A data and knowledge combined driving storm surge early warning method and device

Abstract

The application discloses a data and knowledge combined driving storm surge early warning method and device, and the method comprises the following steps: taking a probability distribution graph converted from a real value of a cyclone center coordinate as a new label; using the converted probability distribution graph as the label to decode a cyclone track prediction feature set to output a cyclone center position probability distribution prediction graph of a preset latitude and longitude range at different time steps; decoding a cyclone intensity prediction feature set to output a grid pressure distribution and a wind speed distribution; traversing the cyclone center position probability distribution prediction graph to obtain a point with the maximum probability value as a cyclone center point; using the predicted cyclone center position and cyclone related knowledge to guide and associate the cyclone pressure and wind speed distribution to obtain a cyclone intensity prediction value; and inputting the cyclone intensity prediction value, cyclone track information and storm surge tide harmonic analysis data input in combination into a long short-term memory recurrent network to predict storm surge intensity at different time steps. The device comprises a processor and a memory.

Description

Technical Field

[0001] This invention relates to the field of storm surge intensity prediction, and in particular to a storm surge early warning method and apparatus driven by data and knowledge. Background Technology

[0002] Storm surge, also known as storm water rise, is an abnormal rise in sea level caused by severe atmospheric disturbances, such as cyclones (tropical cyclones and extratropical cyclones). Generally, the low pressure and strong winds accompanying a cyclone's landfall are the main causes of storm surges. This advancing swell combines with the normal tide to form storm surge. While conventional numerical and empirical methods for storm surge forecasting have been developed to some extent, the variation of storm surges remains a complex marine engineering problem. Many factors influence storm surge variations, including cyclone center pressure, cyclone movement speed, and coastal topography.

[0003] A cyclone is a large vortex in the atmosphere of the Northern (Southern) Hemisphere where horizontal airflow rotates counterclockwise (clockwise). At the same altitude, the air pressure at the center of a cyclone is lower than the surrounding areas, hence it is also called a low-pressure weather system. Cyclones are generally classified into tropical cyclones and extratropical cyclones according to their activity range. Tropical cyclones, in particular, usually bring strong winds and heavy rainfall when they make landfall, leading to severe storm surges. my country has a long coastline and is located in the Northwest Pacific Ocean, a region prone to cyclones. For coastal areas, timely and accurate forecasting of storm surges and monitoring of tidal changes are crucial for my country's natural disaster prevention and control efforts and are of great significance for protecting the lives and property of people in coastal areas.

[0004] With the establishment of ocean observation stations, ground stations, and meteorological satellites, the amount of available data is constantly increasing. Deep learning has shown great advantages in handling big data problems, capable of extracting hidden features from massive amounts of data. Significant progress has been made in applying deep learning to storm surge prediction: Lee et al. [1] By utilizing artificial neural networks and relying on data-driven methods to predict storm surges, the powerful potential of deep learning in the field of storm surge prediction has been initially demonstrated; Chen et al. [2] Using long short-term memory neural networks to extract temporal information from data, storm surge prediction in the East China Sea was achieved; Adeli et al. [3] This paper proposes a spatiotemporal convolutional recurrent neural network to fully capture the temporal and spatial correlations of data, thereby improving the accuracy of storm surge prediction. However, current storm surge intensity prediction relies solely on data-driven approaches and lacks physical correlation and knowledge guidance. Based on relevant research... [4] For storm surge prediction, the intensity and trajectory of the cyclone are the most important factors. Accurate cyclone intensity prediction and trajectory tracking play an extremely important role in storm surge prediction, which helps to improve the accuracy of storm surge prediction and enhance the detection and prevention capabilities of natural disasters in my country's coastal areas. Summary of the Invention

[0005] This invention provides a storm surge early warning method and device driven by data and knowledge. This invention improves the accuracy of storm surge prediction, which is beneficial for enhancing the natural disaster prevention and control capabilities of southeastern coastal areas of my country, reducing flood disasters, and protecting the lives and property of people in coastal areas of my country. See the description below for details:

[0006] Firstly, a storm surge early warning method driven by data and knowledge, the method comprising:

[0007] 1) Use cross-task attention mechanism to update the features of the separated cyclone intensity-specific and common feature sets and cyclone trajectory-specific and common feature sets to obtain the predicted feature sets of cyclone intensity and cyclone trajectory;

[0008] 2) Convert the true values ​​of the cyclone center coordinates in the cyclone dataset into a probability distribution map within a preset latitude and longitude range, and use the probability distribution map after the conversion of the true values ​​of the cyclone center coordinates as the new label;

[0009] 3) Using the transformed probability distribution map as labels, decode the cyclone trajectory prediction feature set and output the probability distribution prediction map of the cyclone center position within a preset latitude and longitude range at different time steps; decode the cyclone intensity prediction feature set and output gridded pressure distribution and wind speed distribution.

