Method for predicting tropical cyclone intensity based on decision-driven memory update network

Abstract

The application discloses a method for predicting the intensity of a tropical cyclone based on a decision-driven memory updating network, and particularly relates to the technical field of predicting the intensity of a tropical cyclone, which is used for predicting the intensity of a tropical cyclone at a current moment, inputting a data set composed of the intensities of tropical cyclones at multiple historical moments and current observation as input features into a splicing module for vector splicing and fusion processing to obtain a context feature vector, simultaneously inputting the context feature vector into a decision branch for use, sequentially processing in the decision branch through a soft decision forest and a forest representation module to obtain a decision vector, sequentially inputting the decision vector into a decision perception memory updating unit, a cell state updating module and a hidden state calculation module, and finally obtaining the intensity of the tropical cyclone at the current moment. According to the application, the historical memory and candidate memory are adaptively fused according to the decision vector, the cell state updating is realized in a decision-driven manner, and finally the intensity of the tropical cyclone at the current moment is accurately predicted.

Description

Technical Field

[0001] This invention relates to the field of cyclones, and more specifically to a method for predicting the intensity of tropical cyclones. Background Technology

[0002] Existing tropical cyclone intensity prediction models have relatively simple memory update mechanisms and lack the ability to make conditional decisions based on the current state and the ability to update memory adaptively. This leads to insufficient retention of historical effective information or accumulation of invalid information during the complex evolution process, which affects the accuracy and stability of tropical cyclone intensity prediction. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting tropical cyclone intensity based on a decision-driven memory update network, thereby solving the problems of poor accuracy and stability in current model predictions of cyclone intensity.

[0004] To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0005] A method for predicting tropical cyclone intensity based on a decision-driven memory update network uses a dataset consisting of tropical cyclone intensities from multiple historical moments. and current observations The input features are fed into the concatenation module for vector concatenation and fusion processing, resulting in... and Context feature vector Simultaneously, the context feature vector is input directly into the decision branch, where it is processed sequentially through the soft decision forest and forest representation modules to obtain the decision vector. Another line of input is fed into the gating branch to obtain candidate memories. ; Decision vector and candidate memories All data are sequentially input into the decision-aware memory update unit, the cell state update module, and the hidden state calculation module; ultimately, the intensity of the tropical cyclone at the current moment is obtained. .

[0006] Furthermore, the aforementioned for The dataset consists of cyclone intensities from 7 moments prior to the current moment. ;

[0007] ;

[0008] In the formula: for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time.

[0009] Furthermore, the context feature vector is fed into a gating branch, which includes an input gate, a candidate gate, and an output gate. The output vector of the input gate is then... and the output vector of the candidate gate Perform Hadamard product processing to obtain the final candidate memory. Input into the decision perception memory update unit.

[0010] Furthermore, the cell state update module receives the candidate memory generated by the gating branch and the decision vector generated by the decision branch. Then, the current state of the memory cell is obtained through iterative calculation in the decision-aware memory update unit. The specific calculation formula is as follows:

[0011] ;

[0012] In the above formula, The state of the cell in the previous moment is remembered; ⋅ This is a vector concatenation and fusion function. This is for the Hadamard product operation.

[0013] Furthermore, the current cell state obtained by the cell state update module is used to remember the cell state. The input is fed into the hidden state calculation module, which then stores the cell state at the current time. Features obtained from the output gate in the gated branch The process is then performed to obtain the current tropical cyclone intensity. .

[0014] Furthermore, the hidden state calculation module remembers the cell state at the current moment. With output gate characteristics The specific calculation formula is as follows:

[0015] ;

[0016] In the above formula: ⋅ It is the hyperbolic tangent activation function; This is for the Hadamard product operation.

[0017] Furthermore, the soft decision forest will generate Each soft decision tree is fed with a context feature vector. Soft routing calculations are performed within each soft decision tree node to obtain the leaf probability of each leaf node. The leaf probability is then used to calculate the leaf embedding vector corresponding to each leaf node. Weighted summation, outputting the feature vector of the corresponding soft decision tree. Aggregate the feature vectors output by all soft decision trees to obtain the decision vector. .

[0018] Furthermore, the soft routes in the aforementioned soft decision forest all adopt... The function performs the processing.

[0019] Furthermore, the decision vector The range of values ​​is .

[0020] Furthermore, a multilayer perceptron is used to aggregate the feature vectors output by all soft decision trees to obtain the decision vector. .

