CN116307283A - Precipitation prediction system and method based on MIM model and space-time interaction memory - Google Patents
Precipitation prediction system and method based on MIM model and space-time interaction memory Download PDFInfo
- Publication number
- CN116307283A CN116307283A CN202310565194.9A CN202310565194A CN116307283A CN 116307283 A CN116307283 A CN 116307283A CN 202310565194 A CN202310565194 A CN 202310565194A CN 116307283 A CN116307283 A CN 116307283A
- Authority
- CN
- China
- Prior art keywords
- space
- time
- mim
- information
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000015654 memory Effects 0.000 title claims abstract description 99
- 238000001556 precipitation Methods 0.000 title claims abstract description 39
- 230000003993 interaction Effects 0.000 title claims abstract description 32
- 238000000034 method Methods 0.000 title claims abstract description 15
- 238000010586 diagram Methods 0.000 claims abstract description 31
- 239000000284 extract Substances 0.000 claims abstract description 12
- 230000007774 longterm Effects 0.000 claims description 41
- 230000007787 long-term memory Effects 0.000 claims description 13
- 238000013507 mapping Methods 0.000 claims description 13
- 230000001965 increasing effect Effects 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 4
- 239000004973 liquid crystal related substance Substances 0.000 description 12
- 230000033001 locomotion Effects 0.000 description 5
- 238000000605 extraction Methods 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000002592 echocardiography Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/10—Devices for predicting weather conditions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Environmental & Geological Engineering (AREA)
- General Physics & Mathematics (AREA)
- Atmospheric Sciences (AREA)
- Artificial Intelligence (AREA)
- Quality & Reliability (AREA)
- General Business, Economics & Management (AREA)
- Operations Research (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Biodiversity & Conservation Biology (AREA)
- Ecology (AREA)
- Environmental Sciences (AREA)
- Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention belongs to the technical field of weather, and relates to a precipitation prediction system and method based on an MIM model and space-time interaction memory. A precipitation prediction system based on MIM model and space-time interaction memory comprises an ST-LSTM model and at least one ISTC-SA-MIM model which are connected in sequence; the ST-LSTM model extracts the space-time information in the radar echo diagram at the previous moment, generates a hidden state and a space-time memory state, and transmits the hidden state and the space-time memory state to the ISTC-SA-MIM model; and generating a new hidden state by the ISTC-SA-MIM model through the input space-time memory state and the hidden state, and predicting a radar echo diagram at the current moment. The ISTC-SA-MIM model provided by the invention can correspondingly predict short-term precipitation through a radar echo diagram, and has the characteristics of low error, high accuracy and low false alarm rate.
Description
Technical Field
The invention belongs to the technical field of weather, and relates to a precipitation prediction system and method based on an MIM model and space-time interaction memory.
Background
The prediction of precipitation within an hour is often referred to as a short-term precipitation prediction. Short-term precipitation predictions are closely related to people's life and travel. The accurate short-time rainfall prediction can help related departments to take corresponding measures in time against different conditions, and avoid harm caused by emergency. The radar echo diagram intuitively shows the intensity and the spatial distribution of precipitation cloud. Traditional short-term precipitation prediction predicts the change trend of an echo region in a radar echo diagram through an optical flow method. However, the radar returns may not be traveling at a fixed rate in a series of movements such as dispersion, concentration, and movement. Therefore, there is a limitation in predicting precipitation by an optical flow method.
The space-time model ConvLSTM can simultaneously capture time information and space information in the radar echo diagram. For radar echo maps, convLSTM is predicted with higher accuracy than conventional optical flow methods. The change information of some motions in the radar echo diagram is stationarity. The change rule of the stationarity information is unchanged, and the stationarity information cannot change along with the change of time. The stationarity information is therefore easy to memorize and predict. Some movements in the radar echo map, such as sudden overlapping and dissipation of precipitation clouds, are non-stationarity information. The law of motion in non-stationary information may change over time, which is difficult to predict and memorize for a space-time model. However ConvLSTM cannot memorize non-stationarity information in the radar echo. The stationarity information is designed to solve the above problem, and a space-time MIM (Memory In Memory) model is designed. The MIM model is based on memorizing non-stationarity information in the radar echo map. However, there is an internal correlation in the echo region in the radar echo map, and there is a correlation between the trend of the change in intensity at each point and the change at other points. Such internal correlation information is spatiotemporal context information. The higher the intensity of the echo region, the more rapidly the echo region changes. In order to accurately predict echo regions of different intensities, the model needs to memorize the spatio-temporal context information of the different regions. The MIM model cannot fully learn short-term and long-term space-time context information contained in the radar echo diagram through the space-time state of the MIM model, so that most important space-time context information cannot be learned by the MIM model to be lost, and the model cannot fully master the change trend of echoes with different intensities. Areas of high echo intensity change rapidly and for prediction of such areas, the model needs to rely largely on the spatio-temporal context information that the area exists to make a more accurate prediction of such areas. Therefore, the greater the intensity of the echo region, the greater the variation trend error of the echo generated by the MIM model, and the lower the accuracy of short-term precipitation prediction.
Disclosure of Invention
The invention aims to solve the problem of low accuracy of an MIM model in short-term precipitation prediction in the prior art, and utilizes an improved space-time MIM model to construct a precipitation prediction system and a precipitation prediction method based on the MIM model and space-time interaction memory.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: a precipitation prediction system based on MIM model and space-time interaction memory comprises an ST-LSTM model and at least one ISTC-SA-MIM model which are connected in sequence; the ST-LSTM model extracts space-time information in the radar echo diagram at the previous moment, generates a hidden state and a space-time memory state, and transmits the hidden state and the space-time memory state to the ISTC-SA-MIM model; the ISTC-SA-MIM model consists of an ISTC-SA module and an MIM module; the ISTC-SA-MIM model learns long-term and short-term space-time context information through input space-time memory states and hidden states; and learning the nonstationary information through the long-term memory state to generate a new hidden state containing the short-term space-time context information and the nonstationary information, and predicting a radar echo diagram at the current moment.
Preferably, the ISTC-SA module performs iterative interaction on the hidden state and the space-time memory state through space-time interaction operation, so as to increase the information quantity of the two states on space-time context information.
Preferably, the ISTC-SA module extracts important short-term spatiotemporal context information in the spatiotemporal memory state by convolution operation, maps the important information into a numerical form suitable for hiding state memory by nonlinear mapping; the hidden state fuses the values through the Hadamard product and nonlinear mapping, and the information quantity of the hidden state relative to short-term space-time context information is increased.
Preferably, the ISTC-SA module extracts the spatiotemporal context information in the updated hidden state through convolution operation, and increases the information amount of the spatiotemporal memory state through nonlinear mapping.
Preferably, the ISTC-SA module updates the spatiotemporal memory state and the importance degree of different spatiotemporal context information in the hidden state through a self-attention mechanism, enhances the expression capability of the hidden state for the long-term context information, and increases the information amount of the long-term context information in the spatiotemporal memory state.
Preferably, the ISTC-SA module generates a series of gates through a gating mechanism to help the space-time memory state to selectively forget redundant repeated space-time context information contained in the space-time memory state, memorize new long-term space-time context information and generate new space-time memory state.
Preferably, the MIM module comprises MIM-N and MIM-S; the MIM-N is used for memorizing non-stationarity information; the MIM-S is used for memorizing the stationarity information and integrating the non-stationarity information and the stationarity information to generate fusion information.
Preferably, the radar echo map is a real radar echo map or a radar echo map generated by system prediction.
The invention further provides a precipitation prediction method based on the MIM model and the space-time interaction memory, which comprises the following steps:
inputting a real radar echo diagram or a radar echo diagram generated by prediction at the previous moment into the precipitation prediction system;
and outputting a radar echo diagram at the current moment through prediction.
The invention also provides a computer storage medium which comprises the precipitation prediction system based on the MIM model and the space-time interaction memory, and the precipitation prediction is realized when the system operates.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention provides a new precipitation prediction model ISTC-SA-MIM, which can memorize non-stationarity information existing in a radar echo diagram, fully learn space-time context information contained in the radar echo diagram and increase the prediction capability of the model on the change trend of echo areas with different intensities;
2. the ISTC-SA-MIM model provided by the invention can correspondingly predict short-term precipitation through a radar echo diagram, and the ISTC-SA-MIM prediction has the characteristics of low error, high accuracy and low false alarm rate;
3. the ISTC-SA module provided by the invention can capture and increase the information quantity of the short-term and long-term space-time context information in different space-time states through space-time interaction operation and a self-attention mechanism, selectively memorize the short-term and long-term space-time context information through a series of gates and enhance the learning and memorizing capability of a model for the space-time context information.