[0010] 4) Traverse the probability distribution prediction map of the cyclone center location, and take the point with the highest probability value as the cyclone center point. Using the predicted cyclone center location and cyclone-related knowledge, combine and guide the correlation of cyclone pressure and wind speed distribution to obtain the predicted cyclone intensity value.

[0011] 5) Input the predicted cyclone intensity, cyclone trajectory information, and the jointly input storm surge level harmonic analysis data into the long short-term memory cyclic network to predict the storm surge intensity at different time steps.

[0012] In a second aspect, a storm surge early warning device driven by data and knowledge, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the steps of the method described in any of the first aspects.

[0013] The beneficial effects of the technical solution provided by this invention are:

[0014] 1. This invention focuses on two tasks: cyclone intensity prediction and cyclone trajectory prediction. It separates the representation of the two tasks by using unique features and common features respectively, utilizes adversarial domain adaptation to deeply mine the common features of different tasks, and uses cross-task attention to achieve coupling and correlation between the tasks.

[0015] 2. This invention utilizes the relevant knowledge and physical correlation between cyclone intensity and trajectory, and combines cyclone intensity and trajectory using the cyclone center point to achieve joint prediction of the spatial distribution of cyclone intensity and trajectory;

[0016] 3. This invention utilizes the predicted cyclone intensity, trajectory, harmonic analysis of storm surge level, and topographic features to predict the storm surge intensity at different time steps in the future.

[0017] Therefore, this invention enables the predicted cyclone intensity and trajectory output to have physical meaning and spatial distribution correlation. By integrating the predicted cyclone intensity, trajectory, and harmonic analysis of tide level and topographic features, it improves the accuracy of storm surge and tide level change prediction, enhances the ability of my country's southeastern coastal areas to prevent and control natural disasters, reduces floods, and protects the lives and property of people in my country's coastal areas. Attached Figure Description

[0018] Figure 1 A flowchart of a storm surge early warning method driven by data and knowledge;

[0019] Figure 2 A schematic diagram of a storm surge early warning method driven by data and knowledge;

[0020] Figure 3 This is another schematic diagram of a storm surge early warning method driven by data and knowledge. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0022] Example 1

[0023] A data- and knowledge-driven storm surge early warning method, see [link to relevant documentation]. Figure 1 The method includes the following steps:

[0024] Step 101: Use a convolutional neural network-based encoding module to encode the features of the multi-source environmental field data and extract the spatiotemporal features of the multi-source environmental field;

[0025] Step 102: Based on the characteristics and commonalities among different tasks, the spatiotemporal feature sets of multiple multi-source environmental fields are mapped to different feature spaces to obtain cyclone intensity-specific feature sets and cyclone trajectory-specific feature sets; then, the two specific feature sets are mapped to the same feature space to obtain cyclone intensity-shared feature sets and cyclone trajectory-shared feature sets. The differences between the cyclone intensity-shared feature sets and cyclone trajectory-shared feature sets are continuously reduced by using adversarial domain adaptation.

[0026] Step 103: Use cross-task attention mechanism to update the features of the separated cyclone intensity-specific and common feature sets and cyclone trajectory-specific and common feature sets, strengthen the coupling and correlation between the two tasks, and obtain the cyclone intensity prediction feature set and the cyclone trajectory prediction feature set.

[0027] Step 104: Convert the true values ​​of the cyclone center coordinates in the cyclone dataset into a 31°×31° probability distribution map, and use the probability distribution map after the conversion of the true values ​​of the cyclone center coordinates as the new label;

[0028] The aforementioned cyclone datasets typically come from publicly available data sources such as the National Meteorological Administration.

[0029] Step 105: For the cyclone trajectory prediction task, using the transformed probability distribution map as a label, decode the cyclone trajectory prediction feature set and output the 31°×31° cyclone center position probability distribution prediction map at different time steps; for the cyclone intensity prediction task, decode the cyclone intensity prediction feature set and output the 31°×31° gridded pressure distribution and 31°×31° gridded wind speed distribution at different time steps.

