CN117741821A - Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data - Google Patents
Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data Download PDFInfo
- Publication number
- CN117741821A CN117741821A CN202311736347.8A CN202311736347A CN117741821A CN 117741821 A CN117741821 A CN 117741821A CN 202311736347 A CN202311736347 A CN 202311736347A CN 117741821 A CN117741821 A CN 117741821A
- Authority
- CN
- China
- Prior art keywords
- model
- time
- arpredrnn
- layer
- short
- 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
- 238000001556 precipitation Methods 0.000 title claims abstract description 50
- 238000013277 forecasting method Methods 0.000 title claims abstract description 9
- 239000010410 layer Substances 0.000 claims abstract description 72
- 238000000034 method Methods 0.000 claims abstract description 29
- 239000011229 interlayer Substances 0.000 claims abstract description 8
- 230000007246 mechanism Effects 0.000 claims description 10
- 230000007774 longterm Effects 0.000 claims description 9
- 230000015654 memory Effects 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 7
- 230000002123 temporal effect Effects 0.000 claims description 6
- 230000006886 spatial memory Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012549 training Methods 0.000 abstract description 19
- 238000002592 echocardiography Methods 0.000 abstract description 16
- 230000000694 effects Effects 0.000 abstract description 9
- 238000005516 engineering process Methods 0.000 abstract description 6
- 230000008901 benefit Effects 0.000 abstract description 3
- 238000012937 correction Methods 0.000 abstract description 3
- 239000013589 supplement Substances 0.000 abstract description 3
- 230000002238 attenuated effect Effects 0.000 abstract description 2
- 238000013213 extrapolation Methods 0.000 description 13
- 230000006870 function Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 6
- 238000013526 transfer learning Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 230000006872 improvement Effects 0.000 description 4
- 230000007423 decrease Effects 0.000 description 3
- 238000010276 construction Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000008520 organization Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000006424 Flood reaction Methods 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 230000008929 regeneration Effects 0.000 description 1
- 238000011069 regeneration method Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- 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
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a short-time strong precipitation minute-level forecasting method based on an SFGAN-ARPredRNN model and multi-layer radar data, relates to a weather forecasting technology, and provides a scheme for solving the problem that radar echoes are gradually attenuated in the prior art. Embedding a time attention module and an interlayer attention module into the PredRNN model, adding historical key information to supplement original input information of the multilayer radar, and adding a residual structure to obtain an ARPredRNN model; generating an countermeasure network by the ARPredRNN model based on sequence information and each frame of picture to obtain the SFGAN-ARPredRNN model; finally, the SFGAN-ARPredRNN model is utilized to quantitatively forecast the short-time strong precipitation in minute scale. The method has the advantages that the problem that the radar echo gradually attenuates along with time is promoted to be solved. And constructing a short-time strong precipitation radar echo prediction model and a differential training model according to the classification and the season. And the forecasting effect of short-time strong rainfall is improved. On-line training, and real-time correction of strong precipitation radar echo prediction.
Description
Technical Field
The invention relates to a weather forecast technology, in particular to a short-time strong precipitation minute-level forecast method based on an SFGAN-ARPredRNN model and multi-layer radar data.
Background
Short-time strong rainfall often causes major disasters such as urban waterlogging, mountain floods, debris flows, landslide and the like, and causes major property loss, casualties and the like.
The short-time strong precipitation is mostly caused by a medium-small scale system, has the characteristics of small scale, strong burst and short life span, and is always the key point and the difficulty of weather forecast. The current objective prediction of the short-time strong precipitation is mainly performed for 6-12 hours according to the numerical mode calculation environment physical quantity, and the requirement of quantitative prediction of the short-time strong precipitation with rapid change or rapid generation and elimination is difficult to meet. Weather radar becomes the main equipment for short-time strong precipitation monitoring and short-time early warning. The forecaster mainly carries out subjective early warning of short-time strong rainfall according to the echo characteristics of the weather radar reflectivity factors, the intensity, the position change and the like, and the objective forecasting method is rare.