Drawings
FIG. 1 is a schematic diagram of a precipitation prediction system based on MIM model and space-time interaction memory in an embodiment of the invention;
FIG. 2 is a block diagram of an ISTC-SA-MIM model in an embodiment of the invention;
FIG. 3 is an interactive flow chart of hidden states and space-time memory states of an ISTC-SA module according to embodiments of the invention;
FIG. 4 is a flow chart showing the interaction of the self-attention score and the space-time memory state of the ISTC-SA module according to the embodiment of the invention.
Detailed Description
In order to facilitate an understanding of the present study, the present study will be described in more detail below with reference to the drawings and specific examples. This study may, however, be embodied in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
The invention provides a precipitation prediction system based on MIM model and space-time interactive memory, which can not only memorize non-stationarity information in radar echo, but also fully learn space-time context information in radar echo. When the system predicts rainfall through a radar echo diagram, an ISTC-SA-MIM model needs to be built. The system consists of an ST-LSTM model and three ISTC-SA-MIM models, the architecture is shown in figure 1, the ST-LSTM model and the three ISTC-SA-MIM models are connected in sequence to form a four-layer network structure, and the first layer, the second layer, the third layer and the fourth layer are arranged from bottom to top in sequence. The operation flow and steps of the system are described in detail below.
The following steps are involved: hidden state, use
A representation; space-time memory state, use->
A representation; long-term memory state, use->
Indicating the superscript +/for each state>
Representing the number of network layers, < >/generated by the state>
The method comprises the steps of carrying out a first treatment on the surface of the Subscript->
Representing the moment of this state generation, for example: />
The superscript 1 of (1) denotes the first layer and the subscript denotes the +.>
Time; />
Subscript representation of (2)
The moment, namely the moment before the current moment; />
The superscript 4 of (2) indicates the fourth layer.
In the pair of
When the radar echo diagram at the moment is predicted, the input of the ST-LSTM model can be the real radar echo diagram at the last moment, and can also be the radar echo diagram predicted and generated by the system at the last moment.
The inputs to the ST-LSTM model include the spatiotemporal memory states generated by the ISTC-SA-MIM model of the fourth layer at the previous time
Input real radar echo picture +.>
Hidden state generated at one time on the layer +.>
And long-term memory state
. The ST-LSTM model first generates a forgetting gate by inputting information and spatiotemporal memory states>
Input door->
And candidate state->
Helping model learn important spatiotemporal information contained in echo pictures to generate new spatiotemporal memory state +.>
：
Wherein, the liquid crystal display device comprises a liquid crystal display device,
、/>
、/>
three convolution kernels for extracting features of an input radar echo picture;
、/>
、/>
is to add to the time-space memory state>
Three convolution kernels for feature extraction; />
、/>
、/>
Is to update the bias of the three state gates; as indicated by the symbol, hadamard product was indicated.
Meanwhile, the ST-LSTM model generates an input gate required for updating the long-term memory state through input information and a long-term change rule in a hidden state memory radar echo diagram generated at the last moment
Amnesia door->
And candidate State->
Updating the long-term memory state:
wherein, the liquid crystal display device comprises a liquid crystal display device,
is a long-term memory state generated by the last time pattern through forgetting the door +.>
The unimportant long-term information is forgotten selectively. />
、/>
、/>
Three convolution kernels for extracting features of an input radar echo picture; />
、/>
、/>
Three convolution kernels for extracting features of the hidden state; />
、/>
、/>
Is to update bias of three state gatesAnd (5) placing.
The ST-LSTM model memorizes information contained in the long-term memory state and the space-time memory state through an output gate, generating a hidden state for a change rule of a region containing a large amount of space-time information and echoes:
wherein, the liquid crystal display device comprises a liquid crystal display device,
the output gate is generated by the current time model, and the output gate control model is used for helping the model to generate prediction information by the information quantity of the memory information; />
Is a convolution kernel for extracting characteristics of input information, hidden state, long-term memory state and space-time memory state, respectively,/-for the input information and the hidden state>
Is the bias required to produce the output gate. />
Is a convolution kernel size +.>
Is responsible for extracting non-stationarity information and long-term and short-term space-time context information in long-term and space-time memory states.