[0030] Step 106: Traverse the 31°×31° cyclone center location probability distribution prediction map, and obtain the point with the highest probability value as the cyclone center point, thereby obtaining the trajectory of the cyclone center at different time steps; use the predicted cyclone center location and cyclone-related knowledge to guide the association of cyclone pressure and wind speed distribution to obtain the predicted value of cyclone intensity (including cyclone center pressure and maximum wind speed near the center).

[0031] Step 107: Input the predicted cyclone intensity (including cyclone center pressure and maximum wind speed near the center) and cyclone trajectory information together with the storm surge level harmonic analysis data into the Long Short-Term Memory Network (LSTM) to predict the storm surge intensity at different time steps.

[0032] In summary, the embodiments of the present invention improve the accuracy of cyclone intensity and trajectory through the above steps 101-107, and achieve accurate prediction of storm surge intensity by accurately predicting cyclone intensity and trajectory, which is conducive to enhancing the natural disaster prevention and control capabilities of southeastern coastal areas of my country.

[0033] Example 2

[0034] The scheme in Example 1 will be further described below with specific calculation formulas and examples:

[0035] 201: Using encoding operations based on convolutional neural networks, the spatiotemporal features of multi-source environmental fields are extracted;

[0036] in, d1 represents the feature dimension at each time step, R represents a real number, F represents the feature set of the multi-source environmental field, and f i This represents the characteristic of the i-th time interval.

[0037] 202: To address the characteristics of cyclone intensity and trajectory prediction tasks, the feature set F of the multi-source environmental field is mapped to different feature spaces through different linear layers;

[0038] Assume the feature set specific to the cyclone intensity prediction task is V1:

[0039]

[0040] in, This represents the mapping from the feature set of a multi-source environmental field to the cyclone intensity-specific feature set. Let the cyclone trajectory prediction task-specific feature set be V2:

[0041]

[0042] in, This involves mapping from a feature set of multi-source environmental fields to a feature set specific to cyclone trajectories. Based on the commonalities between tasks, the specific feature sets V1 and V2 are mapped to the same feature space:

[0043]

[0044]

[0045] in, This is a mapping from a set of cyclone-specific features to a set of common features. V1′ is the set of common cyclone intensity features mapped from set V1, and V2′ is the set of common cyclone trajectory features mapped from set V2, thus achieving representational separation between tasks.

[0046] Furthermore, domain adversarial approaches are implemented for V1′ and V2′ to align their distributions. Specifically, binary labels for cyclone feature attributes are added to determine whether elements in sets V1′ and V2′ belong to the cyclone intensity prediction task or the cyclone trajectory prediction task.

[0047] Using discriminator G d Constrained training continuously reduces the difference between feature sets V1′ and V2′, ensuring that the elements of sets V1′ and V2′ are common to both tasks (i.e., cyclone intensity prediction and trajectory prediction). The discriminator G... d The classification loss is:

[0048]

[0049]

[0050] l d =l d1 +l d2 (7)

[0051] Where d1 and d2 are the binary labels of cyclone feature attributes, n is the total number of training samples, and l d1 For cyclone intensity feature classification loss, l d2 The loss is used for classifying cyclone trajectory features.

[0052] 203: To enhance the interactivity between the two tasks and achieve more accurate embedding of feature representations, cross-task attention is calculated on the separated representations after feature mapping;

[0053] For the cyclone intensity prediction task, to supplement the current task with more feature information, attention weights are calculated using the shared features V1′ and V2′ after distribution alignment. The resulting weight matrix is ​​then used to update the trajectory-specific feature V2, obtaining supplementary information highly correlated with the cyclone intensity task. Similarly, a similar operation is used for the cyclone trajectory prediction task.

[0054] Specifically, attention weights are calculated using the shared representation of cyclone intensity V1′ and the shared representation of cyclone trajectory V2′:

[0055]

[0056] Where, q s Let k be the query obtained by linear projection of the common characterization set V1′ of cyclone intensity. s The key, d, is obtained by linear projection of the common representation set V2′ of cyclone trajectories. m For the feature dimension of V2′, the attention weight Attn obtained through formula (8) ti With v ti Feature updates are performed through multiplication, where v ti The value is obtained by linear projection of V2, a set of representations unique to cyclone trajectories.

[0057]

[0058] Integrate the intensity-specific representation set V1 with the cross-task coupled feature set To splice:

[0059]

[0060] Among them, F TI This is the set of predicted cyclone intensity obtained after feature splicing.