The radar echo short-term prediction related to the method is gradually a new development direction along with the development of the deep learning technology, and the radar echo extrapolation prediction is performed by using the deep learning technology. However, most radar echo extrapolation models based on deep learning face the difficulty that radar echo gradually decays with time increase, namely echo intensity gradually weakens and the predicted hit proportion of a strong echo region drastically decreases. The extrapolation result of the traditional RNN type space-time sequence prediction model is gradually distorted and blurred, the definition is greatly reduced compared with a real radar echo image, and the requirements of clear and accurate prediction cannot be met.
Disclosure of Invention
The invention aims to provide a short-time strong precipitation minute-level forecasting method based on an SFGAN-ARPredRNN model and multi-layer radar data, so as to improve quantification, objectification and refinement level of short-time strong precipitation forecasting and forecasting timeliness and improve the problem of gradual attenuation of radar echo.
According to the short-time strong precipitation minute-level forecasting method based on the SFGAN-ARPredRNN model and the multi-layer radar data, a time attention module and an interlayer attention module are embedded into the PredRNN model, historical key information is added to supplement original input information of the multi-layer radar, and a residual error structure is added to obtain the ARPredRNN model; generating an countermeasure network by the ARPredRNN model based on sequence information and each frame of picture to obtain the SFGAN-ARPredRNN model;
finally, the SFGAN-ARPredRNN model is utilized to quantitatively forecast the short-time strong precipitation in minute scale.
The ARPredRNN model includes a TAM module to capture the attention of different time frames.
The TAM module includes processing long-term time correlation information.
The ARPredRNN model includes LAM modules for capturing the attention between the different layers.
The LAM module includes processing spatial correlation information for different levels.
The PredRNN model is provided with a TAM module and a LAM module to obtain an AttST-LSTM model, and the model expression is as follows:
wherein X is the tensor of the time, T is the time, g is the processing gate, i is the updating gate, f is the forgetting gate, W is the weight, l is any layer, H is the hidden state, M is the spatial memory characteristic tensor, C is the temporal memory characteristic tensor, o is the output gate, Q is the query tensor, K is the key tensor, V is the value tensor, T is the name of TAM, and S is the name of LAM.
After the interlayer attention module is added, the output of any layerThe method comprises the following steps:
output per layerGenerating a current time query Q through convolution and deformation T ，K T Is history input information->Obtained by respective convolution, V T Is to input information +.>Is obtained through deformation; when l is 1, ">For all entered history mapsImage information X 0：t The method comprises the steps of carrying out a first treatment on the surface of the When l is not 1->Conceal the status for all outputs of the previous layer +.>To obtain long-term spatio-temporal information.
Taking into account hidden state tensor information generated from the previous L-1 layer in LAMAcquire attention and add to the top level output +.>The final output is obtained in a self-adaptive manner to obtain the space information in different layers, wherein t is the current time, S is the name of the LAM module, and l is any layer; within the same time frame, top layer output +.>Generating query Q by convolution and morphing S And K is S And V S Hidden state tensor information generated through the front L-1 layer +.>Generating K through respective convolution and deformation S And V S New output->May be updated by the attention mechanism; after generating a new output, combining the original output with the new output and performing layer-by-layer normalization to obtain +.>
The short-time strong precipitation minute-level forecasting method based on the SFGAN-ARPredRNN model and the multi-layer radar data has the following advantages:
1) The SFGAN-ARPredRNN model is innovatively provided, so that the problem that radar echo gradually decays along with time is solved.
Aiming at the problem that radar echoes are gradually attenuated along with time, attention mechanisms are introduced first, a time attention module (Temporal Attention Module, TAM) and an interlayer attention module (Layer Attention Module, LAM) are embedded on the basis of a PredRNN model, and an ARPredRNN model is proposed. The model retains more features from both temporal and spatial dimensions, respectively, and the goal of the two modules is to capture the attention of different time frames and different layers and to take into account long-term history information. TAMs consider long-term time correlation information, while LAMs consider spatial correlation information for different levels.