At this time, the ST-LSTM model completes the learning process for the spatio-temporal information in the radar echo map.
To help the ISTC-SA-MIM model to memorize space-time context information and non-stationarity information, the ST-LSTM model passes through hidden states
And space-time memory state->
The spatiotemporal information is passed to the ISTC-SA-MIM model of the second layer.
As shown in FIG. 2, which is a structural diagram of an ISTC-SA-MIM model, the ISTC-SA-MIM model is composed of an ISTC-SA module and MIM-N, MIM-S, the MIM-N is used for memorizing non-stationarity information, and the MIM-S is used for memorizing stationarity information and integrating the non-stationarity information and stationarity information. The ISTC-SA-MIM models of the second layer, the third layer and the fourth layer are identical, and the algorithm principle of the ISTC-SA-MIM model is elaborated by taking the ISTC-SA-MIM model of the second layer as an example.
For the ISTC-SA-MIM model of the second layer, the input information is respectively hidden state
Space-time memory state->
Long-term memory state->
Hidden state->
. Wherein->
And->
ST-LSTM model from the upper layer.
And->
The two states are the two states generated by the layer of ISTC-SA-MIM model at the previous time.
The ISTC-SA-MIM model first passes through the hidden state
And->
Generating input gates and candidate states：
Meanwhile MIM-N extracts and memorizes the memorable part in the non-stationarity information through differential operation to generate the memory information at the current moment
. MIM-S integrates non-stationarity information and stationarity information to generate fusion information +.>
. Since MIM-N and MIM-S are ConvLSTM-like memory cells, therefore +.>
Long-term change characteristic of non-stationarity information representing MIM-N memory at last moment, ++>
Representing the long-term law of the non-stationarity information and the stationarity information memorized by MIM-S at the last moment, and the superscript also represents the network layer number:
the ISTC-SA-MIM model helps the long-term memory state to memorize and learn the non-stationarity information and the space-time information in the radar echo diagram through inputting the gate, the candidate state and the fusion information, finishes the learning of the non-stationarity information at the current moment, and generates a new long-term memory state
：
Because of the importance of non-stationary information, the ISTC-SA-MIM model requires two memories for non-stationary information. When non-stationarity information is memorized for the first time, the ISTC-SA-MIM model generates an output gate through input information:
wherein, the liquid crystal display device comprises a liquid crystal display device,
is an output gate for generating non-stationarity information of the ISTC-SA-MIM model memory of the current layer;
is the convolution kernel adopted by the ISTC-SA-MIM for extracting the characteristics of the input information, the hidden state, the long-term memory state and the space-time memory state, +.>
Is the bias required to produce the output gate.
The ISTC-SA-MIM model generates hidden states containing non-stationarity information and short-term spatiotemporal context information by outputting gate-selective memory of important non-stationarity information related to predictions
：
At this time, the model only memorizes non-stationarity information in the radar echo map, and does not memorize space-time context information yet. To learn long-short term spatiotemporal context information, the ISTC-SA-MIM model is developed by combining
And->
The input ISTC-SA module extracts important short-term and long-term spatio-temporal context information.
Due to the hidden state
Is ISTC-SA-MIM model learningThe main source of spatio-temporal context information, therefore the ISTC-SA module is first added +.>
The space-time context information amount specifically includes: the ISTC-SA module extracts important space-time context information in the space-time memory state through convolution operation, and maps the important space-time context information into a numerical form suitable for hiding state memory through nonlinear mapping; the hidden state increases the information amount of the space-time context information in the hidden state by memorizing important information generated through nonlinear mapping:
wherein, the liquid crystal display device comprises a liquid crystal display device,
is to add to the time-space memory state>
Performing convolution kernel adopted by feature extraction; />
The first parameter of the superscript represents the number of layers (described above) that the hidden state is generated in, and the second parameter represents the number of times the hidden state has interacted with in the ISTC-SA module. />
Representing the new hidden state that is generated after a time-space interaction has been performed.
At this time, only the information quantity of the hidden state about the space-time context information is increased, and the ISTC-SA module generates the hidden state by the method
Performing convolution operation to extract important space-time context information, and helping space-time memory state to memorize the space-time context information through nonlinear mapping: />
；
Wherein, the liquid crystal display device comprises a liquid crystal display device,
is to be a->
And performing convolution kernel adopted by feature extraction. />
Representing the state of spatiotemporal memory +.>
And completing a new state generated after one time of space-time interaction operation.