[0061] Similarly, attention weights are calculated using the shared trajectory representation V2′ and the shared intensity representation V1′:

[0062]

[0063] Where, q t Let k be the query obtained by linear projection of the common representation set V2′ of cyclone trajectories. t The key, d, is obtained by linear projection of the common characterization set V1′ of cyclone intensity. m For the feature dimension of V1′, the attention weight Attn obtained through formula (11) traj With v traj Feature updates are performed through multiplication, where v traj The value is obtained by linear projection of V1, a set of characteristics unique to cyclone intensity.

[0064]

[0065] Combine the trajectory-specific representation set V2 with the cross-task coupled feature set To splice:

[0066]

[0067] Among them, F Traj This is the set of predicted cyclone trajectories obtained after feature splicing.

[0068] The embodiments of the present invention realize the deep interaction and correlation of cyclone intensity and trajectory prediction tasks through the cross-attention mechanism described above. The new feature set obtained can supplement the feature information of this task, thereby improving the accuracy and robustness of prediction.

[0069] 204: Transform the observed values ​​of the cyclone center coordinates in the cyclone dataset into a 31°×31° probability distribution map, and use the probability distribution map after the transformation of the true values ​​of the cyclone center coordinates as the new label;

[0070] The aforementioned cyclone datasets typically come from publicly available data sources such as the National Meteorological Administration.

[0071] Specifically, based on the distribution range of the cyclone, the latitude and longitude range of the above probability distribution map is 20°-40°. In this embodiment of the invention, a probability distribution map of 31°×31° is selected for special explanation.

[0072] Construct a 31×31 all-zero matrix M, normalize the true values ​​of the cyclone center coordinates from the cyclone dataset to obtain coordinate values ​​(x, y), set M[x][y] = 1, and apply Gaussian filtering to matrix M to obtain a 31°×31° probability distribution map Y. pos :

[0073]

[0074] Where (μ1, μ2) are the normalized true values ​​of the cyclone center position coordinates.

[0075] 205: For the cyclone trajectory prediction task, the transformed probability distribution map is used as a label to decode the cyclone trajectory prediction feature set and output the probability distribution prediction map of the 31°×31° cyclone center position at different time steps; for the cyclone intensity prediction task, the cyclone intensity prediction feature set is decoded and output the gridded pressure distribution and gridded wind speed distribution at different time steps.

[0076] The embodiments of the present invention address F TI Decode and output a 31°×31° gridded pressure distribution and a 31°×31° gridded wind speed distribution, for F Traj Decoding yields a 31°×31° cyclone center location probability distribution prediction map. Assume the outputs are... in, For the predicted air pressure grid distribution, For the predicted wind speed grid distribution, This is a probability distribution prediction map of the cyclone center location. The prediction output is minimized from the true value (Y) using the mean absolute error loss. pres Y wind Y pos Differences between:

[0077]

[0078]

[0079]

[0080] Where n is the total number of cyclone training samples.

[0081] 206: Traverse the 31°×31° cyclone center location probability distribution prediction map, and obtain the point with the highest probability value as the cyclone center point, thereby obtaining the cyclone trajectory at different time steps. With the combined guidance of the predicted cyclone center trajectory and cyclone-related knowledge, the correlation between cyclone pressure and wind speed distribution is used to obtain the predicted cyclone intensity value. The cyclone intensity includes the cyclone center pressure and the maximum wind speed near the center.

[0082] According to cyclone-related knowledge, cyclone intensity is defined as the air pressure at the cyclone center or the maximum wind speed near the cyclone center. A cyclone is a low-pressure system, with the lowest air pressure at the cyclone center. Wind speed gradually increases from the outer periphery of the cyclone towards the center. Due to the difficulty of air flow at the cyclone center, there is almost no wind at the center, but the wind speed reaches its maximum near the center.

[0083] The points with high surrounding air pressure and lowest central air pressure, as well as points with high surrounding wind speed and low central wind speed, were identified from the predicted air pressure distribution grid and wind speed distribution grid, respectively.

[0084] In identifying the aforementioned local minimum points, this embodiment of the invention selects a 3°×3° region for each of the 31°×31° gridded pressure distribution and the 31°×31° gridded wind speed distribution, traversing the entire grid from left to right and top to bottom with a step size of 1. The nine values ​​within the selected region are sorted; if the center value of the 3°×3° region is the minimum, then this point is considered the local minimum pressure point or the local minimum wind speed point. It is worth noting that the region size and step size are not unique; to obtain as many local minimum points as possible, this embodiment of the invention uses a region size of 3°×3° and a step size of 1 for explanation.

[0085] Let the set of the above points be P = {p1, p2, ..., pn}. m}, Q = {q1, q2...q n}. Where p i ∈P, where i is a point in the pressure distribution grid with a lower pressure than its surroundings, i≤m, and m is the total number of local pressure minimum points traversed; q j ∈Q represents a point in the wind speed distribution grid where the wind speed is lower than the surrounding area, j≤n, and n is the total number of local points with the minimum wind speed traversed.

[0086] Construct 5° × 5° rectangles centered on elements of set P and set Q respectively, forming set A = {a1, a2, ..., a3}. m}, B={b1,b2...b n}, where a i ∈A, for p i A rectangle centered at point i, where i ≤ m; b j ∈B, for q j A rectangle centered at point j, where j≤n. A rectangle of the same size, denoted as s, is drawn with the predicted cyclone center location as the center. The ratios of the intersection and union of rectangle s with the rectangle elements in sets A and B, respectively.

[0087]

[0088]

[0089] It should be noted that the size of the rectangle is not unique; in this embodiment of the invention, a 5°×5° rectangle is selected for special explanation.

[0090] Among them, g i For set A, p i Rectangle a centered at point i The intersection-union ratio with rectangle s, t jFor set B, q j Rectangle b centered at point j The intersection-union ratio with rectangle s. Choose a suitable threshold θ (0 < θ < 1) when g i ≥θ,t j When ≥θ, then rectangle a is considered to be i b j The set G may contain the cyclone center point, constructing a set G = {g1, g2, ..., g...}. k}, T={t1,t2...t l}, k≤m, l≤n. G and T respectively contain the ratios of the intersection-union ratios of elements in sets A and B with rectangle s that are not lower than the threshold.

[0091] Sort all elements in sets G and T in ascending order, and select the point p corresponding to the element with the largest value in each set. u and q v p u ∈P, q v ∈Q. Based on knowledge in the field of cyclones, p u The pressure value at a point represents the pressure at the center of the cyclone. (Iterating through q...) v The wind speed within a 5°×5° radius is used, and the maximum wind speed is selected as the cyclone intensity. By linking the cyclone intensity and trajectory through the cyclone center point, a correlation between physical and spatial distribution is achieved.

[0092] 207: The predicted cyclone intensity (including cyclone center pressure and maximum near-center wind speed), cyclone trajectory information, and harmonic analysis data of the storm surge level are input into a total of four sets of data. These four sets of input data are processed through a Long Short-Term Memory Network (LSTM) to obtain four sets of time-series features. These four sets of time-series features are then concatenated and passed through a linear layer to output predicted storm surge intensity values ​​at different time steps. Storm surge intensity refers to the maximum storm surge increase during a single storm surge event. Training is performed using the mean absolute error loss constraint.

[0093]

[0094] in, h is the predicted storm surge intensity. s is the true value of storm surge intensity, and m is the total number of storm surge training samples.

[0095] According to the storm surge intensity classification table, storm surges can be divided into five levels based on their intensity. The predicted storm surge intensity values ​​and their levels can provide relevant professionals with a basis for flood prevention and disaster relief, enabling timely emergency responses and protecting the lives and property of people in coastal areas of my country.

[0096] Table 1. Classification of Storm Surge Intensity Levels

[0097]

[0098]

[0099] Example 3

[0100] The feasibility of Examples 1 and 2 is verified through specific experiments, as detailed below:

[0101] For example, this invention selects multi-source environmental data, including atmospheric and oceanic data, from 2000 to 2014 as input. It uses a convolutional neural network-based encoding module to extract spatiotemporal features. The encoded feature set is then processed through different linear layers for feature mapping, resulting in a set of cyclone intensity-specific features and a set of common features, as well as a set of cyclone trajectory-specific features and a set of common features. A discriminator G is added. d Align the distribution of shared feature sets for cyclone intensity with those for cyclone trajectory. Perform cross-task attention calculation on the four separated feature sets to enable deep interaction and coupling between the two tasks. For the cyclone intensity prediction task, output a 31°×31° gridded pressure distribution and a 31°×31° gridded wind speed distribution for the next 6, 12, 18, and 24 hours. For the cyclone trajectory prediction task, transform the true latitude and longitude coordinates of the cyclone center in the dataset into a 31°×31° probability distribution. Using the transformed probability distribution as labels, output 31°×31° cyclone trajectory probability distribution prediction maps for the next 6, 12, 18, and 24 hours. Iterate through the cyclone trajectory probability distribution prediction maps at different time steps; the point with the highest probability value is the predicted cyclone center trajectory. Points with high surrounding pressure and low central pressure, as well as points with high surrounding wind speed and low central wind speed, are identified from the predicted pressure and wind speed grids at different time steps. A 5°×5° rectangle is drawn centered on the found local minimum pressure and local minimum wind speed points. A 5°×5° rectangle is then drawn with the predicted cyclone center trajectory as the origin. By calculating the overlap ratio between the rectangles, the distribution range of the cyclone center in the pressure and wind speed grids is determined. Finally, guided by cyclone-related knowledge, the cyclone center pressure and near-center maximum wind speed (i.e., cyclone intensity) for the next 6, 12, 18, and 24 hours are obtained. Using the obtained cyclone intensity and trajectory data for the next 6, 12, 18, and 24 hours, along with harmonic analysis input data of the storm surge level, the storm surge intensity and level for the next 6, 12, 18, and 24 hours are jointly predicted. When a cyclone approaches or makes landfall, the predicted storm surge intensity can provide a basis for relevant technical personnel to make timely preparations for flood prevention, thereby avoiding major property losses and casualties, and enhancing the natural disaster prevention and control capabilities of my country's southeastern coastal areas.

[0102] Example 4

[0103] A storm surge early warning device driven by data and knowledge, the device comprising: a processor and a memory, wherein the memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the device to perform the following method steps:

[0104] 1) Use cross-task attention mechanism to update the features of the separated cyclone intensity-specific and common feature sets and cyclone trajectory-specific and common feature sets to obtain the predicted feature sets of cyclone intensity and cyclone trajectory;

[0105] 2) Convert the true values ​​of the cyclone center coordinates in the cyclone dataset into a probability distribution map within a preset latitude and longitude range, and use the probability distribution map after the conversion of the true values ​​of the cyclone center coordinates as the new label;

[0106] 3) Using the transformed probability distribution map as labels, decode the cyclone trajectory prediction feature set and output the probability distribution prediction map of the cyclone center position within a preset latitude and longitude range at different time steps; decode the cyclone intensity prediction feature set and output gridded pressure distribution and wind speed distribution.

[0107] 4) Traverse the probability distribution prediction map of the cyclone center location, and take the point with the highest probability value as the cyclone center point. Using the predicted cyclone center location and cyclone-related knowledge, combine and guide the correlation of cyclone pressure and wind speed distribution to obtain the predicted cyclone intensity value.

[0108] 5) Input the predicted cyclone intensity, cyclone trajectory information, and the jointly input storm surge level harmonic analysis data into the long short-term memory cyclic network to predict the storm surge intensity at different time steps.

[0109] Prior to step 1), the method further includes:

[0110] By mapping the spatiotemporal feature sets of multiple multi-source environmental fields to different feature spaces, we can obtain the unique feature sets of cyclone intensity and cyclone trajectory, and then map them to the same feature space to obtain the common feature set of cyclone intensity and cyclone trajectory. We can then use adversarial domain adaptation to reduce the differences between the common feature sets.

[0111] Specifically, the feature update using the cross-task attention mechanism on the separated sets of unique and common features of cyclone intensity and the sets of unique and common features of cyclone trajectory is as follows:

[0112] For the cyclone intensity prediction task, attention weights are calculated using the common features V1′ and V2′ after distribution alignment. The resulting weight matrix is ​​used to update the trajectory-specific feature V2 to obtain supplementary information that is highly relevant to the cyclone intensity task.

[0113] Attention weights are calculated using the shared representation of cyclone intensity V1′ and the shared representation of cyclone trajectory V2′:

[0114]

[0115] Where, q s Let k be the query obtained by linear projection of the common characterization set V1′ of cyclone intensity. s The key, d, is obtained by linear projection of the common representation set V2′ of cyclone trajectories. m For the feature dimension, the obtained attention weights Attn ti With v ti Feature updates are performed through multiplication, where v ti The value is obtained by linear projection of V2, a set of unique representations of cyclone trajectories.

[0116]

[0117] Integrate the intensity-specific representation set V1 with the cross-task coupled feature set To splice:

[0118]

[0119] Calculate the attention weights using the shared trajectory representation V2′ and the shared intensity representation V1′:

[0120]

[0121] Where, q t Let k be the query obtained by linear projection of the common representation set V2′ of cyclone trajectories. t The key is obtained by linearly projecting the common characterization set V1′ of cyclone intensity, and the attention weights Attn are obtained from this key. traj With v traj Feature updates are performed through multiplication, where v traj The value is obtained by linear projection of V1, a set of characterizations specific to cyclone intensity.

[0122]

[0123] Combine the trajectory-specific representation set V2 with the cross-task coupled feature set Then, the parts are assembled.

[0124] Furthermore, by traversing the probability distribution prediction map of the cyclone center location, the point with the highest probability value is identified as the cyclone center point. The predicted cyclone center location, combined with cyclone-related knowledge, is used to guide the correlation of cyclone pressure and wind speed distributions to obtain the predicted cyclone intensity value:

[0125] Let the set of points be P = {p1, p2, ..., pn}.m}, Q = {q1, q2...q n}, where p i ∈P, where P is a point in the pressure distribution grid with a lower pressure than its surroundings, and m is the total number of local pressure minimum points traversed; q j ∈Q, is a point in the wind speed distribution grid where the wind speed is lower than the surrounding area, j≤n, where n is the total number of local minimum wind speed points traversed;

[0126] Construct 5° × 5° rectangles centered on elements of set P and set Q respectively, forming set A = {a1, a2, ..., a3}. m}, B={b1,b2...b n};b j ∈B, for q j A rectangle centered at point j≤n; a rectangle of the same size centered at the predicted cyclone center position, denoted as s, is constructed, and the ratios of the intersection and union of rectangle elements in sets A and B with rectangle s are calculated respectively.

[0127]

[0128]

[0129] Among them, g i For set A, p i Rectangle a centered at point i The intersection-union ratio with rectangle s, t j For set B, q j Rectangle b centered at point j The intersection-union ratio with rectangle s; when g i ≥θ,t j When ≥θ, then rectangle a i b j Construct a set G = {g1, g2, ..., g...} containing the cyclone center point. k}, T={t1,t2...t l}, G and T respectively contain the ratios of the intersection-union ratios of elements in sets A and B with rectangle s that are not lower than the threshold;

[0130] Select the point p corresponding to the element with the largest value in sets G and T. u and q v p u The pressure value at a point represents the pressure at the center of the cyclone. (Iterating through q...) v The wind speed values ​​within a 5°×5° radius are used to determine the cyclone intensity. By linking the cyclone intensity and trajectory through the cyclone center point, the physical and spatial distribution can be correlated.

[0131] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0132] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0133] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0134] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0135] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0136] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0137] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0138] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0139] References

[0140] [1]Lee T L.Neural network prediction of a storm surge[J].OceanEngineering,2006,33(3-4):483-494.

[0141] [2]Chen K, Kuang C, Wang L, et al. Storm surge prediction based on longshort-term memory neural network in the East China Sea[J]. Applied Sciences, 2022, 12(1):181.

[0142] [3]Adeli E, Sun L, Wang J, et al. An advanced spatio-temporalconvolutional recurrent neural network for storm surge predictions[J]. arXivpreprint arXiv:2204.09501,2022.

[0143] [4]Kohno N,Dube SK,Entel M,et al.Recent progress in storm surgeforecasting[J].Tropical Cyclone Research and Review,2018,7(2):128-139.

[0144] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A storm surge early warning method driven by data and knowledge, characterized in that, The method includes: 1) Use cross-task attention mechanism to update the features of the separated cyclone intensity-specific and common feature sets and cyclone trajectory-specific and common feature sets to obtain the predicted feature sets of cyclone intensity and cyclone trajectory; 2) Convert the true values ​​of the cyclone center coordinates in the cyclone dataset into a probability distribution map within a preset latitude and longitude range, and use the probability distribution map after the conversion of the true values ​​of the cyclone center coordinates as the new label; 3) Using the transformed probability distribution map as labels, decode the cyclone trajectory prediction feature set and output the probability distribution prediction map of the cyclone center position within a preset latitude and longitude range at different time steps; decode the cyclone intensity prediction feature set and output gridded pressure distribution and wind speed distribution. 4) Traverse the probability distribution prediction map of the cyclone center location and find the point with the highest probability value as the cyclone center point. With the combined guidance of the predicted cyclone center location and cyclone-related knowledge, correlate the cyclone pressure and wind speed distribution to obtain the predicted cyclone intensity value. 5) Input the predicted cyclone intensity, cyclone trajectory information, and the jointly input storm surge level harmonic analysis data into the long short-term memory cyclic network to predict the storm surge intensity at different time steps.

2. The storm surge early warning method driven by data and knowledge according to claim 1, characterized in that, Before step 1), the method further includes: By mapping the spatiotemporal feature sets of multiple multi-source environmental fields to different feature spaces, we can obtain the unique feature sets of cyclone intensity and cyclone trajectory, and then map them to the same feature space to obtain the common feature set of cyclone intensity and cyclone trajectory. We can then use adversarial domain adaptation to reduce the differences between the common feature sets.

3. The storm surge early warning method driven by data and knowledge according to claim 1, characterized in that, The specific steps of using a cross-task attention mechanism to update the features of the separated sets of unique and common features of cyclone intensity and the sets of unique and common features of cyclone trajectory are as follows: For cyclone intensity prediction tasks, the common feature V1 after distribution alignment is utilized. ′ V2 ′ Calculate the attention weights, and use the resulting weight matrix to update the trajectory-specific feature V2 to obtain supplementary information that is highly relevant to the cyclone intensity task. The common characterization of cyclone intensity V1 ′ V2 is a common feature of cyclone trajectories ′ Calculate attention weights: Where, q s V1 is the common characterization set of cyclone intensity. ′ The query, k, obtained by linear projection s V2 is the common representation set of cyclone trajectories ′ key, d obtained by linear projection m For the feature dimension, the obtained attention weights Attn ti With v ti Feature updates are performed through multiplication, where v ti The value is obtained by linear projection of V2, a set of unique representations of cyclone trajectories. Integrate the intensity-specific representation set V1 with the cross-task coupled feature set To splice: Trajectory Common Representation V2 ′ Characterized by shared intensity V1 ′ Calculate attention weights: Where, q t V2 is the common representation set of cyclone trajectories ′ The query, k, obtained by linear projection t V1 is the common characterization set of cyclone intensity. ′ The key obtained through linear projection will be used to acquire the attention weights Attn. traj With v traj Feature updates are performed through multiplication, where v traj The value is obtained by linear projection of V1, a set of characterizations specific to cyclone intensity. Combine the trajectory-specific representation set V2 with the cross-task coupled feature set Then, the parts are assembled.

4. The storm surge early warning method driven by data and knowledge according to claim 1, characterized in that, The probability distribution prediction map of the cyclone center location is traversed, and the point with the highest probability value is identified as the cyclone center point. The predicted cyclone center location and related cyclone knowledge are used to guide the correlation between cyclone pressure and wind speed distributions to obtain the predicted cyclone intensity value: Let the set of points be P = {p1, p2, ..., pn}. m }, Q = {q1, q2...q n }, where p i ∈P, where P is a point in the pressure distribution grid with a lower pressure than its surroundings, and m is the total number of local pressure minimum points traversed; q j ∈Q, is a point in the wind speed distribution grid where the wind speed is lower than the surrounding area, j≤n, where n is the total number of local minimum wind speed points traversed; Construct 5° × 5° rectangles centered on elements of set P and set Q respectively, forming set A = {a1, a2, ..., a3}. m }, B={b1,b2...b n };b j ∈B, for q j A rectangle centered at point j≤n; a rectangle of the same size centered at the predicted cyclone center position, denoted as s, is constructed, and the ratios of the intersection and union of rectangle elements in sets A and B with rectangle s are calculated respectively. Among them, g i For set A, p i Rectangle a centered at point i The intersection-to-union ratio with rectangle s, t j For set B, q j Rectangle b centered at point j The intersection-union ratio with rectangle s; when g i ≥θ,t j When ≥θ, then rectangle a i b j Construct a set G = {g1, g2, ..., g...} containing the cyclone center point. k }, T={t1,t2...t l }, G and T respectively contain the ratios of the intersection-union ratios of elements in sets A and B with rectangle s that are not lower than the threshold; Select the point p corresponding to the element with the largest value in sets G and T. u and q v p u The pressure value at a point represents the pressure at the center of the cyclone. (Iterating through q...) v The wind speed values ​​within a 5°×5° radius are used to determine the cyclone intensity. The cyclone intensity and trajectory are then correlated through the cyclone center point to establish a connection between physical and spatial distribution.

5. A data and knowledge co-driven storm surge early warning device, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method steps of any one of claims 1-4.

Images