Secondly, adding a GAN module behind the ARPredRNN model to obtain a final SFGAN-ARPredRNN model. The model improves the blurring problem caused by the MSE loss function by introducing a plurality of antagonism generation network improvement loss functions, and generates extrapolation results through a game of a generator and a multi-dimensional discriminator.
(2) And constructing a short-time strong precipitation radar echo prediction model and a differential training model according to the classification and the season.
According to the influence system and by combining with radar echo organization forms, short-time strong rainfall types in a designated area can be classified into typhoons, mason types (mason type I and mason type II) and common monsoon types, and three types of four-time short-time rainfall radar echoes have the characteristics and obvious differences in the aspects of space structure, moving direction, airflow guiding, rainfall mechanism and the like.
And (3) constructing short-time strong precipitation historical process data sets of the designated areas in a classified manner, respectively constructing three types of four-type short-time strong precipitation radar echo prediction models, and respectively carrying out parameter adjustment and optimization. Compared with the prior art that most radar echo extrapolation prediction technologies adopt the same model, the method disclosed by the invention improves the prediction effect of the radar echo model, and the actual verification and evaluation prove the same, and the method has the further advantage that the requirement of a deep learning model on data volume can be reduced.
(3) And the forecasting effect of short-time strong rainfall is further improved by adopting multi-layer radar data.
By adopting the multi-layer radar data, on one hand, the vertical structures of the strong precipitation radar echoes of different types can be effectively distinguished, the commonality characteristics of the strong precipitation echoes of the same type are refined, and the divergence is reduced. On the other hand, the change characteristics of the echo intensity and the vertical structure of the strong precipitation radar can be effectively monitored and refined, and the method is favorable for predicting the intensity change trend of the increase or decrease of the convection storm echo which generates strong precipitation, and further predicting the evolution trend of the beginning, increase, decrease or extinction of precipitation. Compared with the prior art that most radar echo extrapolation technologies adopt single-layer or combined reflectivity image data, the method acquires more information of the vertical structure and intensity change of strong precipitation, and can further improve the prediction accuracy of the method.
(4) On-line training, and real-time correction of strong precipitation radar echo prediction.
The construction of the strong rainfall echo model is pretrained by using historical data, so that a plurality of common rules are learned, but a plurality of properties are also processed averagely. The radar echo running in real time has the characteristics of model history commonality and individuation in the process. Therefore, the characteristics of the process are learned online, and the method is proved to be effective indeed according to the real-time correction model of the echo live condition, so that the method has certain improvement and promotion effects on solving key problems such as fast echo attenuation, echo regeneration and the like in deep learning.
Drawings
Fig. 1 is a schematic diagram of a structure of a PredRNN network in the prior art.
Fig. 2 is a schematic diagram of a prior art ST-LSTM circulation unit.
Fig. 3 is a schematic diagram of the structure of the AttST-LSTM cell according to the present invention.
Fig. 4 is a schematic view of the construction of the TAM module unit of the present invention.
Fig. 5 is a schematic view of the LAM module unit according to the present invention.
Fig. 6 is a schematic diagram of the global arbiter according to the present invention.
Fig. 7 is a schematic diagram of the structure of the local discriminator in the invention.
FIG. 8 is a schematic diagram of the structure of the SFGAN-ARPredRNN model according to the present invention.
Detailed Description
The SFGAN-ARPredRNN model in the invention comprises the following steps: sequence-Frame Generative Adversarial Networks, ARPredRNN model: predRNN with Attention module and Residual module.
Description of ARPredRNN model: the core of the PredRNN model is a space-time LSTM (ST-LSTM) unit modified on the basis of the ConvLSTM (Convolutional Long Short-Term Memory) model, as shown in fig. 1 and 2. The module can extract and memorize the space and time characteristic states at the same time, and vertically and horizontally transmit the memory states in the hierarchical structure, and the memory states extracted at the highest layer are transmitted to the bottom layer of the next frame. The memory state flows in the zigzag direction in the whole network, so that the space-time memory interacts with the original long-time memory and the prediction capability is improved.
Aiming at the echo extrapolation attenuation problem, based on the PredRNN model, from the viewpoint of an attention mechanism, historical key information is added to improve the prediction effect. In order to obtain the supplement of the original input information, the addition of the residual error structure brings about further improvement, and then an ARPredRNN model is provided.
The TAM (Temporal Attention Module) and LAM (Layer Attention Module) modules designed in the ARPredRNN model function differently from other attention modules or self-attention modules, the goal of both modules is to capture attention at different time frames and between different layers and to take into account long-term history information. TAMs consider long-term time correlation information, while LAMs consider spatial correlation information for different levels. In TAM, historical input information is associated as keys and values in the attention mechanism. Similarly, in LAM, the hidden state of each layer also serves as a corresponding key and value input model. Finally, the current output is used as a query to form the corresponding TAM and LAM outputs by means of an attention mechanism function.
The AttST-LSTM (ST-LSTM with Attention module) model obtained by modification of TAM and LAM modules after application to PredRNN is shown in fig. 3 to 5. Wherein X is the tensor of the time, T is the time, g is the processing gate, i is the updating gate, f is the forgetting gate, W is the weight, l is any layer, H is the hidden state, M is the spatial memory characteristic tensor, C is the temporal memory characteristic tensor, o is the output gate, Q is the query tensor, K is the key tensor, V is the value tensor, T is the name of TAM, and S is the name of LAM.
The time attention module is arranged, and from the viewpoint of frame-by-frame prediction, the PredRNN state update only considers the characteristic tensor mapping of the adjacent frames, and longer historical information is forgotten with the increase of time. In TAM, input of information through history is consideredObtain attention and add into each layer of output +.>Wherein T is the current time, and T is TAM moduleIn the name of (a), l is any layer. In layer 1, +.>For all the input historical image information X 0：t While at other layers all outputs of the previous layer are hiddenTo obtain long-term spatio-temporal information. Output per layer->Generating a current time query Q through convolution and deformation T And K is T And V T Input information +.>Generating K through respective convolution and deformation T And V T The new output of the layer->The attention mechanism can be updated as:
after the new output is generated, the original output is combined with the new output and normalized layer by layer. TAM only works when the current time t > k, i.e. the predicted image phase, the layer conceals the stateHistory input information->As input to the TAM.
The feature tensor is difficult to propagate to the top layer due to the layer-by-layer propagation from layer 1 to the top layer. Although the inter-layer propagation in PredRNN is improved, spatial memory characteristic tensors are proposedPreserving spatial information but it has been updatedThe process still accompanies the loss of spatial information, which results in that the decoding stage cannot acquire local detail information of the predicted image of the previous frame from the top layer, and thus, the predicted image is weakened. Thus, an interlayer attention module is provided, taking into account hidden state tensor information generated from the previous L-1 layer in the LAM +.>Acquire attention and add to the top level output +.>And (3) generating final output to adaptively acquire spatial information in different layers, wherein t is the current time, S is the name of the LAM module, and l is any layer. Within the same time frame, top level outputGenerating query Q by convolution and morphing S And K is S And V S Hidden state tensor information generated through the front L-1 layer +.>Generating K through respective convolution and deformation S And V S New output->May be updated by the attention mechanism. After generating a new output, combining the original output with the new output and performing layer-by-layer normalization to obtain +.>
Description of SFGAN-ARPredRNN model:
the model architecture is shown in fig. 6 to 8, and in order to simultaneously utilize global and local characteristics and simultaneously consider the best performance of a sequence angle and each frame angle, the invention comprehensively considers different dimensions and cross designs a sequence global discriminator SeqGlobalGAN, a sequence local discriminator seqpatch gan, a frame global discriminator FrmGlobalGAN and a frame local discriminator FrmPatchGAN. SFGAN is Sequence-Frame Generative Adversarial Networks, which means that the SFGAN-ARPredRNN model is finally obtained by the network considering the Sequence and the generation of each frame of picture.
The smaller the receptive field is, the larger the finally generated characteristic diagram is, the more detail information is contained, the true and false probabilities are better judged from each part of the image, and more high-frequency detail contents are complemented, so that the receptive field of the input image of the local discriminator is designed to be 7 multiplied by 7, the receptive field of the global discriminator corresponds to 2 layers of convolution layers, and the receptive field of the global discriminator covers the whole input image; the sequence discriminators all take the radar echo prediction sequence as the input of the discriminators, and the frame discriminators randomly extract a certain frame of each sequence as the input of the discriminators, so that the extrapolation effect of each frame is better improved.
Table 1 global arbiter network parameters
Table 2 local discriminant network parameters
The ARPredRNN model cascade structure is adopted as a generator for generating an countermeasure network, the generated sequence and the real echo sequence are input into a discriminator together for classification, and the discriminator and the generator are updated after the loss is calculated. The predicted sequence is made closer to the true sequence by successive iterations of the arbiter and generator.
Loss function improvement the SFGAN-ARPredRNN model loss function consists of two parts: the first part is a loss function L used by the model contained in the generator p Pixel comparison loss, such as MAE and MSE loss, is typically used, where p is the name of pixel, i.e., the pixel comparison loss function. The loss function used in the method is improved, and besides the integral loss consideration of the sequence, the pixel comparison loss of a certain random frame is added, so that the extrapolation effect of each frame is better improved.
While the second part is the generator loss L for the global picture/sequence input discriminator and the local picture/sequence input discriminator adv The receptive field of the global arbiter covers the whole input image, whereas the input image receptive field of the local arbiter is initially designed to be 7 x 7 in size. Wherein adv is a generic name of adversal, namely the overall loss function of the pixel comparison loss function generator is as follows.
L(G)＝L p +λ adv L adv (3-1)
Where i is a time of random extraction in the sequence, l is WMSE penalty (Weighted MSE), WMAE penalty (Weighted MAE), λ l Is a scale factor, D is a discriminator, G is a generator, and E is an average value. For the convenience of experiments, MSE loss scaling factor lambda is set MSE =1, mae loss scale factor λ MAE =0.1, the generator loses the scaling factor λ adv =0.001。
And establishing a short-time strong precipitation radar echo model for weather prediction by applying the SFGAN-ARPredRNN model in classification and season division. The embodiment of the invention takes Guangdong area as an example, the Guangdong short-time strong precipitation type can be roughly classified into typhoons (spiral echoes including low pressure of a monsoon), lines (strong convection type (called line I for short) under the influence of cold air and warm area linear convection type (called line II for short) under the influence of monsoon) and common monsoon (block echoes except for linear echoes and the like) according to the large-scale influence system and radar echo organization, and the spatial structure, the moving direction, the guiding airflow, the precipitation mechanism and the like of three types of four-time strong precipitation radar echoes have the characteristics and large differences and are obvious in seasonality.
And establishing a data set according to the classification by utilizing the Guangdong short-time strong precipitation history process, constructing three types of four short-time strong precipitation radar echo prediction models, and respectively carrying out parameter adjustment and optimization.
And (3) quantitatively forecasting the minute level of the short-time strong precipitation, obtaining precipitation quantity distribution based on a minute level dynamic relation of a variation method, outlining an area larger than 20 mm, and realizing the minute level quantitative forecasting of the short-time strong precipitation. And selecting a precipitation area larger than 20 mm to obtain a minute drop zone and an order forecast of short-time strong precipitation.
Three types of four-type short-time strong precipitation radar echo model division standards:
typhoons: typhoons, low-pressure monsoon and the like, radar echoes of the typhoons and the monsoon are distributed in a vortex or spiral mode and move in a rotary mode. This type of model begins when the typhoon center live location is located 350 km (settable) from the Guangdong province land boundary (i.e., coastline). The configuration file can modify the parameters and can also manually select the model.
Line class: including strong convection line echoes under the influence of cold air (type I, line I), and warm zone line convection echoes under the influence of a monsoon (type II, line II).
Common monsoon: bulk echoes other than linear echoes are scattered or accumulated. In practical application, typhoons are preferentially used for 5-10 months; 3-4 months, mainly form I; and (5) using common quaternary winds initially for 5-10 months, and switching to the line II type if the aspect ratio is judged to be larger than 4:1.
The SFGAN-ARPredRNN model online training method comprises the following steps: the online training belongs to isomorphic transfer learning, and is one kind of transfer learning. The transfer learning (Transfer Learning) is a capability of learning to be three-way in popular terms, and is characterized in that the existing knowledge is used for learning new knowledge, the core is to find the similarity between the existing knowledge and the new knowledge, and the purpose of transfer learning is achieved through the transfer of the similarity. Everything has commonality, how to reasonably find the similarity between everything, and further use the bridge to help learn new knowledge is a core problem of transfer learning. By online training, the characteristics and evolution rules of the current latest radar echo are obtained in real time, and the prediction model is finely adjusted in time so as to improve the prediction effect
Specifically, the on-line training strategy is added by taking the extrapolation prediction model of the three types of precipitation constructed as the reference model. In real-time operation, the judgment is carried out once at the beginning of 00 minutes or 30 minutes of each time, if the radar echo exists in the first 5 hours, an online training strategy is used, the reference model is subjected to fine adjustment through online training, then extrapolation forecast is carried out, and otherwise, the reference model is adopted for extrapolation forecast.
Training on data preprocessing and model: radar echo is a type of spatiotemporal data that has both time inertia and spatial interactions. In order to obtain the forecast data of the 3Km height layer, the invention uses the first 10 hours (1 hour) to extrapolate 20 hours (2 hours). Echoes of the 3Km height layer are extrapolated spatially with echoes of the 2,3,4Km height layer.
Of the data, 35794 parts of effective data are collected after the data cleaning, data arrangement and other common processes according to three types of precipitation types. And then preprocessing the data with the required heights of 2,3 and 4km, outputting the data into npy tensor files, facilitating the frequent reading of the data during training the model, and improving the training efficiency.
In order to achieve the best extrapolation effect of each type of precipitation, according to the opinion of service specialists, three models are classified respectively: the network structure of the model of the monsoon, the line model and the typhoon model is identical, and the training data are different.
The training data set and the test data set are divided according to the 8:2 proportion, and the specific data set number is shown in the following table.
TABLE 3 partitioning of training data sets and validation data sets
| Model type | Total data set | Training data set | Test data set |
| Quaternary wind model | 4320 | 3456 | 864 |
| Line model | 29075 | 23260 | 5815 |
| Typhoon model | 2399 | 1920 | 479 |
The SFGAN-ARPredRNN model can be trained on Ubuntu 18.04LTS operating system of Linux, the model of a server CPU is Intel (R) Xeon (R) Gold 6226R CPU@2.90GHz, yCharm is used as a code editor, codes are developed based on a deep learning framework PyTorch 1.10 in Python, and radar echo image visualization is achieved by using OpenCV. In order to improve the experiment parallel computing speed, the project uses a multi-GPU server environment for training, the used GPU is a GTX 2080Ti GPU manufactured by Inlet-Vicat (NVIDIA) company, and the single GPU video memory is 12G.
Table 4 training environment configuration table
It will be apparent to those skilled in the art from this disclosure that various other changes and modifications can be made which are within the scope of the invention as defined in the appended claims.
Claims (8)
1. The short-time strong precipitation minute-level forecasting method based on the SFGAN-ARPredRNN model and the multi-layer radar data is characterized in that a time attention module and an interlayer attention module are embedded into the PredRNN model, original input information of the multi-layer radar is supplemented by increasing historical key information, and a residual error structure is added to obtain the ARPredRNN model; generating an countermeasure network by the ARPredRNN model based on sequence information and each frame of picture to obtain the SFGAN-ARPredRNN model;
finally, the SFGAN-ARPredRNN model is utilized to quantitatively forecast the short-time strong precipitation in minute scale.
2. The method for short-time, strong precipitation minute level forecasting based on SFGAN-ARPredRNN model and multi-layer radar data of claim 1, wherein the ARPredRNN model includes a TAM module for capturing attention of different time frames.
3. The method of short-time, strong precipitation minute level forecasting based on SFGAN-ARPredRNN model and multi-layer radar data of claim 2, wherein the TAM module includes processing long-term time correlation information.
4. A method of short time, intense precipitation minute level forecasting based on the SFGAN-ARPredRNN model and multi-layer radar data of claim 3, wherein the ARPredRNN model includes LAM modules for capturing attention between different layers.
5. The method of short-time, strong precipitation minute-level forecasting based on the SFGAN-ARPredRNN model and multi-layer radar data of claim 4, wherein the LAM module includes processing spatial correlation information for different layers.
6. The short-time strong precipitation minute-level forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data according to claim 5, wherein the PredRNN model is provided with a TAM module and a LAM module to obtain an AttST-LSTM model, and the model expression is as follows:
wherein X is the tensor of the time, T is the time, g is the processing gate, i is the updating gate, f is the forgetting gate, W is the weight, l is any layer, H is the hidden state, M is the spatial memory characteristic tensor, C is the temporal memory characteristic tensor, o is the output gate, Q is the query tensor, K is the key tensor, V is the value tensor, T is the name of TAM, and S is the name of LAM.
7. The method for short-time, high-intensity precipitation minute-scale prediction based on SFGAN-ARPredRNN model and multi-layer radar data as claimed in claim 6, wherein after adding the interlayer attention module, the output of any layerThe method comprises the following steps:
output per layerGenerating a current time query Q through convolution and deformation T ，K T Is history input information->Obtained by respective convolution, V T Is to input information +.>Is obtained through deformation; when l is 1, ">For all the input historical image information X 0：t The method comprises the steps of carrying out a first treatment on the surface of the When l is not 1->Conceal the status for all outputs of the previous layer +.>To obtain long-term spatio-temporal information.
8. According to the weightsThe method for short-time, strong precipitation minute-level prediction based on SFGAN-ARPredRNN model and multi-layer radar data as claimed in claim 7, wherein hidden state tensor information generated from the previous L-1 layer is taken into consideration in the LAMAcquire attention and add to the top level output +.>The final output is obtained in a self-adaptive manner to obtain the space information in different layers, wherein t is the current time, S is the name of the LAM module, and l is any layer; within the same time frame, top layer output +.>Generating query Q by convolution and morphing S And K is S And V S Hidden state tensor information generated through the front L-1 layer +.>Generating K through respective convolution and deformation S And V S New output->May be updated by the attention mechanism; after generating a new output, combining the original output with the new output and performing layer-by-layer normalization to obtain +.>
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311736347.8A CN117741821A (en) | 2023-12-18 | 2023-12-18 | Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311736347.8A CN117741821A (en) | 2023-12-18 | 2023-12-18 | Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117741821A true CN117741821A (en) | 2024-03-22 |
Family
ID=90254038
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311736347.8A Pending CN117741821A (en) | 2023-12-18 | 2023-12-18 | Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117741821A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112200349A (en) * | 2020-09-16 | 2021-01-08 | 平衡机器科技(深圳)有限公司 | Remote sensing image heat island effect prediction method based on single window algorithm and PredRNN |
| CN115719036A (en) * | 2022-11-04 | 2023-02-28 | 中国地质大学（武汉） | Space-time process simulation method and system based on stacked space-time memory unit |
| CN116660849A (en) * | 2023-05-16 | 2023-08-29 | 西安理工大学 | Space-time LSTM network radar echo sequence prediction method based on self-attention mechanism |
| CN116755095A (en) * | 2023-03-09 | 2023-09-15 | 南京信息工程大学 | Short-term precipitation prediction method and device, terminal equipment and storage medium |
-
2023
- 2023-12-18 CN CN202311736347.8A patent/CN117741821A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112200349A (en) * | 2020-09-16 | 2021-01-08 | 平衡机器科技(深圳)有限公司 | Remote sensing image heat island effect prediction method based on single window algorithm and PredRNN |
| CN115719036A (en) * | 2022-11-04 | 2023-02-28 | 中国地质大学（武汉） | Space-time process simulation method and system based on stacked space-time memory unit |
| CN116755095A (en) * | 2023-03-09 | 2023-09-15 | 南京信息工程大学 | Short-term precipitation prediction method and device, terminal equipment and storage medium |
| CN116660849A (en) * | 2023-05-16 | 2023-08-29 | 西安理工大学 | Space-time LSTM network radar echo sequence prediction method based on self-attention mechanism |
Non-Patent Citations (1)
| Title |
|---|
| 马志峰 等: "基于深度学习的短临降水预报综述", 《计算机工程与科学》, vol. 45, no. 10, 31 October 2023 (2023-10-31), pages 1731 - 1747 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108416250B (en) | People counting method and device | |
| Min et al. | A binary classification method for bankruptcy prediction | |
| CN104408900B (en) | Neutral net flood warning devices and methods therefor based on dynamic optimization | |
| Elluri et al. | Developing machine learning based predictive models for smart policing | |
| CN112149887A (en) | PM2.5 concentration prediction method based on data space-time characteristics | |
| US20220405856A1 (en) | Property hazard score determination | |
| CN115482491A (en) | Bridge defect identification method and system based on transformer | |
| CN115392752B (en) | Subway short-time passenger flow prediction method, system, electronic equipment and storage medium | |
| CN112862177A (en) | Urban area concentration degree prediction method, equipment and medium based on deep neural network | |
| CN113628164A (en) | Pavement crack detection method based on deep learning and web end positioning | |
| Dai et al. | MSTCGAN: Multiscale time conditional generative adversarial network for long-term satellite image sequence prediction | |
| Wang et al. | Automatic detection of indoor occupancy based on improved YOLOv5 model | |
| Sun et al. | GAN-based LUCC prediction via the combination of prior city planning information and land-use probability | |
| CN113255620B (en) | Abnormal event detection method based on vehicle group track analysis in video | |
| Feng et al. | Spatiotemporal effects of urban ecological land transitions to thermal environment change in mega-urban agglomeration | |
| Zhou et al. | Exploration of the hidden influential factors on crime activities: A big data approach | |
| Zhou et al. | Automatic responsive-generation of 3D urban morphology coupled with local climate zones using generative adversarial network | |
| CN117741821A (en) | Short-time strong precipitation minute-scale forecasting method based on SFGAN-ARPredRNN model and multi-layer radar data | |
| CN116740495A (en) | Training method and defect detection method for defect detection model of road and bridge tunnel | |
| CN115631216B (en) | Multi-feature filter fusion-based holder target tracking system and method | |
| CN114781696B (en) | Model-free accident influence range prediction method for urban road network | |
| CN116050619A (en) | Multi-meteorological-element space-time sequence prediction and correction method by using Diffuse and 3DTCN | |
| CN112070075A (en) | Human body detection method based on collaborative regression | |
| Shen et al. | Lfnet: Lightweight fire smoke detection for uncertain surveillance environment | |
| CN116152699B (en) | Real-time moving target detection method for hydropower plant video monitoring system |
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 |