The hidden state and the space-time memory state increase the information amount of the space-time context information at this time, however, the information amount learned by the two space-time interactive operations is insufficient to help the model to memorize the space-time context information. Thus the ISTC-SA module will
And->
Iterative three time-space interactions, generating +.>
And->
The specific flow is shown in fig. 3.
And through convolution operation, the ISTC-SA module simultaneously extracts time information and space information in the space-time state to obtain important space-time context information. However, the importance of the information extracted by the ISTC-SA module through the convolution operation is the same. The spatiotemporal context information of equal importance can help the model predict the trend of the echo region over a short period. These spatio-temporal context information extracted by convolution are thus short-term spatio-temporal context information. The ISTC-SA module can only increase the information amount of two states with respect to short-term spatiotemporal context information through spatiotemporal interaction. For long-term prediction, the importance of the spatio-temporal context information is not the same. The ISTC-SA module updates the importance degree of different space-time context information on long-term prediction through a self-attention mechanism to obtain long-term space-time context information which is helpful for helping long-term prediction of a model.
Because the ISTC-SA-MIM model updates the spatiotemporal memory state primarily through hidden states, the ISTC-SA module first helps the hidden states to memorize long-term spatiotemporal context information through a self-attention mechanism. The ISTC-SA module will first operate by convolution
Mapping to different feature spaces to obtain Query: />
、 Key：/>
And Value: />
Three different characteristic information. Wherein W, H, C are three dimensions of the feature, respectively. N is obtained by multiplying the dimension W by H. />
Is a Query value obtained by convolution mapping of hidden states,>
the importance of each feature of the hidden state is queried by matrix multiplication. />
Key value reflects the importance degree of different features in the hidden state.
The ISTC-SA module is realized by letting
And->
Multiplication results in concealmentSimilarity score for all information of the information itself
. The dimension of the similarity score is n×n. The self-attention score represents the extent to which long-term prediction of N features requires attention to the other N features. The stronger the similarity score between two features, the stronger the correlation between the two features for long-term prediction:
the value of the similarity score is not in one dimension at this time. To facilitate the memorization and computation of the model, the ISTC-SA module unifies all values into one dimension through normalization operations. Normalized similarity score and corresponding
Multiplying to obtain the self-attention score of the hidden state about itself>
。/>
Contains a large amount of long-term spatiotemporal context information:
in the formula
Representing a logarithmic base, the value is approximately equal to 2.71./>
And->
Is a real number, representing +.>
Personal specialSyndrome, no%>
The value range of the characteristic is [1, N ]]。
Through a self-attention mechanism, the hidden state updates the importance degree of different time-space context information for long-term prediction, and a large amount of long-term time-space context information is obtained.
To further increase the amount of information of the long-term spatiotemporal context information in the hidden state, the ISTC-SA structure helps the hidden state learn the long-term information in the spatiotemporal memory state through a self-attention mechanism.
For the following
The ISTC-SA structure firstly maps the ISTC-SA structure to different feature spaces through convolution operation to obtain Key: />
，Value：/>
。/>
Key value reflects the importance degree of different features in the space-time memory state. Hidden status pass->
And->
Multiplying to obtain importance degree of different time space information in time space memory state, and generating similarity score +.>
：
The hidden state at this time yields long-term spatiotemporal context information in itself and in the spatiotemporal memory state. The ISTC-SA module fuses two kinds of information (self and long-term space-time context information in space-time memory state) with the hidden state, thereby enhancing the information quantity of the long-term space-time context information in the hidden state and obtaining a new hidden state
。
The ISTC-SA module needs to increase the space-time memory state after increasing the information amount of the long-term space-time context in the hidden state through the self-attention mechanism
With respect to the amount of information of the long-term space-time context, a specific flow is shown in fig. 4.
