2 (en) | Estimating physical parameters of a physical system based on a spatial-temporal emulator | |
| Aswin et al. | Deep learning models for the prediction of rainfall | |
| US11720727B2 (en) | Method and system for increasing the resolution of physical gridded data | |
| Larraondo et al. | Automating weather forecasts based on convolutional networks | |
| Xu et al. | AM-ConvGRU: A spatio-temporal model for typhoon path prediction | |
| Zheng et al. | Weather image-based short-term dense wind speed forecast with a ConvLSTM-LSTM deep learning model | |
| Dutt et al. | Artificial intelligence and technology in weather forecasting and renewable energy systems: emerging techniques and worldwide studies | |
| CN114548595A (en) | Strong convection weather physical characteristic quantity prediction method and system based on attention mechanism | |
| Wang et al. | Multi‐scale network for remote sensing segmentation | |
| Han et al. | Fengwu-ghr: Learning the kilometer-scale medium-range global weather forecasting | |
| Charlton-Perez et al. | Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán | |
| CN117194926A (en) | Method and system for predicting hoisting window period of land wind power base | |
| Li et al. | LandBench 1.0: A benchmark dataset and evaluation metrics for data-driven land surface variables prediction | |
| Ashesh et al. | Accurate and clear quantitative precipitation nowcasting based on a deep learning model with consecutive attention and rain-map discrimination | |
| Zhou et al. | A station-data-based model residual machine learning method for fine-grained meteorological grid prediction | |
| CN112784477A (en) | WRF-LES and BP-PSO-Bagging combined wind power prediction method | |
| Wang et al. | LLNet: Lightweight network with a channel and spatial attention mechanism for local climate zone classification from Sentinel-2 image | |
| Pasero et al. | Artificial neural networks for meteorological nowcast | |
| Gan et al. | W-MRI: A Multi-output Residual Integration Model for Global Weather Forecasting | |
| Wang et al. | Tiny-RainNet: A Deep CNN-BiLSTM Model for Short-Term Rainfall Prediction | |
| CN114355482B (en) | Strong convection weather identification method and system based on optical flow and meteorological numerical prediction | |
| HADJI | A coupled models Hydrodynamics-Multi headed Deep convolutional neural network for rapid forecasting large-scale flood inundation | |
| Sigg et al. | Photographic Visualization of Weather Forecasts with Generative Adversarial Networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |