CN115840261A - Typhoon precipitation short-term prediction model optimization and prediction method - Google Patents
Typhoon precipitation short-term prediction model optimization and prediction method Download PDFInfo
- Publication number
- CN115840261A CN115840261A CN202211506522.XA CN202211506522A CN115840261A CN 115840261 A CN115840261 A CN 115840261A CN 202211506522 A CN202211506522 A CN 202211506522A CN 115840261 A CN115840261 A CN 115840261A
- Authority
- CN
- China
- Prior art keywords
- time
- space
- attention
- typhoon
- term prediction
- 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
Images
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
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a typhoon precipitation short-term prediction model optimization and prediction method, and belongs to the field of artificial intelligence application. The method designs a precipitation long-distance space-time modeling method based on Cross-batch multilayer semantic attention, and improves the understanding of the model to the large-scale high-order space-time characteristics of the complex weather system through long-distance space-time modeling. Aiming at the problem of huge resource consumption caused by a pixel-based space-time attention modeling mode in a large-scale scene, the method considers the influence of high-order space-time characteristics representing a space-time evolution mode of a large-scale weather system, realizes the improvement of the perception and modeling capacity of the model on long-distance space-time dependence, and comprehensively and effectively improves the precision of rainfall intensity prediction. Meanwhile, the model can effectively deal with precipitation short-term prediction of typhoon scenes of different types, and has certain generalization capability. The method has important practical application value for model optimization and rapid application of typhoon precipitation short-term prediction.
Description
Technical Field
The invention belongs to the field of artificial intelligence, and particularly relates to a typhoon precipitation short-term prediction model optimization and prediction method.
Background
The large-scale atmospheric circulation, water vapor transmission and other contemporary weather systems have obvious space-time correlation with typhoon precipitation, and high-order space-time characteristics representing macroscopic weather systems have important significance for constructing long-distance space-time dependence of typhoon precipitation. However, the spatial coding receptive field based on the convolutional neural network is relatively limited, and the time sequence coding based on the cyclic neural network also has the problem of long-term attenuation, and both the problems are difficult to effectively model the long-distance space-time dependency relationship. Some studies develop long-distance spatiotemporal modeling by combining a self-attention mechanism with position coding, but in a high-dimensional spatiotemporal domain, the modeling manner of the self-attention mechanism pixel-to-pixel brings huge calculation amount and increases the fitting difficulty. On the other hand, the self-attention only considers the spatiotemporal correlation among pixels in the calculation process, and ignores the correlation among the spatiotemporal features at higher levels. How to improve the perception and modeling capability of the model on long-distance space-time dependence at the cost of small computational power and further improve the prediction capability of the optimization model is a main target in this chapter.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a typhoon precipitation short-term prediction model optimization and prediction method.
In order to realize the purpose of the invention, the technical scheme is as follows:
in a first aspect, the invention provides a typhoon precipitation short-term prediction model optimization method, which is used for optimizing a typhoon precipitation short-term prediction model of an encoder-decoder structure, and comprises the following steps:
s1: acquiring respective original feature maps of an encoder and a decoder in a typhoon precipitation short-term prediction model to be optimized, carrying out region division on the original feature maps and acquiring the representation of each region feature in a high-dimensional feature space, so that each original feature map respectively forms corresponding space-time semantic features;
s2: performing multilayer semantic full-time-space information interaction based on the space-time semantic features obtained in the step S1, constructing multilayer semantic space-time features from the space-time semantic features to full-time-space visual fields, and providing multi-scale space-time semantic features with full-time receptive fields for subsequent calculation of space-time attention;
s3: based on multi-scale space-time semantic features with multi-level semantic full-time space vision, an attention mechanism is introduced to construct a space, time and channel three-head attention module, long-distance space-time attention calculation of precipitation is carried out, space-time attention with blocks as units is calculated, the final result is subjected to block decoding, fusion is carried out through a residual error structure and an original feature map, time sequence space-time feature expression of prediction is adaptively enhanced, and prediction accuracy of a typhoon precipitation short-term prediction model is improved.
Based on the technical scheme, the steps are preferably realized in the following specific mode. The preferred implementation manners of each step can be combined correspondingly without conflict, and are not limited.
In the first aspect, in the step S1, the typhoon precipitation short-term prediction model includes an encoder and a decoder; the encoder comprises three cascaded block modules, wherein each block module is a group of cascaded Convolation layers and ConvLSTM layers; the decoder also comprises three cascaded block modules, wherein each block module is a group of cascaded ConvLSTM operation and cascaded Deconvolition operation; in the step S1, the time-series feature maps of six timing feature maps of a first block module in the encoder and the six predicted timing feature maps of a last block module in the decoder are extracted to calculate the spatio-temporal semantic features, the time-series feature maps extracted in the encoder are timing feature codes output by a constraint layer and are used for constructing keys and values in the self-attention mechanism, and the predicted timing feature maps extracted in the decoder are predicted timing feature maps output by a constraint layer and are used for constructing queries in the self-attention mechanism.
As a preferable aspect of the first aspect, in step S1, the processing is performed in accordance with steps S11 to S13 on the feature maps of the encoder and the decoder in the original typhoon precipitation transient prediction model:
s11: performing region division on each feature map with the original size of c × h × w, performing regular division and clipping on the space region through a preset p × p block size (patch size), and obtaining an h/p × w/p block (patch) region;
s12: performing parallel coding on each block region based on c 'convolution kernels with the dimensionality of c multiplied by p, wherein c' > c, and performing parameter sharing on coding convolution kernels of all block regions in order to ensure that the extracted modes of each block region are kept consistent, obtaining h/p multiplied by w/p block region characteristic graphs with the dimensionality of c 'multipliedby 1 multiplied by 1, and recombining the h/p multiplied by w/p to obtain the c' multipliedby h/p multiplied by w/p dimensionality space-time semantic characteristics;
s14: and (3) converting the feature maps of the six input time sequences of the coder and the feature maps of the six prediction time sequences of the decoder according to S11 and S12 respectively to obtain the space-time semantic features with the dimension of c' × h/p × w/p.
As a preferred aspect of the first aspect, the specific method of step S2 is as follows:
s21: respectively connecting and fusing six space-time semantic features corresponding to the encoder and six space-time semantic features corresponding to the decoder in series to obtain a dimension(s) in X c', h/p, w/p) and(s) out Two space-time characteristics CPF of x c', h/p, w/p) in And CPF out An initial value of (1);
in the formula: concat represents a tandem fusion operation; s is in And s out Characteristic diagrams respectively representing input timings in the encoder and the decoder, both of which are 6;
s22: using multi-layer stacking Depth Separable Convolution (DSC), the initial space-time feature CPF obtained in S21 in And CPF out Are respectively in the airCross convolution of cross-patch is carried out in the inter-layer and feature depth direction, and the output feature of the ith-layer stacking depth separable convolution is represented as CPFi;
s23: obtaining n +1 characteristic graphs CPF after separable convolution of common n-layer stacking depth in S22 i ，i∈[0,n]Respectively representing space-time semantic features to space-time features of different semantic levels with global receptive fields, then performing weighted fusion on the space-time features of different semantic levels by adopting a pixel-by-pixel addition mode, and finally obtaining a multi-scale space-time semantic feature TCPF:
wherein λ is 0 ,λ 1 ...λ n Represents n +1 characteristic maps CPF i The weights of (a) can be optimized by model back propagation.
S24: reversely segmenting the multi-scale space-time semantic features according to the serial splicing mode adopted in S21 to obtain the separated multi-scale space-time semantic features TCPF at each moment with full space-time vision and each level of semantic information i ，i∈[1,12]Indicating the time, the first six times i e [1,6 ∈]Corresponding to the input sequence, the last six time instants i are in the same size as the 7,12]Corresponding to the output sequence.
As a preferable aspect of the first aspect, in S22, each layer of stacking depth separable convolution includes a two-layer network of Depthwise convolution and Pointwise convolution, and the calculation process of each layer of stacking depth separable convolution is represented as:
CPF i ＝DSC(CPF i-1 )＝PW(DS(CPF i-1 ))
wherein CPFi represents an output characteristic of the i-th layer stack depth separable convolution, PW convolution is a weighted combination operation in the depth direction of the characteristic map, and the number of convolution kernels thereof is equal to the number of output channels.
As a preferred aspect of the first aspect, the specific method of step S3 is as follows:
s31: first, convoluting Conv using two 1*1 k And Conv v Multi-scale space-time semantic characteristic TCP for six input moments respectivelyF i Performing convolution to obtain key K and value V characteristics of attention mechanism, and performing convolution operation Conv used at different time k Parameter sharing between, convolution operations Conv used at different times v Parameters are also shared between:
K＝Conv k (TCPF 1 ，TCPF 2 …TCPF 6 )
V＝Conv v (TCPF 1 ，TCPF 2 …TCPF 6 )
meanwhile, conv is operated by adopting 1*1 convolution operation q Multi-scale space-time semantic feature TCPF for six output sequences i Performing convolution to obtain the query Q characteristics of the attention mechanism, and performing convolution operation Conv used at different moments q Parameter sharing with each other:
Q＝Conv q (TCPF 7 ，TCPF 8 …TCPF 12 )
the dimensions of K, Q, V obtained above are the same and are all (T, C, H, W).
S32: in order to calculate attention in three different feature spaces of space, time and channel, K, Q, V is respectively subjected to three-dimensional conversion to obtain three groups of K, Q, V with the dimensions of (T × C, H × W), (C × H × W, T) and (T × H × W, C); wherein the first group K, Q, V is Ks, qs, vs with dimensions (T C, H W) for spatial attention calculation; a second group K, Q, V is Kt, qt, vt with dimension (C x H x W, T) for temporal attention calculation; a third set K, Q, V is Kc, qc, vc, with dimensions (T × H × W, C) for channel attention calculations;
through the calculation of space attention, time attention and channel attention, the attention vector Att is obtained s 、Att t And Att c ；
S33: the calculated spatial attention vector Att s Time attention vector Att t And channel attention vector Att c Uniformly restoring to (T, C, H, W) dimensionality through dimensionality transformation, performing weighted fusion by combining weight coefficients in a pixel-by-pixel addition mode to obtain a final space-time attention calculation result Att fusion ：
Wherein, att re Representing the Att vector after dimension transformation; the weight coefficients alpha, beta, gamma can be optimized by model back propagation;
final spatiotemporal attention calculation result Att fusion After the block (patch) decoding, the decoding result is fused into a feature map output by a decoder in a typhoon precipitation short-term prediction model through a residual error structure, so as to adaptively enhance the space-time feature expression of six prediction time sequences.
Preferably, in S32, the internal calculation logics of the space, time and channel attention modules are consistent, and the formula is as follows:
Attention(Q,K,V)＝softmax(Q T /K)V
in the formula: attention denotes the calculated Attention vector.
Preferably, the inputs of the typhoon precipitation short-term prediction model are precipitation grid data and typhoon feature data in the target area at 6 moments.
Preferably, in the first aspect, the typhoon is characterized by a center air pressure and a maximum wind speed of the typhoon.
In a second aspect, the invention provides a typhoon precipitation short-term prediction method, which is implemented by optimizing a typhoon precipitation short-term prediction model of an encoder-decoder structure by using the optimization method of any one of the schemes in the first aspect, and training the typhoon precipitation short-term prediction model to predict typhoon precipitation short-term.
Compared with the prior art, the invention has the following beneficial effects:
the invention provides a typhoon rainfall short-term prediction model optimization method, which combines a rainfall long-distance space-time modeling method based on Cross-batch multilayer semantic attention, adopts a Cross-batch multilayer semantic space-time coding strategy and multilayer semantic space-time attention calculation, and improves the understanding of the model to the large-scale high-order space-time characteristics of a complex weather system through long-distance space-time modeling. Aiming at the problem of huge resource consumption caused by a pixel-based space-time attention modeling mode in a large-scale scene, the method considers the influence of high-order space-time characteristics representing a space-time evolution mode of a large-scale weather system, realizes the improvement of the perception and modeling capacity of the model on long-distance space-time dependence, and comprehensively and effectively improves the precision of rainfall intensity prediction. Meanwhile, the method can effectively deal with the short-term prediction of rainfall in different types of typhoon scenes, and has certain generalization capability. Therefore, the method has important practical application value for model optimization and rapid application of typhoon precipitation short-term prediction.
Drawings
FIG. 1 is a flow chart of long-distance spatiotemporal modeling based on cross-patch multi-layer semantic attention.
FIG. 2 is a flowchart of the Patch Embedding process.
FIG. 3 is a Cross-Patch multi-layer semantic full-time space information interaction flow chart.
FIG. 4 is a flow chart of spatiotemporal attention calculation.
Detailed Description
The invention is further illustrated and described below with reference to the drawings and the detailed description.
In a preferred embodiment of the present invention, a typhoon precipitation short-term prediction model optimization method is provided for optimizing a typhoon precipitation short-term prediction model of a coder-decoder structure. The typhoon precipitation short-term prediction model optimization method mainly comprises 3 steps of S1-S3:
s1: the method comprises the steps of obtaining respective original feature maps of an encoder and a decoder in a typhoon precipitation short-term prediction model to be optimized, carrying out region division on the original feature maps, obtaining the representation of each region feature in a high-dimensional feature space, and enabling each original feature map to form corresponding space-time semantic features respectively. Compared with the original characteristic diagram, the characteristic diagram obtained by the Patch Embedding reduces the space dimension, enlarges the channel dimension, and obtains the high-order expression of the original characteristic diagram in a high-dimensional characteristic space by compressing redundant fine-grained precipitation information;
s2: and further performing multilayer semantic full-time-space information interaction based on the space-time semantic features PEF obtained by Patch Embedding in the S1, constructing multilayer semantic space-time features from space-time semantic features to full-time-space visual fields, and providing a multi-scale space-time semantic feature TCPF with full-time receptive fields for the subsequent calculation of space-time attention.
S3: based on a multi-scale spatiotemporal semantic feature TCPF with a multi-level semantic full spatiotemporal view, an attention mechanism is introduced to construct a space, time and channel three-head attention module, long-distance spatiotemporal attention calculation of precipitation is carried out, and spatiotemporal attention with a block (patch) as a unit is calculated. And finally, performing patch decoding on the result, fusing the residual error structure with the original characteristic diagram, and adaptively enhancing the expression of the six predicted time sequence space-time characteristics.
In the methods shown in S1 to S3, a Cross-Patch multi-layer semantic attention-based long-distance spatio-temporal modeling mechanism is designed for the problem of huge resource consumption caused by a global pixel point-based spatio-temporal attention modeling mode in a large-scale scene, and considering the influence of high-order spatio-temporal features representing a large-scale weather system spatio-temporal evolution mode, and the principle of the mechanism is shown in fig. 1. On one hand, the invention provides space-time attention calculation based on patch aiming at huge calculation amount based on global pixel point attention in a high-dimensional space-time scene. By performing patch division and encoding on the feature map, the aggregation neighborhood information obtains a high-order feature expression based on the patch. And while the redundant fine-grained information in the characteristic diagram is compressed, the resource consumption of subsequent space-time attention calculation is greatly reduced. On the other hand, in order to further fully extract and fuse different levels of space-time characteristics and provide multi-scale space-time characteristics with full space-time visual fields for calculation of space-time attention, the invention provides a Cross-Patch multi-layer semantic full space-time information interaction module, so that the subsequent space-time attention can adaptively select different levels of space-time characteristics to carry out long-distance modeling in the calculation process.
The following describes in detail specific implementations of S1 to S3 and effects thereof in this embodiment.
It should be noted that, theoretically, the short-term typhoon precipitation prediction model of the optimization object in the present invention may be any deep neural network model adopting an encoder-decoder structure, as long as the model can predict the short-term precipitation in the typhoon generation process.
In an embodiment of the present invention, the typhoon precipitation short-term prediction model includes an encoder (performing an Encoding process) and a decoder (performing a Forecasting process). As shown in fig. 1, the encoder includes three cascaded block modules, each block module is a set of cascaded Convolution layers and ConvLSTM layers; the decoder also contains three cascaded block modules, each block module being a set of cascaded ConvLSTM and Deconvolation operations. And finally outputting by a decoder through double-layer convolution to obtain a short rainfall prediction result.
In the above typhoon precipitation short-term prediction model, the input of the model needs to be time sequence data related to typhoon precipitation, and in this embodiment, precipitation grid data and typhoon feature data in the target region at 6 moments can be selected. The specific typhoon characteristics in the typhoon characteristic data can be optimized according to the actual condition, and the typhoon characteristics selected in the embodiment are the central air pressure and the maximum wind speed of the typhoon. The precipitation grid data and the typhoon feature data of the input model both need to be resampled to the same spatial reference and spatial resolution, the multisource input needs to be fused, and the simplest and intuitive fusion method is to splice three grid data (localization) together to be used as a three-channel feature input encoder. In addition, two coding branches can be respectively arranged in the encoder, one coding branch is used for inputting precipitation grid data, the other coding branch is used for inputting typhoon characteristic data (splicing center air pressure and maximum wind speed), the outputs of the two coding branches are fused through splicing operation, and the fusion result is used as the input of the decoder. In this embodiment, based on the typhoon precipitation short-term prediction model composed of the 6 block modules, the time-space semantic features may be calculated by extracting six timing feature maps of the first block module in the encoder and six predicted timing feature maps of the last block module in the decoder, and the timing feature maps extracted in the encoder are timing feature codes output by the convergence layer to construct a Key (K) and a Value (V) in the self-attention mechanism, and the predicted timing feature maps extracted in the decoder are predicted timing features output by the ConvLSTM layer to construct a Query (Q) in the self-attention mechanism.
In the embodiment of the present invention, in step S1, the processing is performed in steps S11 to S13 in a manner of Patch Embedding shown in fig. 2 for each feature map of the encoder and the decoder in the original typhoon precipitation short-term prediction model:
s11: performing region division on each feature map with the original size of c × h × w, regularly dividing and cutting a space region through a preset p × p block size (patch size), and obtaining an h/p × w/p block (patch) region;
s12: performing parallel coding on each block region based on c 'convolution kernels with the dimensionality of c multiplied by p, wherein c' > c, and performing parameter sharing on coding convolution kernels of all block regions in order to ensure that the extracted modes of each block region are kept consistent, obtaining h/p multiplied by w/p block region characteristic graphs with the dimensionality of c 'multipliedby 1 multiplied by 1, and recombining the h/p multiplied by w/p to obtain the c' multipliedby h/p multiplied by w/p dimensionality space-time semantic characteristics;
s14: compared with the original characteristic diagram, the characteristic diagram obtained by the Patch Embedding reduces the space dimension, enlarges the channel dimension, and obtains the high-order expression of the original characteristic diagram in the high-dimensional characteristic space by compressing redundant fine-grained precipitation information. Based on the process, six feature maps of six input time sequences of the encoder and six feature maps of six prediction time sequences of the decoder are converted according to S11 and S12 respectively to obtain the space-time semantic features with the dimension of c' × h/p × w/p.
Compared with the original characteristic diagram, the characteristic diagram obtained by Patch Embedding reduces the space dimension, enlarges the channel dimension, and obtains the high-order expression of the original characteristic diagram in a high-dimensional characteristic space by compressing redundant fine-grained precipitation information. By dividing the patch on the feature map and coding the region representation of each patch at each moment, the feature map compression is realized, and the resource consumption in the subsequent attention calculation process is greatly reduced. Meanwhile, the patch coding aggregates local information, blurs low-order detail features and obtains high-order feature expression based on the patch, so that space-time attention calculation is based on higher-level space-time feature expansion, and a model is facilitated to capture a higher-order global space-time evolution mode. In order to ensure that the coding modes of the input sequence and the output sequence are consistent, the input sequence and the output sequence keep parameter sharing in the process of Patch Embedding. Compared with the original characteristic diagram, the characteristic diagram obtained by Patch Embedding reduces the space dimension, enlarges the channel dimension, and obtains the high-order expression of the original characteristic diagram in a high-dimensional characteristic space by compressing redundant fine-grained precipitation information. Based on the process, the feature maps of the six input time sequences and the feature maps of the six predicted time sequences can be respectively converted into the feature PEF based on the patch with six dimensions of c' × h/p × w/p, and refined and efficient space-time semantic features are provided for the following full space-time information interaction and space-time attention calculation.
In the embodiment of the present invention, in the step S2, a Cross-Patch multi-layer semantic full-time-space information interaction process is designed, as shown in fig. 3, the specific method is as follows:
s21: respectively connecting and fusing six space-time semantic features corresponding to the encoder and six space-time semantic features corresponding to the decoder in series along the time dimension direction to obtain a dimension(s) in X c', h/p, w/p) and(s) out Two space-time characteristics CPF of x c', h/p, w/p) in And CPF out Is denoted by the subscript 0, i.e., an initial value of;
in the formula: concat represents a tandem fusion operation; s in And s out Characteristic diagrams respectively showing input timings in the encoder and the decoder, both of which are 6 in the present embodiment.
S22: using multi-layer stacking Depth Separable Convolution (DSC), the initial space-time feature CPF obtained in S21 in And CPF out Cross-convolution of cross-patch, i-th layer stack depth separable convolution DSC, in the spatial and feature depth directions respectivelyThe output characteristics are denoted CPFi.
Wherein each layer of stacking depth separable convolution comprises a Depthwise (DW) convolution and a Pointwise (PW) convolution two-layer network, and the DSC calculation process of each layer of stacking depth separable convolution is represented as:
CPF i ＝DSC(CPF i-1 )＝PW(DW(CPF i-1 ))
wherein CPFi represents an output characteristic of the i-th layer stacking depth separable convolution DSC, PW convolution is a weighted combination operation in the depth direction of the characteristic diagram, and the number of convolution kernels thereof is equal to the number of output channels.
S23: obtaining n +1 characteristic graphs CPF after separable convolution of common n-layer stacking depth in S22 i ，i∈[0,n]Respectively representing space-time semantic features (patch embedding) to different semantic levels with global receptive fields, and then performing weighted fusion on the space-time features of different semantic levels by adopting a pixel-by-pixel addition mode. Let the weight of n +1 features be λ 0 ,λ 1 ...λ n And finally obtaining the multi-scale space-time semantic feature TCPF as follows:
wherein λ is 0 ,λ 1 ...λ n Representing n +1 signatures CPF i The weights of (a) can be optimized by model back propagation.
S24: reversely segmenting (split) the multi-scale space-time semantic features according to the serial splicing mode adopted in S21 to obtain the separated multi-scale space-time semantic features TCPF at each moment with full space-time visual field and each level of semantic information i ，i∈[1,12]Indicating the time, the first six times i e [1,6 ∈]Corresponding to the input sequence, the last six time instants i epsilon [7,12 ]]Corresponding to the output sequence.
In the embodiment of the present invention, in the step S3, a space-time attention calculation method is introduced, as shown in fig. 4, the specific method is as follows:
s31: first, convoluting Conv using two 1*1 k And Conv v Respectively for six inputsMulti-scale space-time semantic feature TCPF of time i Convolution is carried out to obtain Key (K) and Value (V) characteristics of the attention mechanism, and parameters of the same convolution operation are shared among different moments, namely the convolution operations Conv used at different moments k Parameter sharing between, convolution operations Conv used at different times v The parameters are also shared. The Key (K) and Value (V) characteristics are calculated as follows:
K＝Conv k (TCPF 1 ，TCPF 2 …TCPF 6 )
V＝Conv v (TCPF 1 ，TCPF 2 …TCPF 6 )
meanwhile, conv is operated by adopting 1*1 convolution operation q Multi-scale space-time semantic feature TCPF for six output sequences i Performing convolution to obtain Query (Q) characteristics of attention mechanism, and performing convolution operation Conv used at different time q Share parameters with each other. The calculation formula of Query (Q) features is as follows:
Q＝Conv q (TCPF 7 TCPF 8 …TCPF 12 )
the dimensions of K, Q, V obtained above are the same and are all (T, C, H, W).
S32: in order to calculate attention in three different feature spaces of space, time and channel, K, Q, V is respectively subjected to three-dimensional conversion to obtain three groups of K, Q, V with the dimensions of (T × C, H × W), (C × H × W, T) and (T × H × W, C); the first group K, Q, V is Ks, qs, vs with dimensions (T C, H W) for spatial attention calculation; a second group K, Q, V is Kt, qt, vt with dimension (C x H x W, T) for temporal attention calculation; a third set K, Q, V is Kc, qc, vc with dimensions (T × H × W, C) for channel attention calculations.
The internal calculation logics of the space, time and channel attention modules are consistent, and the formula is as follows:
Attention(Q，K，V)＝softmax(Q T K)V
in the formula: attention denotes the calculated Attention vector.
Calculation of attention through space, time and channelThe Attention vector Attention is respectively obtained and recorded as Att s 、Att t And Att c 。
S33: the calculated spatial attention vector Att s Time attention vector Att t And channel attention vector Att c Uniformly restoring to (T, C, H, W) dimensionality through dimensionality transformation, performing weighted fusion by combining weight coefficients in a pixel-by-pixel addition mode to obtain a final space-time attention calculation result Att fusion ：
Wherein, att re Representing the Att vector after dimension transformation; the weight coefficients α, β, γ can be optimized by model back propagation.
Final spatiotemporal attention calculation result Att fusion After the block (patch) decoding, the decoding result is fused into a feature map output by a decoder in a typhoon precipitation short-term prediction model through a residual error structure, so as to adaptively enhance the space-time feature expression of six prediction time sequences. And inputting the fused features into a double-layer convolution positioned behind a decoder in the model, and outputting a corresponding typhoon precipitation short-term prediction result.
Therefore, the typhoon rainfall short-term prediction model with the encoder-decoder structure is optimized by the optimization method shown in S1-S3, and can be used for typhoon rainfall short-term prediction after model training.
The following is based on the optimization method shown in the above embodiments S1 to S3, and the effect thereof is shown by applying the optimization method to a specific example. The specific process is as described above, and is not described again, and the specific parameter setting and implementation effect are mainly shown below.
Examples
The following specifically describes the present invention with typhoon in the northwest pacific region and the precipitation of "summer waves" in the 2019 typhoon as specific examples, and the specific steps are as follows:
1) Typhoon data used was from the national weather center central weather station, including 85 typhoons from 2017 to 2019 in the northwest pacific region (fig. 1.2). The observed information of the typhoon data includes a cyclone position, a central air pressure, a moving wind speed, a maximum wind speed, and the like. The spatial resolution of the typhoon data is 0.1 deg., and the observation intervals are mainly 3h and 6h. As the typhoon intensity increases, the observation interval is correspondingly shortened, and the minimum time interval is 1h. And based on the data after the time-space matching, constructing a time-space prediction sequence sample by adopting a sliding window method, and further screening and dividing a typhoon precipitation data set on the basis.
2) Since the observation sequence was the past 3 hours (6 frames) of data and the prediction sequence was the future 3 hours (6 frames) of data, the sliding window length was set to 6 hours (12 frames). To obtain as many samples as possible while circumventing the data leakage problem, the sliding window moving step size is set to 2h (4 frames). And sampling the matched time sequence data based on a sliding window method, and finally obtaining a series of precipitation sequence samples with the sequence length of 6 hours.
3) And screening the precipitation sequence samples in the previous step, and deleting the precipitation sequence without the typhoon event. In addition, when typhoon is in the early formation or terminal death phase, the typhoon mark time in the observation sequence (the first 6 frames) in the precipitation sequence samples may be less than 6 frames, and in order to ensure that the samples have enough typhoon marks for the late typhoon information fusion and spatio-temporal inference map modeling process, the sequence samples with the typhoon mark time less than 3 in the observation sequence (the first 6 frames) are further deleted, and 4581 sequence samples are finally obtained.
4) And completely randomly dividing the sequence sample obtained in the second step into a training data set, a verification data set and a test data set according to the proportion of 10. The partitioned typhoon Precipitation Dataset is called TCPD (reliable cycle Precipitation Dataset), and relevant experiments are carried out based on the data set;
5) Model training and experimental validation were performed using the TCPD data set according to the procedure described previously.
The reconstruction method of the present invention is named ConvLSTM + MCPA, which is improved by 1.06% (m = 0.5), 1.83% (m = 2), 2.63% (m = 5), 3.17% (m = 10), 3.52% (m = 20), 3.77% (m = 30), respectively, compared to the original classical ConvLSTM method CSI, by 0.79% (m = 0.5), 1.73% (m = 2), 3.02% (m = 5), 4.14% (m = 10), 5.17% (m = 20), 5.87% (m = 30), respectively, by F1, by 0.99% (m = 0.5), 1.9% (m = 2), 3.12% (m = 5), 4.19% (m = 10), 5.19% (m = 20), 5.88% (m = 30), respectively. The precision lifting range shows the characteristic of increasing along with the increase of the precipitation strength. The multi-level semantic full-time-space information interaction module is added, so that the MCPA model of the multi-level semantic full-time-space information interaction module is further used for fully extracting and fusing different levels of time-space characteristics, multi-scale time-space characteristics with full-time-space visual fields are provided for the calculation of time-space attention, the model can be ensured to be capable of adaptively selecting the time-space characteristic information of reasonable semantic levels for long-distance modeling in the calculation process of the time-space attention, and finally the prediction capability of the model is improved. The experimental result also verifies the effectiveness and superiority of the long-distance modeling method provided by the invention.
The above-described embodiments are merely preferred embodiments of the present invention, which should not be construed as limiting the invention. Various changes and modifications may be made by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present invention. Therefore, the technical scheme obtained by adopting the mode of equivalent replacement or equivalent transformation is within the protection scope of the invention.
Claims (10)
1. A typhoon precipitation short-term prediction model optimization method is characterized by being used for optimizing a typhoon precipitation short-term prediction model of an encoder-decoder structure, and the optimization method comprises the following steps:
s1: acquiring respective original feature maps of an encoder and a decoder in a typhoon precipitation short-term prediction model to be optimized, carrying out region division on the original feature maps and acquiring the representation of each region feature in a high-dimensional feature space, so that each original feature map respectively forms corresponding space-time semantic features;
s2: performing multilayer semantic full-time-space information interaction based on the space-time semantic features obtained in the step S1, constructing multilayer semantic space-time features from the space-time semantic features to full-time-space visual fields, and providing multi-scale space-time semantic features with full-time receptive fields for subsequent calculation of space-time attention;
s3: based on multi-scale space-time semantic features with multi-level semantic full space-time visual fields, an attention mechanism is introduced to construct a space, time and channel three-head attention module, long-distance rainfall space-time attention calculation is carried out, space-time attention with blocks as units is calculated, the final result is subjected to block decoding, fusion is carried out through a residual error structure and an original feature map, time sequence space-time feature expression prediction is enhanced in a self-adaptive mode, and prediction accuracy of a typhoon rainfall short-term prediction model is improved.
2. The typhoon precipitation short-term prediction model optimization method according to claim 1, characterized in that: in the step S1, the typhoon precipitation short-term prediction model includes an encoder and a decoder; the encoder comprises three cascaded block modules, wherein each block module is a group of cascaded Convolition layers and ConvLSTM layers; the decoder also comprises three cascaded block modules, wherein each block module is a group of cascaded ConvLSTM operation and cascaded Deconvolition operation; in the step S1, the time-series feature maps of six timing feature maps of a first block module in the encoder and the six predicted timing feature maps of a last block module in the decoder are extracted to calculate the spatio-temporal semantic features, the time-series feature maps extracted in the encoder are timing feature codes output by a constraint layer and are used for constructing keys and values in the self-attention mechanism, and the predicted timing feature maps extracted in the decoder are predicted timing feature maps output by a constraint layer and are used for constructing queries in the self-attention mechanism.
3. The typhoon precipitation short-term prediction model optimization method as claimed in claim 1, characterized in that: in the step S1, processing is performed according to S11 to S13 for respective feature maps of an encoder and a decoder in the original typhoon precipitation short-term prediction model:
s11: performing region division on each feature map with the original size of c × h × w, regularly dividing and cutting a space region through a preset p × p block size (patch size), and obtaining an h/p × w/p block (patch) region;
s12: performing parallel coding on each block region based on c 'convolution kernels with the dimensionality of c multiplied by p, wherein c' > c, and performing parameter sharing on coding convolution kernels of all block regions in order to ensure that the extracted modes of each block region are kept consistent, obtaining h/p multiplied by w/p block region characteristic graphs with the dimensionality of c 'multipliedby 1 multiplied by 1, and recombining the h/p multiplied by w/p to obtain the c' multipliedby h/p multiplied by w/p dimensionality space-time semantic characteristics;
s14: and (3) converting the feature maps of six input time sequences of the encoder and the feature maps of six prediction time sequences of the decoder according to S11 and S12 respectively to obtain the space-time semantic features with the dimension of c' × h/p × w/p.
4. The typhoon precipitation short-term prediction model optimization method according to claim 1, characterized in that: the specific method of the step S2 is as follows:
s21: respectively connecting and fusing six space-time semantic features corresponding to the encoder and six space-time semantic features corresponding to the decoder in series to obtain a dimension(s) in X c', h/p, w/p) and(s) out Two space-time characteristics CPF of x c', h/p, w/p) in And CPF out An initial value of (1);
in the formula: concat represents a tandem fusion operation; s is in And s out Characteristic diagrams respectively representing input timings in the encoder and the decoder, both of which are 6;
s22: applying a multi-layer stacking Depth Separable Convolution (DSC) to the initial spatio-temporal features CPF obtained in S21 in And CPF out Performing cross convolution of cross-patch in space and feature depth directions respectively, wherein the output feature of the ith layer stacking depth separable convolution is represented as CPFi;
s23: the stacking depth of the common n layer in S22Separating convolution to obtain n +1 characteristic maps CPF i ，i∈[0,n]Respectively representing space-time semantic features to space-time features with different semantic levels of a global receptive field, and then performing weighted fusion on the space-time features with different semantic levels by adopting a pixel-by-pixel addition mode to finally obtain a multi-scale space-time semantic feature TCPF:
wherein λ is 0 ,λ 1 ...λ n Representing n +1 signatures CPF i The weights of (a) can be optimized by model back propagation.
S24: reversely segmenting the multi-scale space-time semantic features according to the serial splicing mode adopted in S21 to obtain the separated multi-scale space-time semantic features TCPF at each moment with full space-time vision and each level of semantic information i ，i∈[1,12]Represents the time, the first six times of i E [1,6 ]]Corresponding to the input sequence, the last six time instants i are in the same size as the 7,12]Corresponding to the output sequence.
5. The typhoon precipitation short-term prediction model optimization method as claimed in claim 1, characterized in that: in S22, each layer of stacking depth separable convolution includes a Depthwise convolution and a Pointwise convolution two-layer network, and the calculation process of each layer of stacking depth separable convolution is represented as:
CPF i ＝DSC(CPF i-1 )＝PW(DW(CPF i-1 )
wherein CPFi represents an output characteristic of the i-th layer stack depth separable convolution, PW convolution is a weighted combination operation in the depth direction of the characteristic map, and the number of convolution kernels thereof is equal to the number of output channels.
6. The typhoon precipitation short-term prediction model optimization method according to claim 1, characterized in that: the specific method of the step S3 is as follows:
s31: first, convoluting Conv using two 1*1 k And Conv v Multiple scales for six input moments respectivelySpatio-temporal semantic features TCPF i Performing convolution to obtain key K and value V characteristics of attention mechanism, and performing convolution operation Conv used at different time k Parameter sharing between them, convolution operations Conv used at different times v Parameters are also shared between:
K＝Conv k (TCPF 1 ,TCPF 2 …TCPF 6 )
V＝Conv v (TCPF 1 ,TCPF 2 …TCPF 6 )
meanwhile, conv is operated by adopting 1*1 convolution operation q Multi-scale space-time semantic feature TCPF for six output sequences i Performing convolution to obtain the query Q characteristics of the attention mechanism, and performing convolution operation Conv used at different moments q Parameter sharing with each other:
Q＝Conv q (TCPF 7 ，TCPF 8 …TCPF 12 )
the dimensions of K, Q, V obtained above are the same and are all (T, C, H, W).
S32: in order to calculate attention in three different feature spaces of space, time and channel, K, Q, V is respectively subjected to three-dimensional conversion to obtain three groups of K, Q, V with the dimensions of (T × C, H × W), (C × H × W, T) and (T × H × W, C); wherein the first group K, Q, V is Ks, qs, vs with dimensions (T C, H W) for spatial attention calculation; a second group K, Q, V is Kt, qt, vt with dimension (C x H x W, T) for temporal attention calculation; a third group K, Q, V is Kc, qc, vc with dimensions (T × H × W, C) for channel attention calculation;
through the calculation of space attention, time attention and channel attention, the attention vector Att is obtained s 、Att t And Att c ；
S33: the calculated spatial attention vector Att s Time attention vector Att t And channel attention vector Att c The dimension is restored to the (T, C, H, W) dimension in a unified way through dimension transformation, and the weighted fusion is carried out by combining weight coefficients in a pixel-by-pixel addition mode to obtain the final space-time attention calculation result Att fusion :
Wherein, att re Representing the Att vector after dimension transformation; the weight coefficients α, β, γ can be optimized by model back propagation;
final spatiotemporal attention calculation result Att fusion After the block (patch) decoding, the decoding result is fused into a feature map output by a decoder in a typhoon precipitation short-term prediction model through a residual error structure, so as to adaptively enhance the space-time feature expression of six prediction time sequences.
7. The typhoon precipitation short-term prediction model optimization method as claimed in claim 1, characterized in that: in S32, the internal calculation logics of the space, time, and channel attention modules are consistent, and the formula is as follows:
Attention(Q，K，V)＝softmax(Q T K)V
in the formula: attention represents the calculated Attention vector.
8. The method for optimizing the typhoon precipitation short-term prediction model according to claim 2, wherein the typhoon precipitation short-term prediction model is inputted with precipitation grid data and typhoon feature data in the target area at 6 moments.
9. The method of optimizing a typhoon precipitation short-term prediction model according to claim 2, wherein the typhoon characteristics are the center air pressure and the maximum wind speed of the typhoon.
10. A typhoon precipitation short-term prediction method is characterized in that a typhoon precipitation short-term prediction model of an encoder-decoder structure is optimized by the optimization method of any one of claims 1-9 and is trained to be used for typhoon precipitation short-term prediction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211506522.XA CN115840261A (en) | 2022-11-28 | 2022-11-28 | Typhoon precipitation short-term prediction model optimization and prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211506522.XA CN115840261A (en) | 2022-11-28 | 2022-11-28 | Typhoon precipitation short-term prediction model optimization and prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115840261A true CN115840261A (en) | 2023-03-24 |
Family
ID=85576170
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211506522.XA Pending CN115840261A (en) | 2022-11-28 | 2022-11-28 | Typhoon precipitation short-term prediction model optimization and prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115840261A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116070676A (en) * | 2023-03-28 | 2023-05-05 | 南京气象科技创新研究院 | Expressway road surface temperature forecasting method based on attention mechanism and self-encoder |
-
2022
- 2022-11-28 CN CN202211506522.XA patent/CN115840261A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116070676A (en) * | 2023-03-28 | 2023-05-05 | 南京气象科技创新研究院 | Expressway road surface temperature forecasting method based on attention mechanism and self-encoder |
| CN116070676B (en) * | 2023-03-28 | 2023-06-30 | 南京气象科技创新研究院 | Expressway road surface temperature forecasting method based on attention mechanism and self-encoder |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109902863B (en) | Wind speed prediction method and device based on multi-factor time-space correlation | |
| CN112347859B (en) | Method for detecting significance target of optical remote sensing image | |
| Yu et al. | Scene learning: Deep convolutional networks for wind power prediction by embedding turbines into grid space | |
| CN109657839B (en) | Wind power prediction method based on deep convolutional neural network | |
| CN109830102A (en) | A kind of short-term traffic flow forecast method towards complicated urban traffic network | |
| CN110570035B (en) | People flow prediction system for simultaneously modeling space-time dependency and daily flow dependency | |
| CN115840261A (en) | Typhoon precipitation short-term prediction model optimization and prediction method | |
| CN114495500A (en) | Traffic prediction method based on dual dynamic space-time diagram convolution | |
| Jeong et al. | Correcting rainfall forecasts of a numerical weather prediction model using generative adversarial networks | |
| CN112598590B (en) | Optical remote sensing time series image reconstruction method and system based on deep learning | |
| CN102256137B (en) | Context-prediction-based polar light image lossless coding method | |
| Tian et al. | Semantic segmentation of remote sensing image based on GAN and FCN network model | |
| Niu et al. | De-Trend First, Attend Next: A Mid-Term PV forecasting system with attention mechanism and encoder–decoder structure | |
| CN111275751B (en) | Unsupervised absolute scale calculation method and system | |
| CN116863379A (en) | Video prediction defense method based on space-time self-attention single-step disturbance | |
| Fang et al. | Stunner: Radar echo extrapolation model based on spatio-temporal fusion neural network | |
| CN115796359A (en) | PM2.5 space-time prediction method based on depth Koopman operator | |
| CN116152263A (en) | CM-MLP network-based medical image segmentation method | |
| Zhao et al. | Label Freedom: Stable Diffusion for Remote Sensing Image Semantic Segmentation Data Generation | |
| CN111666849A (en) | Multi-source remote sensing image water body detection method based on multi-view depth network iterative evolution | |
| Wei et al. | Compression and Storage Algorithm of Key Information of Communication Data Based on Backpropagation Neural Network. | |
| CN114664090B (en) | Traffic data filling method and system based on cyclic neural network | |
| CN116596779B (en) | Transform-based Raw video denoising method | |
| CN116129646B (en) | Traffic prediction method of graph convolution neural network based on feature intersection | |
| CN115984757B (en) | Abnormal event detection method based on global local double-current feature mutual learning |
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 |