ISTC-SA enhances hidden state through five time-space interaction operations
Expression ability for long-term context information while increasing spatiotemporal memory state +.>
Information content of medium-long term upper information, generating new hidden state +.>
And space-time memory state->
. Through five time-space interaction operations, hiding state +.>
And space-time memory state->
Including important long-term as well as short-term spatiotemporal context information. At this time, hidden state->
Through a gating mechanism, corresponding forget gates, input gates and candidate states are generated. Among the candidate states there is a large amount of spatiotemporal context information that requires spatiotemporal memory states to be memorized:
the ISTC-SA module screens the information in the space-time memory state through a forgetting door, helps the space-time memory state to selectively memorize long-term and short-term space-time context information which helps the model to predict the change trend of the echo region, and forgets redundant space-time information.
And meanwhile, the ISTC-SA-MIM model controls the information quantity which needs to be memorized in the space-time memory state in the candidate states through the input gate. The space-time memory state completes the memory of space-time context information through the forgetting gate, the input gate and the candidate state, and a new space-time memory state is generated:
Thus the ISTC-SA module passes
And->
An output gate is created. The ISTC-SA module generates a new hidden state by outputting the information quantity of the space-time context information and the non-stationarity information which need to be memorized by the gate control ISTC-SA-MIM model>
：
；
At this time, the ISTC-SA-MIM of the second layer completes the learning of the space-time context information and the non-stationarity information, and generates the prediction information for the next moment
。
The ISTC-SA-MIM model of the second layer will itself generate
And->
And transmitting the information to an ISTC-SA-MIM model of the third layer, and helping the ISTC-SA-MIM model of the third layer to continuously learn the space-time context information and the non-stationarity information. The ISTC-SA-MIM model of the third layer generates +.>
And->
. Likewise, the ISTC-SA-MIM model of the third layer will itself generate +.>
And->
The ISTC-SA-MIM model passed to the fourth layer is learned. The ISTC-SA-MIM model of the fourth layer is generated by learning>
And->
. The ISTC-SA-MIM model of the fourth layer extracts +.>
Important time-space information of the current time to obtain a prediction result of the current time: />
。
And meanwhile, the ISTC-SA-MIM model of the fourth layer transmits the generated space-time memory state to the ST-LSTM model of the first layer at the next moment, so that the ST-LSTM model is helped to learn the global space-time context information learned at the current moment. The system now completes the prediction of the radar echo map at a moment. In order to continuously predict the trend of the echo region in the radar echo map, the prediction result at that time may be input to the system to continuously predict.
Claims (10)
1. A precipitation prediction system based on MIM model and space-time interaction memory comprises an ST-LSTM model and at least one ISTC-SA-MIM model which are connected in sequence; the ST-LSTM model extracts space-time information in the radar echo diagram at the previous moment, generates a hidden state and a space-time memory state, and transmits the hidden state and the space-time memory state to the ISTC-SA-MIM model; the method is characterized in that: the ISTC-SA-MIM model consists of an ISTC-SA module and an MIM module; the ISTC-SA-MIM model learns long-term and short-term space-time context information through input space-time memory states and hidden states; and learning the nonstationary information through the long-term memory state to generate a new hidden state containing the short-term space-time context information and the nonstationary information, and predicting a radar echo diagram at the current moment.
2. The precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 1, wherein: and the ISTC-SA module carries out iterative interaction on the hidden state and the space-time memory state through space-time interaction operation, and increases the information quantity of the two states on space-time context information.
3. The precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 2, wherein: the ISTC-SA module extracts important short-term space-time context information in the space-time memory state through convolution operation; mapping the important information into a numerical form suitable for hiding state memory through nonlinear mapping; the hidden state fuses the values through the Hadamard product and nonlinear mapping, and the information quantity of the hidden state relative to short-term space-time context information is increased.
4. A precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 3, wherein: and the ISTC-SA module extracts the space-time context information in the updated hidden state through convolution operation, and increases the information quantity of the space-time memory state through nonlinear mapping.
5. The precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 2, wherein: the ISTC-SA module updates the importance degree of different space-time context information in the space-time memory state and the hidden state through a self-attention mechanism, enhances the expression capability of the hidden state on the long-term context information, and increases the information quantity of the long-term context information in the space-time memory state.
6. The precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 5, wherein: the ISTC-SA module generates a series of gates through a gating mechanism, helps the space-time memory state to selectively forget redundant repeated space-time context information contained in the space-time memory state, memorizes new long-period space-time context information and generates a new space-time memory state.