[0021] By adopting the above technical solution, the beneficial technical effects of the present invention are as follows:

[0022] This application generates decision vectors through soft decision forests and uses these decision vectors to drive the state update of memory cells, enabling the model to adaptively determine the degree of fusion between historical memory and candidate memory based on the current state, thereby improving memory management capabilities and enhancing the accuracy and stability of tropical cyclone intensity prediction. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for predicting tropical cyclone intensity based on a decision-driven memory update network.

[0024] Figure 2 This is a flowchart for decision-making branches. Detailed Implementation

[0025] like Figure 1 As shown, the method for predicting tropical cyclone intensity based on decision-driven memory update networks obtains a dataset of tropical cyclone intensities from the IBTRASC database, consisting of data from multiple historical moments. And current observations obtained from ERA5 reanalysis data The input features are fed into the concatenation module for vector concatenation and fusion processing, resulting in... and Context feature vector Simultaneously, the context feature vector is input directly into the decision branch, where it is processed sequentially through the soft decision forest and forest representation modules to obtain the decision vector. , decision vector The data is sequentially input into the decision-aware memory update unit, the cell state update module, and the hidden state calculation module; finally, the intensity of the tropical cyclone at the current moment is obtained. .

[0026] The For the data obtained from the IBTrACS database The dataset consisting of cyclone intensities from seven moments prior to the current moment. , for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time.

[0027] ;

[0028] In the formula: This is a dataset consisting of sea surface temperatures at seven different times. ; for The sea surface temperature at that moment. for The sea surface temperature at that moment. for The sea surface temperature at that moment. for The sea surface temperature at that moment. for The sea surface temperature at that moment. for The sea surface temperature at that moment. for The sea surface temperature at that moment;

[0029] This is a data set consisting of 850 hPa zonal winds at seven different times. ; for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment. for The zonal wind value of 850 hPa at that moment;

[0030] The dataset consists of radial winds at 850 hPa at seven different times. ; for The radial wind value of 850 hPa at that moment. for The radial wind value of 850 hPa at that moment. for The radial wind value of 850 hPa at that moment. for The radial wind value of 850 hPa at that moment. for The radial wind value of 850 hPa at that moment. for The radial wind value of 850 hPa at that moment. for Radial wind value of 850 hPa at that moment;

[0031] The data set consists of 200 hPa zonal winds at 7 different times; ; for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment. for The 200 hPa zonal wind value at that moment;

[0032] The dataset consists of radial winds at 200 hPa at seven different times. ; for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment. for The radial wind value at 200 hPa at that moment;

[0033] This is a dataset consisting of 700 hPa relative humidity at 7 different times. ; for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment. for The relative humidity value at 700 hPa at that moment;

[0034] VWS The dataset consists of vertical wind shear at seven time points; ; for The vertical wind shear value at time _____ for The vertical wind shear value at time _____ for The vertical wind shear value at time _____ for The vertical wind shear value at time _____ for The vertical wind shear value at time _____ for The vertical wind shear value at time _____ for The vertical wind shear value at any given time.

[0035] Furthermore, the context feature vector is fed into a gating branch, which includes an input gate, a candidate gate, and an output gate. The output vector of the input gate is then... and the output vector of the candidate gate Perform Hadamard product processing to obtain the final candidate memory. Input into the decision perception memory update unit.

[0036] Furthermore, the cell state update module receives the candidate memory generated by the gating branch and the decision vector generated by the decision branch. Then, the current state of the memory cell is obtained through iterative calculation in the decision-aware memory update unit. The specific calculation formula is as follows:

[0037] ;

[0038] In the above formula, The state of the cell in the previous moment is remembered; ⋅ This is a vector concatenation and fusion function. This is for the Hadamard product operation.

[0039] It is not a single meteorological variable, but rather a network that has passed through... A high-dimensional potential state vector is formed after processing information at each moment. It reflects: historical evolution memory + current environmental driving force + decision-making mechanism screening, rather than a specific meteorological physical quantity.

[0040] It can comprehensively encode information from a past period of time, for example:

[0041] The previous intensity level of the tropical cyclone;

[0042] Intensity change trend;

[0043] Rate of intensity change;

[0044] Preliminary environmental conditions;

[0045] Cyclone response to environmental changes;

[0046] Historical dependencies are strengthened or weakened during the process.

[0047] Furthermore, the current cell state obtained by the cell state update module is used to remember the cell state. The input is fed into the hidden state calculation module, which then stores the cell state at the current time. Features obtained from the output gate in the gated branch The process is then performed to obtain the current tropical cyclone intensity. .

[0048] Furthermore, the hidden state calculation module remembers the cell state at the current moment. With output gate features The specific calculation formula is as follows:

[0049] ;

[0050] In the above formula: ⋅ It is the hyperbolic tangent activation function; This is for the Hadamard product operation.

[0051] Furthermore, such as Figure 2 As shown, the soft decision forest will generate Each soft decision tree is fed with a context feature vector. Soft routing calculations are performed within each soft decision tree node to obtain the leaf probability of each leaf node. The leaf probability is then used to calculate the leaf embedding vector corresponding to each leaf node. Weighted summation, outputting the feature vector of the corresponding soft decision tree. The decision vector is obtained by aggregating the feature vectors output by all soft decision trees using a multilayer perceptron. The soft routes in the aforementioned soft decision forest all adopt... The function processes the decision vector. The range of values ​​is The decision vector The hidden layer dimension is The leaf embedding vector Specifically represented as the first The first soft decision tree The leaf embedding vectors corresponding to each leaf node are aggregated with the feature vectors output by all soft decision trees to obtain the decision vector. .

[0052] use: ;

[0053] In the formula: MLP stands for Multilayer Perceptron; For feature concatenation function; The feature vector of the first soft decision tree, Let be the feature vector of the Kth tree.

[0054] Data Comparison:

[0055] 410 data points were obtained from storm cyclone intensity data obtained from IBTRASC and satellite observation datasets and reanalysis datasets obtained from ERA5 environmental fields, including: marine environment, low-level wind field, high-level wind field, humidity, and vertical wind shear. The above data were input into different prediction models, resulting in the following table.

[0056] Table 1 Comparison of Error Data for Multiple Models

[0057] .

[0058] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting tropical cyclone intensity based on a decision-driven memory update network, characterized in that, A dataset consisting of data on tropical cyclone intensity at multiple historical moments. and current observations The input features are fed into the concatenation module for vector concatenation and fusion processing, resulting in... and Context feature vector Simultaneously, the context feature vector is fed directly into the decision branch, where it is processed sequentially through the soft decision forest and forest representation modules to obtain the decision vector. Another line of input is fed into the gating branch to obtain candidate memories. ; Decision vector and candidate memories All data are sequentially input into the decision-aware memory update unit, the cell state update module, and the hidden state calculation module; ultimately, the intensity of the tropical cyclone at the current moment is obtained. .

2. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 1, characterized in that, The for The dataset consists of cyclone intensities from 7 moments prior to the current moment. ; ； In the formula: for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time; for Cyclone intensity value at any given time.

3. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 1, characterized in that, The gated branch includes an input gate, a candidate gate, and an output gate, and then the output vector of the input gate is... and the output vector of the candidate gate Perform Hadamard product processing to obtain the final candidate memory. Input into the decision perception memory update unit.

4. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 1, characterized in that, The cell state update module receives candidate memories generated by the gating branch and decision vectors generated by the decision branch. Then, the current state of the memory cell is obtained through iterative calculation in the decision-aware memory update unit. The specific calculation formula is as follows: ； In the above formula, The state of the cell in the previous moment is remembered; ⋅ This is a vector concatenation and fusion function; This is for the Hadamard product operation.

5. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 1, characterized in that, The cell state update module obtains the current cell state at that moment. The input is fed into the hidden state calculation module, which then stores the cell state at the current time. Features obtained from the output gate in the gated branch The process is then performed to obtain the current tropical cyclone intensity. .

6. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 5, characterized in that, The hidden state calculation module remembers the cell state at the current moment. With output gate characteristics The specific calculation formula is as follows: ； In the above formula: ⋅ The hyperbolic tangent activation function is used. This is for the Hadamard product operation.

7. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 1, characterized in that, The soft decision forest will generate Each soft decision tree is fed with a context feature vector. Soft routing calculations are performed within each soft decision tree node to obtain the leaf probability of each leaf node. The leaf probability is then used to calculate the leaf embedding vector corresponding to each leaf node. Weighted summation, outputting the feature vector of the corresponding soft decision tree. ; By aggregating the feature vectors output from all soft decision trees, we obtain the decision vector. .

8. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 7, characterized in that, The soft routes in the soft decision forest all adopt... The function performs the processing.

9. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 7, characterized in that, The decision vector The range of values ​​is .

10. The method for predicting tropical cyclone intensity based on a decision-driven memory update network according to claim 6, characterized in that, The feature vectors output by all soft decision trees are aggregated using a multilayer perceptron to obtain the decision vector. .

Images