7. The precipitation prediction system based on MIM model and spatiotemporal interaction memory according to claim 1, wherein: the MIM module comprises MIM-N and MIM-S; the MIM-N is used for memorizing non-stationarity information; the MIM-S is used for memorizing the stationarity information and integrating the non-stationarity information and the stationarity information to generate fusion information.
8. Precipitation prediction system based on MIM model and spatiotemporal interaction memory according to any of claims 1-7, characterized in that: the radar echo diagram is a real radar echo diagram or a radar echo diagram generated by prediction.
9. A precipitation prediction method based on MIM model and space-time interaction memory comprises the following steps:
inputting a real radar echo map or a radar echo map generated by prediction at a previous moment into a precipitation prediction system according to any of claims 1-7;
and outputting a radar echo diagram at the current moment through prediction.
10. A computer storage medium, characterized by: precipitation prediction system comprising a MIM model and a spatiotemporal interaction memory based precipitation prediction system according to any of claims 1-7, whereby the system is operated to achieve precipitation prediction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310565194.9A CN116307283B (en) | 2023-05-19 | 2023-05-19 | Precipitation prediction system and method based on MIM model and space-time interaction memory |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310565194.9A CN116307283B (en) | 2023-05-19 | 2023-05-19 | Precipitation prediction system and method based on MIM model and space-time interaction memory |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116307283A true CN116307283A (en) | 2023-06-23 |
| CN116307283B CN116307283B (en) | 2023-08-18 |
Family
ID=86824330
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310565194.9A Active CN116307283B (en) | 2023-05-19 | 2023-05-19 | Precipitation prediction system and method based on MIM model and space-time interaction memory |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116307283B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116719002A (en) * | 2023-08-08 | 2023-09-08 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH10227872A (en) * | 1996-12-13 | 1998-08-25 | Nippon Telegr & Teleph Corp <Ntt> | Meteorological image forecast method, equipment and record medium recording meteorological image forecast program |
| CN107657627A (en) * | 2017-08-24 | 2018-02-02 | 北京理工大学 | Space-time contextual target tracking based on human brain memory mechanism |
| DE102018102027A1 (en) * | 2017-01-31 | 2018-08-02 | GM Global Technology Operations LLC | EFFICIENT SITUATION AWARENESS THROUGH EVENT GENERATION AND EPISODIC STORAGE FOR AUTONOMOUS DRIVE SYSTEMS |
| CN112698427A (en) * | 2020-12-09 | 2021-04-23 | 最美天气(上海)科技有限公司 | Short-term forecasting method and system based on space-time forecasting model |
| CN113610329A (en) * | 2021-10-08 | 2021-11-05 | 南京信息工程大学 | Short-time rainfall approaching forecasting method of double-current convolution long-short term memory network |
| CN115310724A (en) * | 2022-10-10 | 2022-11-08 | 南京信息工程大学 | Precipitation prediction method based on Unet and DCN _ LSTM |
| CN115346101A (en) * | 2022-08-22 | 2022-11-15 | 南京信息工程大学 | Method for realizing radar echo extrapolation model based on depth space-time fusion network |
| CN115390164A (en) * | 2022-10-27 | 2022-11-25 | 南京信息工程大学 | Radar echo extrapolation forecasting method and system |
| 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 |
| CN115792913A (en) * | 2022-05-16 | 2023-03-14 | 湖南师范大学 | Radar echo extrapolation method and system based on time-space network |
| CN115862319A (en) * | 2022-11-08 | 2023-03-28 | 贵州民族大学 | Traffic flow prediction method for space-time diagram self-encoder |
-
2023
- 2023-05-19 CN CN202310565194.9A patent/CN116307283B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH10227872A (en) * | 1996-12-13 | 1998-08-25 | Nippon Telegr & Teleph Corp <Ntt> | Meteorological image forecast method, equipment and record medium recording meteorological image forecast program |
| DE102018102027A1 (en) * | 2017-01-31 | 2018-08-02 | GM Global Technology Operations LLC | EFFICIENT SITUATION AWARENESS THROUGH EVENT GENERATION AND EPISODIC STORAGE FOR AUTONOMOUS DRIVE SYSTEMS |
| CN107657627A (en) * | 2017-08-24 | 2018-02-02 | 北京理工大学 | Space-time contextual target tracking based on human brain memory mechanism |
| CN112698427A (en) * | 2020-12-09 | 2021-04-23 | 最美天气(上海)科技有限公司 | Short-term forecasting method and system based on space-time forecasting model |
| CN113610329A (en) * | 2021-10-08 | 2021-11-05 | 南京信息工程大学 | Short-time rainfall approaching forecasting method of double-current convolution long-short term memory network |
| CN115792913A (en) * | 2022-05-16 | 2023-03-14 | 湖南师范大学 | Radar echo extrapolation method and system based on time-space network |
| CN115346101A (en) * | 2022-08-22 | 2022-11-15 | 南京信息工程大学 | Method for realizing radar echo extrapolation model based on depth space-time fusion network |
| CN115310724A (en) * | 2022-10-10 | 2022-11-08 | 南京信息工程大学 | Precipitation prediction method based on Unet and DCN _ LSTM |
| CN115390164A (en) * | 2022-10-27 | 2022-11-25 | 南京信息工程大学 | Radar echo extrapolation forecasting method and system |
| 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 |
| CN115862319A (en) * | 2022-11-08 | 2023-03-28 | 贵州民族大学 | Traffic flow prediction method for space-time diagram self-encoder |
Non-Patent Citations (3)
| Title |
|---|
| ZHIFENG MA等: "MM-RNN: A Multimodal RNN for Precipitation Nowcasting", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》, vol. 61, pages 1 - 14 * |
| 张晋文: "基于循环网络的气象雷达回波外推研究", 《中国优秀硕士学位论文全文数据库（电子期刊）信息科技辑》, no. 3, pages 136 - 1688 * |
| 胡强等: "基于改进人工蜂群算法的云制造服务组合优化方法", 《通信学报》, vol. 44, no. 1, pages 200 - 210 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116719002A (en) * | 2023-08-08 | 2023-09-08 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
| CN116719002B (en) * | 2023-08-08 | 2023-10-27 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116307283B (en) | 2023-08-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wayne et al. | Unsupervised predictive memory in a goal-directed agent | |
| Arietta et al. | City forensics: Using visual elements to predict non-visual city attributes | |
| Quilodrán-Casas et al. | Digital twins based on bidirectional LSTM and GAN for modelling the COVID-19 pandemic | |
| CN116307283B (en) | Precipitation prediction system and method based on MIM model and space-time interaction memory | |
| CN111444243A (en) | User behavior prediction image method and system based on track information | |
| US20230400301A1 (en) | Tropical instability wave early warning method and device based on temporal-spatial cross-scale attention fusion | |
| CN108090686B (en) | Medical event risk assessment analysis method and system | |
| CN110781970B (en) | Classifier generation method, device, equipment and storage medium | |
| Sun et al. | PGCNet: patch graph convolutional network for point cloud segmentation of indoor scenes | |
| CN114943365A (en) | Rainfall estimation model establishing method fusing multi-source data and rainfall estimation method | |
| Qin et al. | DeepFR: A trajectory prediction model based on deep feature representation | |
| CN115719036A (en) | Space-time process simulation method and system based on stacked space-time memory unit | |
| Qian et al. | A fire monitoring and alarm system based on channel-wise pruned YOLOv3 | |
| CN117197632A (en) | Transformer-based electron microscope pollen image target detection method | |
| Mao et al. | A hybrid deep learning network for forecasting air pollutant concentrations | |
| CN114611990A (en) | Method and device for evaluating contribution rate of element system of network information system | |
| Liu et al. | 3DCNN landslide susceptibility considering spatial-factor features | |
| Ji | Research on Recognition Effect of DSCN Network Structure in Hand-Drawn Sketch | |
| Zhang | Deep learning-based hybrid short-term solar forecast using sky images and meteorological data | |
| Nejad et al. | SERT: A Transfomer Based Model for Spatio-Temporal Sensor Data with Missing Values for Environmental Monitoring | |
| Huang et al. | Spatiotemporal prediction in three-dimensional space by separating information interactions | |
| Tekin | Spatiotemporal Series Analysis and Forecasting: New Deep Learning Architectures On Weather and Crime Forecasting | |
| Wang | Big Data-Based Analysis and Prediction of International Events | |
| Yang et al. | Convolutional Self-Attention Neural Network for Multi-Step Forecasting of Environmental Image Series | |
| Shi et al. | An Attention-based Context Fusion Network for Spatiotemporal Prediction of Sea Surface Temperature |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |