Meteorological early warning publishing method based on satellite fusion with Internet of Things

Abstract

The invention relates to the technical field of meteorological monitoring, in particular to a satellite fusion internet-of-things meteorological early warning issuing method, which comprises the following steps of: firstly, acquiring data of a stationary satellite, a low-orbit satellite, millimeter wave communication, ground optics and an electric field, and calibrating and registering to form a unified observation set; a multi-modal diagram is constructed, the optimal transmission mapping meeting mass conservation is obtained through quantum annealing, and three-dimensional wind speed, specific humidity and temperature fields are reconstructed to generate potential energy tensors; establishing a potential energy function based on the potential energy tensor, and judging convection nucleation through Hamiltonian Monte Carlo sampling minimum action path; a coring event is packaged into an early warning message, a signature and a delay function timestamp are added, then the early warning message is written into a right and interest proof block chain, a named data network controlled by reinforcement learning is used for self-adaptive copying and distribution, and meanwhile, an action quantity threshold value and a regularization coefficient are dynamically adjusted. According to the invention, minute-level advance, second-level issuing and whole-process traceable meteorological early warning are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological monitoring, and particularly relates to a satellite fusion Internet of Things meteorological warning publishing method. BACKGROUND

[0002] The disastrous weather such as short-time heavy rain, strong convective wind, hail and the like has the characteristics of suddenness and high influence on urban traffic, power dispatching and public safety. The traditional radar-manual station network mode has insufficient coverage for the far sea, mountainous area and radar blind area, and the radar-short message link full-process time efficiency is usually greater than 10 minutes, which is difficult to meet the demand of the'minute-level' warning of the wisdom city, space-air-ground integrated disaster prevention. The wide coverage ability of the stationary-low earth orbit satellite and the high resolution blind filling ability of the ground Internet of Things are fused, and then the data driven network publishing is carried out, which can significantly shorten the warning link and improve the spatial resolution, and has a revolutionary significance for the industry. SUMMARY

[0003] In view of the problems existing in the prior art, the satellite fusion Internet of Things meteorological warning publishing method is provided. After the satellite-Internet of Things observation is fused, the global quality flow is solved on the multi-modal graph by using quantum annealing, and then the three-dimensional wind-heat-humidity field with physical non-divergence is obtained by taking the flow as a source term and explicitly iterating the mass conservation equation. Then, the potential energy function is constructed, the lifting kernel is quickly captured by using Hamilton Monte Carlo sampling minimum action path, the blockchain trusted timestamp warning is triggered, and the data is issued at the second level through the adaptive named data network.

[0004] A satellite fusion Internet of Things meteorological warning publishing method, comprising:

[0005] Collecting remote sensing data of stationary satellites or low earth orbit satellites, ground Internet of Things meteorological electric field data, millimeter wave communication attenuation data and ground optical imaging cloud data, performing radiation calibration, space-time registration and integrity processing to form a unified observation data set;

[0006] Based on the unified observation data set, a weighted multi-modal graph is constructed, the source density and the sink density are discretized into a binary optimization model, the quantum annealing algorithm is adopted to obtain the optimal transmission mapping, the three-dimensional wind speed field, the specific humidity field and the temperature field are reconstructed according to the mass conservation equation, and the fusion atmospheric state field and the potential energy tensor are generated;

[0007] The space-time height potential energy function is constructed by taking the fusion atmospheric state field and the potential energy tensor as the input, the candidate air particle path is variational sampled and the action is calculated, when the action is lower than the threshold value and the potential energy decreases, the minimum action path is screened, and the convection kernel event object is generated;

[0008] The convection nucleation event object is encapsulated as an early warning message, a named data name is generated, a digital signature and a verifiable delay function timestamp are written into a blockchain after completion, the early warning message is distributed in the named data network according to a reinforcement learning replication strategy, and the replication level, the action threshold and the regularization coefficient are adjusted in real time according to the network delay and the hit rate.

[0009] Preferably, the stationary satellite or low-orbit satellite remote sensing data is composed of infrared brightness temperature data and lightning imaging data, the infrared brightness temperature data is used to extract the cloud top brightness temperature field, and the lightning imaging data is used to locate the instantaneous discharge activity to supplement the convection trigger information.

[0010] Preferably, the radiation calibration is established by matching the radiation transmission model output on the clear sky pixel, the space-time registration is completed by the exterior orientation correction of the orbit attitude data and the optical interpolation method, and the integrity processing is completed by establishing the Gaussian process prior to spatial inference of the missing pixels.

[0011] Preferably, the observation nodes of the weighted multi-modal graph are composed of satellite pixel nodes, millimeter wave precipitation rate slice nodes and ground cloud body optical flow nodes, the node feature vectors are spliced by brightness temperature, precipitation rate, optical flow velocity and electric field amplitude, and the edge weight is determined according to the exponential decay function of the Euclidean distance of the node feature vectors.

[0012] Preferably, the source density and the sink density are discretized as a binary variable optimization model, and a quantum annealing processor is used to embed the optimization model into a quantum bit network, and the optimal transport mapping is obtained by single pulse annealing and returned to the unique free energy minimum solution.

[0013] Preferably, the mass flow of the optimal transport mapping is used as the source term, the mass conservation partial differential equation is solved by the finite difference method, and the divergence-free constraint is maintained in the iteration process, and the three-dimensional wind speed field, specific humidity field and temperature field are obtained.

[0014] Preferably, the space-time height potential energy function is composed of specific humidity, isentropic potential temperature, millimeter wave precipitation rate and potential energy tensor through linear weighted combination, the specific humidity and isentropic potential temperature are used to describe the thermal state, the millimeter wave precipitation rate is used to describe the liquid water content, and the potential energy tensor is used to describe the kinetic energy and potential energy distribution.

[0015] Preferably, the Hamilton Monte Carlo algorithm is used for variational sampling of candidate air particle paths, the symplectic integral is used to keep the phase space volume unchanged, the minimum action is used as the path screening criterion, and a single minimum action path is output.

[0016] Preferably, the early warning message is embedded in a verifiable delay function timestamp after elliptic curve digital signature, and is written into a blockchain maintained by a proof-of-stake consensus mechanism to ensure the integrity of the data chain.

[0017] Preferably, the cache replication strategy of the named data network is determined by a reinforcement learning model trained with network latency and hit rate as reward functions to determine replication levels in real time, and the action threshold and the regularization coefficient of the optimal transmission mapping are updated synchronously to maintain the balance between the overall warning timeliness and accuracy after the replication level changes.

[0018] Compared with the prior art, the application has the advantages and beneficial effects that:

[0019] By solving the binary optimal transmission mapping through quantum annealing, global mass conservation coupling between multi-modal observations is realized, and the local optimization and energy imbalance problems of traditional heuristic fusion are overcome.

[0020] By reconstructing the three-dimensional wind-heat-humidity field through divergence-free projection finite difference, the dynamic field is updated at a second level, and the lifting channel is captured in advance compared with the radar interpolation method.

[0021] By Hamiltonian Monte Carlo variation sampling and minimum action screening, early identification of convective nucleation is realized, and the warning lead time is improved and the false alarm rate is reduced.

[0022] By elliptic curve signature + verifiable delay function + proof of stake chain storage, the integrity, time credibility and traceability of the warning information are realized, and the problem of difficult evidence of centralized push is solved.

[0023] By the reinforcement learning driven named data network replication strategy, the replication level, action threshold and regularization coefficient are adaptively adjusted according to the network load, and the balance between warning timeliness and bandwidth utilization is maintained. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The flowchart of the method of the application is shown in the figure;

[0025] Figure 2 The weighted multi-modal graph construction schematic diagram in the application is shown in the figure. DETAILED DESCRIPTION

[0026] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, many specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments can be implemented without these specific details. Furthermore, in the following description, descriptions of well-known structures and techniques are omitted to avoid unnecessarily obscuring the concept of the present disclosure.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used herein, the terms "comprises", "comprising", "includes", "including" and the like are specifically intended to be open-ended and to mean that other features, steps, operations, and / or components can be added.

[0028] All terms used herein, including technical and scientific terms, have the meanings commonly understood by one of ordinary skill in the art unless otherwise defined herein. It should be noted that the terms used herein are defined as having meanings that are consistent with the context of the specification in which the terms are used and should not be interpreted in an overly idealized or overly formal manner.

[0029] As shown in Figure 1 A satellite fusion Internet of Things meteorological warning release method, comprising:

[0030] Collecting stationary satellite or low-orbit satellite remote sensing data, ground Internet of Things meteorological electric field data, millimeter wave communication attenuation data, and ground optical imaging cloud data, performing radiation calibration, space-time registration, and integrity processing to form a unified observation data set;

[0031] In the first step of the satellite fusion Internet of Things meteorological warning release method, the core goal is to gather five types of heterogeneous data, including stationary satellite remote sensing, low-orbit satellite remote sensing, ground Internet of Things meteorological and electric field observation, millimeter wave communication link attenuation observation, and ground optical imaging cloud observation, into a unified observation data set, so that subsequent algorithms can be coupled and operated in the same space-time reference system. The principle of this step can be decomposed into three sub-links: radiation calibration, space-time registration, and integrity processing, which collectively solve three typical fusion problems: "scale difference", "time misalignment", and "data missing".

[0032] The radiation calibration link takes the satellite sensor calibration coefficient as the reference to convert the original digital quantization value into a radiation brightness with physical meaning, and further into a brightness temperature. For the infrared channel, the relationship between brightness temperature T and radiation brightness L can be obtained by using the inverse of Planck's equation:

[0033]

[0034] Where c1 and c2 are radiation constants, λ is the channel center wavelength, and L is the calibrated radiation brightness. By using the same formula to correct the infrared sensors of geostationary satellites and sun-synchronous satellites, the consistency of brightness temperatures of different platforms can be ensured, thereby providing comparable dimensions for subsequent multi-modal graph construction. The millimeter wave communication link uses the difference between the actual path loss and the clear sky path loss to retrieve the path precipitation rate; the low-orbit satellite occultation humidity profile is converted into a specific humidity field through the conventional occultation inversion process; and the ground optical imaging adopts the optical flow calculation of multiple frame subtraction to calculate the cloud plane velocity.

[0035] The space-time registration relies on a unified latitude-longitude grid and absolute world time coordinates to map multi-platform data to a regular grid with a resolution of 0.01 degrees. The orbit position error is corrected by matching the satellite subsatellite point trajectory with the ground control points, the millimeter wave link attenuation slice is determined according to the communication antenna beam center point via spherical projection, and the optical flow result is bound to the fixed pose of the camera and converted to the ground latitude-longitude plane through a perspective matrix. All observation times are taken as a unified reference based on the global positioning system second, and the data of different refresh periods are resampled to one-second steps through linear interpolation.

[0036] The integrity processing adopts statistical modeling and keeping null for random missing and systematic occlusion respectively. For random missing pixels, Gaussian process spatial inference is used to fill in the brightness temperature, precipitation rate and optical flow velocity; for systematic occlusion (for example, the communication blind area behind the mountain), null is kept for subsequent coupling algorithms to utilize the conservation regular automatic smoothing. Thus, the unified observation dataset is continuous in key areas and does not introduce artificial false signals in truly no-information areas.

[0037] In the application layer, the unified observation dataset has three effects: first, multi-source data are converted to the same physical dimension, eliminating the coupling error caused by inconsistent dimensions; second, the space-time registration ensures that each grid cell has the closest multi-dimensional observation at the closest real time, so that the subsequent optimal transport mapping based on quantum annealing can accurately project the source-sink relationship to the three-dimensional wind field; third, the integrity processing reserves the real missing boundary for the coupling algorithm, preventing the energy conservation term from being destroyed due to excessive interpolation.

[0038] For example, in a summer afternoon strong convective case in North China, the cloud top temperature observed by the geostationary satellite infrared brightness temperature decreases continuously, the millimeter wave link appears path attenuation increase, and the occultation humidity profile gives the low tropospheric water vapor increase. After the above steps, these complementary information in the unified observation dataset form a space-time overlap, which points to a minimum action path in the subsequent potential function calculation, triggering the nucleation event ten minutes in advance, and finally improving the early warning time by about five minutes compared with the traditional early warning relying on a single radar.

[0039] Preferably, the geostationary satellite or low-orbit satellite remote sensing data is composed of infrared brightness temperature data and lightning imaging data, the infrared brightness temperature data is used to extract the cloud top brightness temperature field, and the lightning imaging data is used to locate the instantaneous discharge activity to supplement the convective trigger information.

[0040] The geostationary satellite remote sensing and low-orbit satellite remote sensing provide complementary information from different orbital altitudes and observation geometries for convective monitoring. The present application preferentially selects infrared brightness temperature data and lightning imaging data as two types of key observations on the satellite side. After radiation calibration and atmospheric correction, the infrared brightness temperature data can be used to retrieve the cloud top radiation brightness, and then the Planck inversion formula is used:

[0041]

[0042] The cloud top brightness temperature field is obtained, where T represents the brightness temperature, L represents the radiance, c1 and c2 are radiation constants, and λ is the central wavelength of the channel. The role of the brightness temperature field in the present application is to identify the deep convection vertical development stage through the cloud top temperature gradient. When the brightness temperature rapidly decreases in consecutive time, it is considered as a potential convection enhancement signal.

[0043] The lightning imaging data is derived from an optical lightning detector carried by a satellite, which can capture the position and energy of a discharge pulse on the order of 0.001 seconds. The present application aggregates the positioning results of a single pulse to generate a transient discharge activity distribution grid. After superimposing the distribution grid and the cloud top brightness temperature field at the same time, pure cold clouds and convection cloud cores carrying charges can be distinguished, reducing the risk of false positives identified by cold top alone.

[0044] In the unified observation data set, the infrared brightness temperature field is stored as a cloud top brightness temperature grid, and the lightning imaging is stored as a pulse count grid. Both are mapped to the same latitude and longitude grid in the space-time registration stage, so that the weighted multi-modal graph can combine the "deep and cold" and "with lightning" two characteristics into a single node feature vector component, thereby improving the accuracy of the optimal transport mapping in describing the real convection source area.

[0045] Embodiment: On July 23, 2024, at about 16:00 Beijing time, a cluster of pixels with a cloud top brightness temperature that decreased by more than 8 Kelvin for 3 minutes appeared near 115 °E, and lightning imaging data gave a dense pulse in the same area. According to the process of the present application, the brightness temperature drop rate is used as a thermal enhancement indicator, and the lightning pulse density is used as an electrical activity indicator. They are encoded into node features and obtain a higher weight in quantum annealing mapping, and finally form a minimum action path across 3 height layers, which is about 5 minutes earlier than the traditional algorithm using only brightness temperature criteria to output the convection nucleation event. This example shows that the fusion of brightness temperature and lightning information can effectively improve the detection rate of weak early discharge clouds and reduce false alarms caused by ice phase cooling in high clouds.

[0046] Through the above design, the infrared brightness temperature data provides a thermal perspective of the vertical development of the cloud, and the lightning imaging data supplements the electrical perspective, both of which provide multi-dimensional priori for the subsequent potential function construction, achieving more reliable trigger identification in the early stage of convection, and thus laying a solid foundation for the issuance of warning messages.

[0047] Preferably, the radiation calibration is established by matching the output of the radiation transfer model on clear sky pixels to establish a brightness temperature baseline, the space-time registration is completed by correcting the exterior orientation angle of the orbit attitude data and combining the optical interpolation method, and the integrity processing is completed by establishing a Gaussian process priori to spatially infer the missing pixels.

[0048] Radiometric calibration, spatio-temporal registration and integrity processing jointly undertake the pre-task of transforming heterogeneous observations into co-computable physical fields. The reason why the present invention emphasizes establishing a brightness temperature baseline in clear-sky pixels is that the gain of infrared detectors of geostationary or low-orbit satellites will drift over time. If the static calibration coefficient given by the manufacturer is directly used, the brightness temperature field will have a systematic deviation when compared across seasons or across satellite platforms. The baseline construction process first extracts clear-sky pixels using an automatic cloud detection algorithm, and then calls a radiative transfer model to calculate the theoretical radiance L RTM . A single-point linear regression is performed on the measured radiance L obs and L RTM to obtain the gain g and the bias b, and then L cal = gL obs +b is used to correct the entire image. In the formula, L cal is the corrected radiance, g is the slope correction coefficient, and b is the intercept correction coefficient. By updating g and b for each orbit, the brightness temperature of different dates and different satellites can be directly compared. The existence of this baseline ensures that the source-sink matching in the subsequent optimal transport mapping is not affected by instrument drift, thereby improving the consistency of wind field retrieval.

[0049] The spatio-temporal registration process handles the misalignment of cross-orbit observations in terms of viewing angle and time. The orbital attitude metadata is provided in the original remote sensing product in the form of quaternions. The present invention maps the subsatellite point trajectory to the Earth reference ellipsoid by correcting the exterior orientation angle. In the spatial dimension, the line-of-sight vector of each pixel is first calculated, and then projected to a regular grid with a resolution of 0.01 degrees using bilinear interpolation; in the time dimension, the data of different refresh periods are linearly resampled to a 1-second time step using the Global Positioning System second as the absolute reference. After adopting a unified grid and a unified time axis, the satellite brightness temperature grid, the millimeter wave precipitation rate slice, and the optical flow plane velocity field can form one-to-one corresponding components in the node feature vector, providing synchronous input for the subsequent edge weight calculation of the weighted multi-modal graph.

[0050] The integrity processing adopts different strategies for random missing and systematic occlusion. Random missing indicates situations such as temporary saturation of brightness temperature sensors and instantaneous interruption of millimeter wave links, and such missing values are filled by Gaussian process spatial inference. The Gaussian process uses a kernel function:

[0051]

[0052] where x i and x j are the longitude and latitude coordinate vectors, is the prior variance, and l is the correlation scale. The mean of the inference result is taken as the missing pixel filling value, and the variance is used as the observation uncertainty in subsequent quantum annealing weight allocation; the millimeter wave blind area caused by systematic shielding such as mountain shadow remains empty, and is automatically smoothed by the conservation constraint through the coupling algorithm. This reduces the fitting of real information-free areas and avoids energy leakage caused by full-image mean filling.

[0053] Effectively, the baseline drift of the radiometric calibrated brightness temperature is less than 0.5 Kelvin, ensuring that the cross-satellite data can be superimposed; the space-time registration aligns all observations within 1 second, reducing the risk of time difference shear of dynamic convection structures; the integrity processing reduces the proportion of random missing pixels from 4% to 0.3%, and the retention of system null values makes the conservation of mass regularization term have actual traction.

[0054] The embodiment shows that in the convection case in North China on July 23, 2024, the above processing makes the minimum action path stably present three-dimensional spiral rising, and the unprocessed data causes the path to be broken due to brightness temperature drift and time misalignment, with an error of up to 4 minutes. The present application establishes a high-credibility, multi-dimensional consistent input basis for quantum annealing coupling and variational path screening through accurate calibration, registration and missing data inference, significantly enhancing the accuracy and advance of early warning.

[0055] Based on the unified observation data set, a weighted multi-modal graph is constructed, the source density and sink density are discretized into a binary optimization model, the quantum annealing algorithm is used to obtain the optimal transport mapping, and the three-dimensional wind speed field, specific humidity field and temperature field are reconstructed according to the mass conservation equation, to generate the fusion atmospheric state field and potential energy tensor;

[0056] After completing the unified observation data set, the present application uses graph theory and optimal transport theory to construct a multi-modal observation graph, and uses quantum annealing to solve the cross-modal mass mapping. First, an observation node is placed at the center of each grid of the latitude and longitude grid. The infrared brightness temperature of the stationary satellite, the millimeter wave precipitation rate, the ground flow horizontal velocity and the electric field strength each form a feature vector component, forming a multi-dimensional node feature. Each pair of nodes with a spatial distance of less than 30 kilometers is connected, and the weight is set as:

[0057]

[0058] where f i and f j are node features, and σ is a prior scale. The weight reflects the observation similarity, and the greater the weight, the more likely the node is located in the same air mass. Then, the brightness temperature-flow combination is defined as the source density, and the precipitation rate-electric field combination is defined as the sink density according to the observation type to which the node belongs.

[0059] To map the continuous density coupling problem to quantum hardware, the invention employs threshold segmentation to discretize the node density into quaternary, and then encodes the flux decision variable x in binary ij . The optimal transport objective is written in binary optimization form:

[0060]

[0061] where c ij = d ij / w ij is the cost, d ij is the Euclidean distance between nodes, s i and t j are the discrete weights of the source and sink densities, respectively. The quadratic and linear terms are packed into an Ising matrix, which is then embedded into a quantum annealing chip. Quantum annealing searches for the global ground state in a millisecond pulse and outputs the unique mapping solution with the minimum free energy.

[0062] After mapping, x ij is the mass flux, which is substituted into the mass conservation equation under the incompressibility assumption:

[0063]

[0064] where u is the three-dimensional wind velocity vector, and S is the divergence of the mass flux within the grid. The continuous wind velocity field is obtained by solving the finite difference. The cloud top height and liquid water content retrieved from satellite brightness temperature and precipitation rate are used to implement the thermodynamic balance interpolation of specific humidity and temperature, forming the fused atmospheric state field. The potential energy tensor is calculated as:

[0065]

[0066] where g is the acceleration of gravity, and z is the terrain height.

[0067] The embodiment shows that in a sea breeze front triggered urban rainstorm process, the optimal mapping obtained by quantum annealing correctly matches the large specific humidity source flow of the sea breeze front nodes to the lightning-rain density of the inland updraft nodes, and the reconstructed wind speed field shows low-level convergence and upward motion along the coast to the urban area. Compared with the traditional linear interpolation wind analysis, the vertical velocity peak value of the fused field in the thunderstorm core area is 15% higher, and the low-level convergence width is 20% more consistent with radar echo observations. Therefore, the multi-modal graph combined with quantum annealing can more accurately restore the mass and energy transport structure before convective triggering, provide more realistic dynamic conditions for subsequent action path screening, and ultimately improve the advance amount and reliability of convective nucleation determination.

[0068] Preferably, as Figure 2As shown in the figure, the observation nodes of the weighted multi-modal graph are composed of satellite pixel nodes, millimeter wave precipitation rate slice nodes and ground cloud body optical flow nodes, the node feature vector is spliced by brightness temperature, precipitation rate, optical flow speed and electric field amplitude, and the edge weight is determined according to the exponential decay function of the Euclidean distance of the node feature vector.

[0069] After the preparation of the unified observation data set is completed, the application adopts a graph theory method to structurally express the multi-source, multi-scale and heterogeneous observation, so as to support the subsequent quantum annealing optimal transmission mapping. The core method is to map each latitude and longitude grid element to an observation node, map the similarity between nodes to an edge weight, and then form a weighted multi-modal graph.

[0070] Node construction. The node type is divided into three categories. The first category is the satellite pixel node, the data source is the infrared brightness temperature field of the stationary satellite or the low-orbit satellite, each node stores the cloud top brightness temperature and the field of view angle information. The second category is the millimeter wave precipitation rate slice node, the data source is the path attenuation inversion result uploaded by the communication base station, which is mapped to the ground grid through beam geometry projection and records the average precipitation rate along the ray path. The third category is the ground cloud body optical flow node, the data source is the continuous shooting of the skyline cloud body by the fixed camera, the plane velocity vector is obtained through the optical flow algorithm, and then projected to the latitude and longitude plane. The three types of nodes are distinguished by different marks when combined at the same grid center, ensuring the complete preservation of multi-source information.

[0071] Feature splicing. The application defines the node feature vector to be spliced in sequence by four components: brightness temperature, precipitation rate, optical flow speed module length and electric field amplitude. The brightness temperature describes the thermal state of the cloud top, the precipitation rate corresponds to the liquid water content, the optical flow speed reflects the plane motion intensity of the cloud body, and the electric field amplitude reflects the degree of electrical instability. Each component is first normalized to zero mean, and then spliced into a fixed length vector, so that the data of different dimensions has comparability when measured in Euclidean space. The advantage of this splicing design is that it can adaptively reflect the comprehensive influence of each component through distance measurement without pre-setting manual weights, thereby providing a unified input for the cost function in the quantum annealing stage.

[0072] Edge weight design. In order to measure the possibility of two nodes belonging to the same air mass, the application adopts an exponential decay function to set the edge weight. The core formula is:

[0073]

[0074] Where w ij represents the edge weight of node i and node j, f i and f jwherein denotes the corresponding node feature vector, and is a prior scale parameter controlling the similarity decay rate. In the equation, the Euclidean distance term characterizes the feature difference, and the exponential decay ensures that the larger the distance is, the smaller the weight is, which is consistent with the spatial correlation law that the meteorological field decreases with distance. To improve sparsity and computational efficiency, only node pairs with a spatial distance less than 30 kilometers and a weight greater than zero point one are retained, and the average degree of the graph is controlled within 20% of the number of parent nodes.

[0075] Model discretization. In the optimal transport modeling link, it is necessary to map the source density and sink density into the flow decision variable. The present application uses four-bit binary coding to characterize the node quality contribution, and sets a binary flow variable for each retained edge. This not only meets the requirement of binary variables for quantum annealing hardware, but also maintains the expressiveness of the mass conservation constraint. The optimization objective consists of two parts: edge cost and conservation penalty. The edge cost is the product of the cost coefficient and the flow variable, and the penalty term enforces source-sink balance through the square form. All coefficients and variables are packaged into an Ising matrix and mapped to a quantum annealing chip to realize global search for low-energy states.

[0076] Physical field reconstruction. After the optimal transport mapping output, the present application writes the flow as a discrete source term into the mass conservation equation, and solves the three-dimensional wind speed, specific humidity and temperature field by finite difference method. Among them, the wind speed calculation adopts explicit Jacobi iteration to ensure convergence; specific humidity and temperature are interpolated by energy balance after the wind speed field is determined, combined with satellite brightness temperature and millimeter wave precipitation rate. Finally, the potential energy tensor is defined according to the kinetic energy and potential energy, which is an important input for the trigger criterion of convection.

[0077] Effect analysis. By introducing multi-modal nodes and adaptive edge weights, the present application enables the quantum annealing cost function to consider thermal, dynamic and electrical information simultaneously, providing a more physically consistent reference direction for flow matching. Comparative experiments show that, under the same quantum annealing time budget, when using the two-component brightness temperature and precipitation rate to construct the graph, the coherent energy of the reconstructed wind speed field in the three-dimensional energy spectrum is about one and a half times lower than that of the complete four-component model, and after using the complete four-component model, the path selected by the minimum action is more continuous, and the average trigger time is three minutes earlier.

[0078] Embodiment. On July 23, 2024, a typical afternoon convection in North China occurred in the southern suburb of Beijing. After constructing the graph using the method of the present application, the edge weights between multiple nodes in the sea breeze front edge are concentrated above 0.9, indicating that the brightness temperature-precipitation-rate-light flow comprehensive characteristics are similar, with rapid temperature drop and frequent lightning. The flow deduced by quantum annealing matches the sea breeze front source node to the inland high electric field node, reflecting the real air mass transport path. The reconstructed three-dimensional wind speed appears a significant convergence area at the eight hundred meter height layer, which is consistent with the radar reflectivity observation, verifying the accuracy of graph construction and mapping solution.

[0079] Preferably, the source density and sink density are discretized into a binary variable optimization model, and the optimization model is embedded into a quantum bit network using a quantum annealing processor, and the optimal transport mapping is obtained by single pulse annealing to return the unique free energy minimum solution.

[0080] After the weighted multi-modal graph is established, the present application uses the optimal transport theory as the framework to express the process of transporting mass from the source node to the sink node as a flow distribution problem with minimum energy. The quantum annealing processor can only process binary spin variables, so first, a binary flow variable x ij is introduced for each reserved edge. ij When x ij = 0, it means that no mass is transported, and when x i = 1, it means that a unit of mass is transported. The source node demand is represented by a discrete weight s j , and the sink node demand is represented by a discrete weight t ij , and the weight comes from the normalized linear combination of the brightness temperature, precipitation rate, electric field amplitude and optical flow velocity in the node. For edge (i, j), the cost coefficient c a is defined as the weighted sum of the spatial distance and the characteristic difference of the nodes, which is used to measure the kinetic energy and thermal consumption required to transport the path.

[0081] Based on the above quantification, the objective energy of the binary optimization model is constructed:

[0082]

[0083] Where λ is the conservation constraint weight; the first term is the total transportation cost; the second and third terms respectively constrain the mass conservation of each source node and sink node. After expanding the objective energy in the Ising matrix form, it is mapped to the quantum bit network of the quantum annealing processor. When embedding, the chain strength adjustment strategy is used to split the logical variables into physical quantum bit chains, and the spin consistency is ensured by chain coupling.

[0084] The quantum annealing execution process uses a single pulse annealing scheme: at the start point t ij , the transverse field is loaded, and then the transverse field is exponentially decayed and the Ising energy term is simultaneously increased, and the annealing ends to obtain a set of candidate spin states. The processor calculates the energy E of each spin state in real time, and the spin state with the minimum energy is the unique free energy minimum solution. Since the mass conservation constraint has been explicitly added to the objective energy, the minimum solution naturally satisfies the source flow and sink flow balance, thereby forming the optimal transport mapping.

[0085] When the optimal transport mapping is output, the edges with x ij = 1 correspond to the mass flow, which is considered as the source term of the three-dimensional mass conservation equation:

[0086]

[0087]

[0088] where u is the wind velocity field, r ij is the edge center coordinate. The wind velocity field is solved by explicit Jacobi iteration; then, combined with the satellite brightness temperature and the millimeter wave precipitation rate, the specific humidity field and the temperature field are obtained by energy balance interpolation; and the three fields are superimposed to form the fused atmospheric state field. The potential energy tensor is defined by linearly adding the kinetic energy term and the potential energy term gz (g represents the gravitational acceleration, and z represents the terrain height), providing a potential field basis for subsequent action path screening.

[0089] Embodiment: In a coastal city rainstorm process, the quantum annealing model contains about 2500 nodes and about 20000 edges. The single pulse annealing time is set to 1 millisecond. After the annealing is completed, the minimum energy solution obtained activates about 30% of the edges of the sea breeze front to the lightning dense area in the inland. The wind speed field reconstructed by using the flow generates a clear convergence line near the ground, and the position is consistent with the weather radar reflectivity observation. Compared with the traditional uniform speed interpolation reconstruction, the maximum vertical velocity is increased by about 15%, and the convection nucleation event is triggered 5 minutes before the real convection outbreak, which reserves sufficient response time for city disaster prevention.

[0090] Effect analysis shows that the global search of quantum annealing avoids the limitation of classical gradient descent falling into local minimum, and ensures that the reconstructed wind speed field meets the energy minimum principle and the conservation equation at the same time. The unified graph structure coding speeds up the variable embedding, and the single annealing can return the unique free energy minimum solution, reducing the overhead of multiple candidate solution screening. In addition, because the source density and the sink density come from the brightness temperature and the precipitation rate, the physical meaning of the electric field strength is combined, and the optimal mapping is more consistent with the real dynamic process of air mass transport in explanation.

[0091] Example supplement: Assuming that node i is located at 35.0°N, 120.0°E, the brightness temperature is 195 Kelvin, the precipitation rate is 10 millimeters per hour, and the electric field amplitude is 1003 volts per meter, the optical flow velocity module length is 12 meters per second; node j is located at 34.5°N, 120.2°E, and the corresponding quantities are 198 Kelvin, 8 millimeters per hour, 110 volts per meter, and 10 meters per second. After four-component standardization, the node characteristic Euclidean distance is about 0.9, and the edge weight obtained by substituting the edge weight formula is about 0.4. If the node spatial distance is 25 kilometers, the cost coefficient c ij is about 62.5. Quantum annealing preferentially selects edges with smaller cost coefficients and larger weights, ensuring that the transport path takes into account the physical synthesis and energy economy.

[0092] Preferably, the mass flow with optimal transport mapping is used as the source term, the mass conservation partial differential equation is solved by the finite difference method, and the divergence-free constraint is maintained in the iteration process, to obtain the three-dimensional wind speed field, the specific humidity field and the temperature field.

[0093] In the power field reconstruction link of the application, the optimal transport mapping provides spatial-time discrete quality flow information. The edges with a value of 1 in the mapping result are regarded as real quality transport paths, and a body source term S can be constructed at the center position of the grid.

[0094]

[0095] In the formula, u=(u, v, w) represents a three-dimensional wind speed vector, u, v, and w are east-west, north-south, and vertical velocity components, respectively, and S is the mass generation rate per unit volume. For numerical solution, the application adopts a second-order central difference to discretize the divergence operator on the cubic grid, and performs zero-mean projection after each iteration to ensure that the divergence-free constraint continues to hold. This projection is equivalent to performing a Poisson equation correction, so that the iteration error is not accumulated on the grid.

[0096] After the wind speed field converges, the specific humidity field is obtained by integrating the source node precipitation rate and the given upper and lower boundaries of the cloud top along the wind speed field streamline through the conservative advection-diffusion equation; the temperature field is superimposed on the basis of dry adiabatic decrease to add specific humidity correction to ensure that the columnar total static energy and mass injection energy are balanced. Then the potential energy tensor is calculated:

[0097]

[0098] Where g is the acceleration of gravity, and z is the terrain height. The potential energy tensor contains both kinetic energy and potential energy, and is the potential field source for subsequent action path evaluation.

[0099] The above process has two key principle advantages. First, the source term directly comes from the global optimal solution of quantum annealing, which can better maintain the source-sink structure of the real convection area compared to classical variational assimilation. Second, the divergence-free projection makes the wind speed iteration not produce a pseudo-source, improving the accuracy of mass conservation. In numerical tests, with a grid horizontal resolution of 1 km and a vertical resolution of 250 m, the Jacobi iteration generally converges within 300 steps, and the overall calculation time is 3 seconds, which can meet the minute-level warning link rhythm.

[0100] Specific embodiments: In the area from 115 degrees east longitude to 117 degrees east longitude and from 39 degrees north latitude to 41 degrees north latitude on July 23, 2024, the optimal transport mapping activates 800 source-sink edges. The wind speed field reconstructed by the above method appears a 2.8 m / s convergence line at the 500 m height layer, which is consistent with the results of the weather radar double Doppler inversion; the vertical velocity peak is 1.6 m / s, which is 15% higher than that of the traditional three-dimensional variational assimilation. The specific humidity field reaches 14 g / kg at the same level, and the temperature negative disturbance is 2 K, which together form a high gradient area of potential energy. The action path screening is completed 4 seconds after the wind speed field converges, and the overall process gives a nucleation judgment 5 minutes before the convection outbreak, which is significantly improved compared with the conventional warning scheme which relies on radar echoes. ​

[0101] A space-time potential energy function is constructed by fusing the atmospheric state field and potential energy tensor, a candidate air mass point path is variational sampled and an action is calculated, when the action is lower than a threshold value and the potential energy is reduced, a minimum action path is screened, and a convection nucleation event object is generated;

[0102] The three-dimensional dynamic-thermal field is the basis for subsequent action path screening, and the accuracy of the field depends on whether the discrete mass flow in the optimal transport mapping can be correctly "injected" into the continuous equation set. The idea of the present application is that when solving the mass conservation partial differential equation, the mass flow generated by the optimal transport mapping is placed on the right side of the equation as a volume source term, and the divergence-free constraint is continuously applied in the iteration process, so that the resulting wind speed field fully satisfies the continuity. Compared with the traditional three-dimensional variational or finite volume interpolation method, the advantages of the present application are: on the one hand, the source term is derived from the quantum annealing global optimal solution, which is inherently consistent with the observation-cost balance; on the other hand, the difference-projection iteration explicitly maintains the divergence to be zero, avoiding the generation of false turbulent energy.

[0103] Source term construction. The optimal transport mapping gives whether each edge transports a unit mass in binary form. All edges with x ij =1 are mapped to line segments in the grid coordinate system, and the line segment length is uniformly distributed to the grid centers passed. After accumulation, the source term grid S is obtained. In order to prevent local isolated sources from causing numerical oscillation, the source term is Gaussian smoothed, and the kernel scale is set to twice the grid resolution. This step will not change the total source amount, but only expand the source influence range, ensuring the stability of the Jacobi iteration initial value.

[0104] Conservation equation. The wind speed vector u needs to satisfy where the symbol represents the three-dimensional divergence operator. The equation is discretized by central difference on the cubic grid to obtain the discrete divergence operator The numerical solution selects the explicit Jacobi iteration: update under the step size Δt:

[0105]

[0106]

[0107] where φ n is the difference between the current divergence and the source term, and ɑ is the relaxation coefficient. After each step, the following is obtained:

[0108]

[0109]

[0110] That is, solve a Poisson equation and correct the velocity with a gradient to ensure that tends to zero again. Such iteration is carried out until max|φ n |<1×10-4 The Poisson equation is solved by multigrid, and the grid order is automatically selected according to the resolution to ensure convergence within 5 rounds. The specific humidity and temperature field. The specific humidity field is approximately conserved by the wet convection, and is solved in the fixed wind field:

[0111]

[0112] Take the lightning-rain rate source node as the lower boundary, and the satellite cloud top height as the upper boundary. The profile distribution is obtained by first-order upwind interpolation integration along the streamline; the temperature field is calculated according to T(z) = T surf -Γ d z-βq(z), where T surf is the ground temperature, Γ d is the dry adiabatic rate, and β is the wet correction coefficient. This expression couples the latent heat release caused by the decrease of specific humidity into the temperature field, improving the accuracy of vertical stability estimation.

[0113] Potential energy tensor. After the atmospheric state field is built, the potential energy tensor is generated according to , where g is 9.81 meters per square second. The kinetic energy term u 2 +v 2 +w 2 reflects the local vortex intensity, and the potential energy term gz represents the mass column height. The potential energy tensor decreases with time, indicating that the convective lifting resistance decreases, which is an important input for the minimum action trajectory discrimination.

[0114] Example. In a one-kilometer grid at 36.2 degrees north latitude and 116.9 degrees east longitude, the optimal mapping activates three source streams pointing to the grid, with a total source amount of 3. After Gaussian smoothing, the source term amplitude is 1.2. The initial wind speed of the Jacobi iteration is set to zero field, and after 280 steps of iteration, the absolute value of the divergence is reduced to 8×10-5. The peak vertical velocity is 1.4 meters per second, the specific humidity is 15 grams per kilogram, and the temperature negative perturbation is 2.3 kelvin. Compared with the same time sounding, the temperature error is 0.7 kelvin, and the humidity error is 1.2 grams per kilogram. The inversion profile shows that the maximum upward movement appears at 1.2 kilometers, while the radar echo top reaches 1.1 kilometers, with a deviation of less than the grid layer thickness.

[0115] Benefit analysis. First, the divergence-free projection avoids the accumulated divergence distortion in explicit iteration, and the numerical error does not overflow to the kinetic energy calculation. Second, the source term has a clear physical meaning, making the wind field show a convergent-divergent structure consistent with observations. Third, the temperature and humidity fields are guided by the wind field rather than simply interpolated, so the vertical distribution is more consistent with cloud physics processes. The spatial and temporal high potential energy function calculated based on this field is more sensitive to the capture of potential convective nuclei. In the statistics of 27 spring convective samples, the minimum action triggering time is 4.6 minutes earlier than the radar-based method on average, and the false alarm rate is reduced by 13%.

[0116] Implementation case: On May 4, 2024, from 13:00 to 14:00, local strong convection occurred in northern Jiangsu. The optimal transmission mapping was output at 12:55, the Jacobian iteration completed the wind field at 12:56, the temperature and humidity field at 12:57, the potential energy tensor was generated at 12:57:30 and triggered the minimum action threshold, 5 minutes before the radar echo first reached 35 decibels. After the named data warning was issued, the ground weather station recorded a 11-meter-per-second gust, which was consistent with the warning content.

[0117] Preferably, the space-time high potential energy function is composed of specific humidity, isentropic temperature, millimeter wave precipitation rate and potential energy tensor through linear weighted combination. The specific humidity and isentropic temperature are used to depict the thermal state, the millimeter wave precipitation rate is used to depict the liquid water content, and the potential energy tensor is used to depict the kinetic energy and potential energy distribution.

[0118] In the core algorithm link of the application, the space-time high potential energy function is a bridge connecting the fusion of atmospheric state field and the minimum action path search. The potential energy function takes specific humidity, isentropic temperature, millimeter wave precipitation rate and potential energy tensor as input, and gives a single-point potential value through linear weighting. Its mathematical expression is written as:

[0119] V = aq - bθ e + cR + dΦ

[0120] Wherein V represents the potential energy function value, q represents the specific humidity, θ e represents the isentropic temperature, R represents the liquid water content obtained by millimeter wave precipitation rate inversion, Φ represents the potential energy tensor of kinetic energy and potential energy; a, b, c, d are dimensionless weight coefficients. The initial value of the weight coefficient comes from the Bayesian optimization of the ten-year strong convection sample, and the triggering advance and false alarm rate are adjusted as double targets; in the deployment stage, the model is readjusted according to the sliding window, so that the model has self-adaptability to regional climate differences.

[0121] The specific humidity and isentropic temperature reflect the thermal state. The increase of specific humidity means that the low-level water vapor is abundant, and the decrease of isentropic temperature represents the increase of instability in the troposphere. The combination of the two appears as positive and negative terms, which can adjust the sensitivity of the potential energy threshold under different humidity backgrounds. The millimeter wave precipitation rate depicts the liquid water content, which can quantify the accumulation of water condensate in the air in advance when strong echoes have not yet appeared but communication links have been attenuated. The potential energy tensor integrates kinetic energy and potential energy, directly measuring the dynamic lifting potential of the local air column. When V continuously decreases with time and a significant minimum value appears at the same time, it means that the atmosphere is evolving towards a more favorable convection state, and the minimum action path prefers to pass through this area.

[0122] The variational sampling adopts the Hamilton Monte Carlo algorithm. In the phase space, the potential energy function is used as the potential energy term, and the Hamiltonian is constructed by the air particle velocity and the gradient of the potential energy function, and the symplectic integral is used to keep the volume of the flow unchanged. One hundred candidate paths are extracted for each time slice, and the action of the path is calculated:

[0123]

[0124] where r is the position vector of a particle, u is the wind velocity, is the particle velocity. The particle trajectory with the minimum action and corresponding potential function decreasing is regarded as the precursor channel of convective nucleation. When the minimum action is lower than the threshold value 3.5 and The algorithm outputs the convective nucleation event object, including the geographical longitude and latitude, the predicted triggering time and the confidence.

[0125] For example: on June 12, 2024, at 15:40, the specific humidity in the Beijing south suburb grid rose to 14 grams per kilogram, the isentropic potential temperature dropped by 3 kelvins, and the millimeter wave precipitation rate increased to 8 millimeters per hour. At the same time, the potential energy tensor reached 1800 joules per kilogram at a height of 1.5 kilometers. According to the weight a:b:c:d=2:1:1:1, the V decreases by 12 units in 4 minutes. The Hamilton Monte Carlo sampling gets the minimum action 3.1, which is less than the threshold value three and a half. The algorithm outputs the nucleation event, and the early warning time is 6 minutes. The subsequent radar observed the new strong echo at 15:46, verifying the triggering accuracy.

[0126] Effect evaluation shows that in 30 independent convective cases, the linearly weighted potential energy function makes the false alarm rate of the minimum action judgment less than 0.7 times that of the single wet convective index scheme. The reason is that specific humidity and precipitation rate can capture water vapor transport, isentropic potential temperature can sense layer instability, and potential energy tensor can constrain dynamic lifting. The four quantities make the potential energy function sensitive to various trigger mechanisms. At the same time, because the weight remains linear, the calculation overhead is only 1.2 times that of the traditional criterion superposition, and the full map update can be completed in seconds, meeting the requirements of the minute-level early warning link.

[0127] Preferably, the Hamilton Monte Carlo algorithm is used for variational sampling of the candidate air particle path, the symplectic integral is used to keep the phase space volume unchanged, the minimum action is used as the path screening criterion, and a single action minimum path is output.

[0128] On the basis of the three-dimensional wind speed field, specific humidity field and potential energy tensor provided by the optimal transport mapping, the invention needs to determine which air particles first break through the suppression layer and evolve into convective nuclei. To this end, a space-time-height potential energy function V(r,t) is constructed, and then a variational idea is used to search for the particle trajectory with the fastest potential energy decrease and the smallest dissipation. The particle motion is regarded as a classical Hamiltonian system, and the position vector r and the conjugate momentum p form the phase space coordinates, corresponding to the Hamiltonian:

[0129]

[0130] Where ||p|| represents the momentum modulus, and V(r,t) is the potential energy function value. To maintain the phase space volume of the Hamiltonian system constant, this invention employs a step-drift-step scheme from symplectic integrals. Specifically, within a fixed time step, the momentum is updated half a step first, then the position is updated whole steps, and finally the momentum is updated half a step again. The symplectic scheme ensures that the numerical trajectory maintains Livial invariance under long-term sampling and does not produce energy drift.

[0131] The Hamiltonian Monte Carlo algorithm was used for candidate trajectory sampling. First, 100 particle paths were randomly initialized within each 3D grid block, and each path was assigned a uniformly distributed initial momentum. Then, 25 symplectic integral transitions were performed to obtain the complete trajectory. The action of the trajectory is calculated as follows:

[0132]

[0133] in Let S be the particle velocity, u be the combined wind speed, and t0 and t1 be the start and end times of sampling, respectively. Integration is implemented using Simpson's formula for discretization, with the time interval consistent with the step size of the simplicial integral to ensure the error is controlled within 1%. The algorithm uses S as the energy criterion, filtering out values ​​below a threshold S. crit The set of paths is used; if the set is not empty, the path with the smallest effect is selected as the final convection triggering path. This path records the spatial trajectory, polygon coverage area, and expected eruption time, and is packaged to generate a convection nucleated event object.

[0134] Example. At 15:40 on June 12, 2024, after initializing the air particle path within the grid in the southern suburbs of Beijing, the symplectic integral step size was set to 0.04 seconds, the total number of transitions was 25, and the action threshold was set to 3.5. The algorithm output a minimum action path S = 3.1, with its terminal height reaching 2.2 kilometers, and the corresponding potential function exhibiting the maximum rate of descent at the terminal point. The nucleation event object was generated 6 minutes earlier than the first radar echo exceeding 35 dB. Post-event verification showed that the deviation between the path center and the thunderstorm reporting time from the ground automatic weather station was less than 2 minutes, proving that the trajectory screening rule of this invention can accurately capture early convective upwelling channels.

[0135] Performance Comparison. In 30 spring and summer convective events, compared with the traditional cold top-lightning criterion based on a fixed threshold, the Hamiltonian Monte Carlo-minimum action scheme improved the average lead time by 4.6 minutes and reduced the false alarm rate by 13%. The main reason for the improvement is that the symplectic integral preserves the trajectory dynamics constraints, and the action comprehensively considers the coupling disadvantages of the wind-heat-water fields, making the selected paths more consistent with the actual air mass ascent conditions. The algorithm runs on a single A100 graphics card, with full-domain sampling and filtering taking less than 8 seconds, and can be embedded in minute-level early warning links.

[0136] The convection nucleation event object is encapsulated as an early warning message, a named data name is generated, and after digital signature and verifiable delay function timestamp, it is written into the blockchain, and the early warning message is distributed in the named data network according to the reinforcement learning replication strategy, and the replication level, action threshold and regularization coefficient are adjusted in real time according to the network delay and hit rate.

[0137] After the convection nucleation event object is generated, the present application needs to reliably and traceably propagate the early warning information to the end user within seconds, while ensuring that the replication level of the information in the network can be dynamically adjusted according to the real-time load. For this purpose, a "named data network-blockchain double-layer publishing mechanism" is proposed, and the core process includes message encapsulation, trusted timestamp, blockchain storage and adaptive replication.

[0138] Message encapsulation. The convection nucleation event object contains event polygon, predicted burst time, minimum action S min , confidence γ. When encapsulating, the fields are first serialized to obtain the Payload byte string. Then a named data name is generated:

[0139] N = / alert / lat_lon / level / utc / h

[0140] Where lat_lon is the event center longitude and latitude code, level is the height layer identifier, utc is the coordinated universal time, and h is the SHA-256 hash value of Payload. The hierarchical structure of the name facilitates the longest prefix matching and cache positioning of the named data network.

[0141] Trusted timestamp. To avoid replay attacks and time tampering, a verifiable delay function is used to generate a timestamp. Given the seed d as the Payload hash value, the delay function outputs:

[0142]

[0143] Where n is the security modulus, and T is the iteration step number. The verifier only needs to perform a modular exponential calculation once to confirm that the delay function has indeed run 2 T times square. The delay function ensures that there is a physically incompressible delay between the signing time and the publishing time, preventing the falsification of early warnings after the fact.

[0144] Blockchain storage. {N, d, y, π} is written as transaction data into the proof-of-stake chain. The on-chain record includes the block hash, timestamp and serial number, and any node can verify the early warning publishing order and data integrity through the chain browser. The storage only records the hash and proof, without revealing the content, balancing privacy and auditability.

[0145] Digital signature. Payloads are signed with the Elliptic Curve Digital Signature Algorithm. The signature is attached to the MetaInfo field of the named data packet, ensuring that the data has not been modified during network forwarding.

[0146] Adaptive replication. Named data networks use an on-demand "pull" mechanism: terminals send interest packets to name prefixes, and routing nodes return data upon cache hit. To improve delivery speed in hotspots, the invention introduces a reinforcement learning replication strategy. Replication level k takes values from 0 to 3, indicating the distribution range of data copies in the node and its several-hop neighbors. The network delay Δt and the cache hit rate η form the state vector s = (Δt, η), and the action set is different k. The reward function is:

[0147] R = -Δt + βη - γk

[0148] where β and γ are weighting coefficients. Through double deep Q-network online training on edge routers, the replication level that can simultaneously reduce latency and improve hit rate is selected. The replication level change is written back to the event center in real time. If k is higher than 2 for a long time, it indicates that the network load is large, and the system automatically increases the action threshold S crit and the regularization coefficient λ, reducing the frequency of subsequent warning triggers and smoothing the quantum annealing mapping, forming a "network-model" closed loop.

[0149] Example. On July 23, 2024, at 15:10, a convective nucleation event was detected in the Beijing urban area, generating a Payload of 512 bytes. The verifiable delay function parameter setting T = 20, and the service node took 1 second to obtain y and π. The transaction was confirmed by the blockchain within 2 seconds after uploading. The initial replication level of the edge router k = 1, the average latency measured was 1.8 seconds, and the hit rate was 0.75; the deep Q-network increased k to 2 according to the reward, and the latency decreased to 1.2 seconds, and the hit rate increased to 0.84. Rain appeared in the urban area before and after 15:16, and the public terminals received the push 4 minutes earlier than the 35 decibel threshold of the radar echo. The on-chain record proves that the warning was indeed issued at 15:11:03, meeting the post-audit requirements.

[0150] Effect comparison. In 24 independent case tests, the reinforcement learning replication strategy shortened the average receiving latency of terminals by 38%, increased the hit rate by 12%, and reduced unnecessary warnings in high-load scenarios by 18% through threshold self-adjustment. Digital signature and delay function ensure that messages cannot be forged and are time-credible, and blockchain storage provides a publicly verifiable publishing sequence. The overall process tightly couples meteorological model output, network distribution, and trusted audit, building a "fast, reliable, and traceable" full-stack warning link in the satellite fusion Internet of Things scenario.

[0151] Preferably, the early warning message is embedded with a verifiable delay function timestamp after elliptic curve digital signature and written into a blockchain maintained by a proof-of-stake consensus mechanism to ensure the integrity of the evidence chain.

[0152] After the early warning message is output from the algorithm side, the present application first performs elliptic curve digital signature to bind the message content and the publisher. The signature process selects the curve parameter secp256r1 recommended by the National Cryptography Bureau, uses the private key k priv to sign the message digest d to calculate (r, s). Where d is the SHA-256 digest, r and s are the results of horizontal and vertical coordinates after finite field operation; any receiver holding the corresponding public key k pub can complete the verification through one elliptic curve multiplication and hash comparison. The signature ensures that the content cannot be tampered with, and the publisher has undeniable.

[0153] In order to prevent the publisher from retroactively forging earlier warnings, the present application uses a verifiable delay function as a timestamp mechanism. Given the message digest d and the security modulus N (the product of two large primes), the delay function calculates:

[0154]

[0155] Where T is the iteration depth parameter (controls the shortest calculation time in logarithmic form). The proof π is obtained through a fast algorithm and can be verified at the other end through one modular exponentiation operation, thus having the characteristics of "slow generation and fast verification". The sequence (d, y, π, T) is serialized into the Meta area, so that any node can check that the timestamp indeed occupies the theoretically minimum square time, blocking the post-forgery window.

[0156] After completing the signature and delay function encapsulation, the message is written into a blockchain maintained by a proof-of-stake consensus mechanism. The chain uses a fixed 2-second block generation Byzantine fault-tolerant algorithm, each new block stores the hash of the previous block, the timestamp and the transaction list, ensuring that the chain structure is tamper-proof. The message on-chain process is divided into two steps: first, (d, y, π, T) is packaged as transaction data tx; then the right to verify the node generates the next block based on the proportion of holding coins, writes tx into and broadcasts. Any post-auditing party can quickly locate the early warning publishing order through the block height and transaction hash to achieve evidence consistency.

[0157] The blockchain storage is also linked to the distribution layer of the named data network. When the edge router receives the early warning data packet, it first locally verifies the signature and delay function; if it passes, it writes the block hash into the data packet MetaInfo and caches it so that the subsequent interest packet can be directly verified. This avoids repeated chain queries while ensuring that the cached copy still has the original tamper-proof credentials.

[0158] Effect example: 2024-07-23 15:11:03 Coordinated Time generates an early warning message, T = 20 corresponds to the shortest delay of about 1 second; calculate y = 5d8f…, π = 0b36…; Signature (r, s) = (3a97…, 72f4…). The transaction tx is packaged into the block with height 5432189 by the stake verification node at 15:11:05. Starting at 15:11:06, the terminal pulls the data packet by prefix / alert / 39.90_116.40 / low / 151103, and can locally verify the signature, delay function and block hash at the same time. If any field is tampered with, the signature or timestamp verification will fail immediately; if the publisher tries to replay the old packet after 15:12, the block height will not match and will also trigger rejection.

[0159] Statistical 24 times of actual measurement cases, the average time consumption of block confirmation is 2.1 seconds, and the average delay of message reaching the edge router and completing the three verifications is 3.4 seconds; compared with the traditional central server signature scheme, the decentralized chain storage record shortens the multi-point consistent confirmation time by 42%, and reduces the single-point failure risk to zero. The on-chain public height and hash also greatly simplify the supervision and post-evaluation process, so that "whether the early warning is timely" can be given a quantitative conclusion within 1 minute by an automatic script.

[0160] The application connects elliptic curve signature, verifiable delay function and proof of stake chain in series to form a three-fold guarantee of "content integrity-time credibility-evidence tamper-proof": the signature ensures that the content has not been tampered with, the delay function ensures that the time sequence cannot be forged, and the block chain provides a publicly auditable sequence. Combined with the prefix caching mechanism of the named data network, even if part of the chain nodes or routing nodes are offline, the terminal can still obtain the early warning data packet containing self-verification information, thereby realizing fast, reliable and traceable full-link publishing in the satellite integrated Internet of Things meteorological early warning scene.

[0161] Preferably, the cache replication strategy of the named data network is determined by a reinforcement learning model trained with network delay and hit rate as the reward function to determine the replication level in real time, and the action threshold and the regularization coefficient of the optimal transmission mapping are updated synchronously after the replication level changes to maintain the balance of the overall early warning timeliness and accuracy.

[0162] The named data network undertakes the real-time distribution of early warning messages on the "cloud-edge-end" path. In order to balance low latency in sudden peak scenarios and resource conservation in daily scenarios, the application dynamically decides the cache replication level by a reinforcement learning model. When the edge router receives the early warning data packet and completes the signature and timestamp verification, the replication level decision process is triggered. The core steps include state representation, reward design, policy update and feedback coupling.

[0163] State representation. Router computes local network latency and cache hit rate every 2 seconds to represent current network load state as a vector s = (At, η). Where At is the round-trip latency from receiving an interest packet to returning a data packet, in milliseconds; η is the proportion of interest packets that hit the cache in the past 60 seconds. Two-dimensional features are standardized and input into a double deep Q network. Q network adopts a fully connected structure, with a hidden layer size of 128 and a rectified linear unit as the activation function; the parameter synchronization period of the main network and the target network is set to 100 steps.

[0164] Action space. Replication level k e {0, 1, 2, 3} represents the broadcast radius of data replicas around the node within 0 hops, 1 hop, 2 hops, and 3 hops. The higher the replication level, the greater the potential for latency reduction, but the more bandwidth and cache resources are occupied.

[0165] Reward function. Considering the three factors of latency, hit rate, and resource consumption, we define:

[0166] R = -At + βη - γk

[0167] Where β balances the hit rate benefit, and γ balances the replication overhead. Through historical sample grid search, we get β = 50 and γ = 10. The reward form ensures that at low load, k is reduced to save resources, and at high load or hot area, k is automatically increased.

[0168] Policy update. The experience replay buffer size is 5000, the batch size is 64, and the discount factor is 0.95. After each decision, (s, k, R, s') is stored in the buffer, and the main network weight is updated by randomly sampling from it. As the network is optimized, the Q value converges to the long-term cumulative reward, and the decision gradually stabilizes. Feedback coupling. When the replication level is greater than 2 for several decision periods, it indicates that the network pressure is high, and the system automatically increases the action threshold S crit :

[0169] S crit,new = S crit,old + α(At - 100)

[0170] α is the step size coefficient 0.05. When the replication level is less than 1 and the hit rate is higher than 0.9, the threshold is reduced to improve sensitivity. Similarly, the optimal transmission mapping regularization coefficient λ is fine-tuned:

[0171] λ new = λ old (1 + 0.02k avg )

[0172] This closed-loop coupling ensures that the algorithm triggering frequency and network carrying capacity change synchronously, achieving "model-network" self-balancing.

[0173] The replication operation is broadcast by rewriting the prefix of the interest packet, the copy TTL is set to 30 minutes, and the timeout is automatically recycled. In order to prevent cache pollution, only the warning packet with the name prefix containing 'alert' and the confidence greater than 0.7 is executed for replication. The router runs an asynchronous CUDA stream, and the forward inference can be completed within 2 milliseconds. The network monitoring module polls the port with the user space data plane development kit, and the error is less than 1 millisecond.

[0174] Embodiment. On July 23, 2024, 15:10, the router in the urban area of Beijing observed that Δt = 190 milliseconds and η = 0.74. The Q network gave the action k = 2, and the reward R = -190 + 50 * 0.74 - 20 = -143. After the replication was completed, the delay was reduced to 120 milliseconds from 15:10 to 15:12, and the hit rate was increased to 0.83. The new state feedback made the network keep k = 2 in the next period. Since k was 2 for 3 consecutive periods, the system adjusted the action threshold from 3.5 to 3.6, reducing the low-confidence warning output. After 16 minutes, the rain ended, Δt returned to 95 milliseconds, and η increased to 0.92, and the algorithm automatically downgraded to k = 1, while λ was reduced by 2%, shortening the quantum annealing solving time.

[0175] Effect evaluation. The actual measurement of the comparison of 10 busy urban areas and 14 sparse towns showed that the reinforcement learning replication strategy reduced the average terminal receiving delay in the busy area from 1.9 seconds to 1.2 seconds, and the cache redundancy in the sparse area was reduced by 45%. The dynamic coupling of the threshold-regularization reduced the total warning amount by 18% compared with the fixed parameter scheme, and the false alarm rate did not increase significantly. This result shows that the application effectively alleviates the network congestion and model over-triggering problem while maintaining the timeliness of the warning.

[0176] Through the design of "reinforcement learning replication-model threshold coupling", the application realizes the perception-decision-feedback closed loop: the change of the network environment is mapped in real time to the adjustment of the replication strategy, the strategy synchronously affects the model trigger threshold and the mapping smoothness, and then changes the subsequent warning traffic, forming a self-stabilizing mechanism. This cross-layer adaptive characteristic improves the overall reliability and scalability of the satellite integrated Internet of Things weather warning system.

[0177] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system or a computer program product. Therefore, the application can adopt a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0178] The above is only an embodiment of the application and is not intended to limit the application. Those skilled in the art can make various modifications and changes to the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the scope of the claims of the application.

Claims

1. A satellite fusion Internet of Things weather warning and release method, characterized in that, The application relates to a method for generating a lightning early warning message. The method comprises the following steps: Collecting static satellite or low-orbit satellite remote sensing data, ground Internet of Things meteorological electric field data, millimeter wave communication attenuation data and ground optical imaging cloud data, performing radiation calibration, space-time registration and integrity processing to form a unified observation data set; Based on the unified observation data set, a weighted multi-modal graph is constructed, the source density and sink density are discretized into a binary optimization model, a quantum annealing algorithm is used to obtain an optimal transmission mapping, a three-dimensional wind speed field, a specific humidity field and a temperature field are reconstructed according to a mass conservation equation, and a fused atmospheric state field and potential energy tensor are generated; A space-time high potential energy function is constructed based on the fused atmospheric state field and potential energy tensor, variation sampling is performed on a candidate air particle path, and an action amount is calculated, when the action amount is lower than a threshold value and the potential energy decreases, a minimum action amount path is screened, and a convection nucleation event object is generated; 2. The method of claim 1, wherein, The convection nucleation event object is encapsulated into a warning message, a named data name is generated, a digital signature and a verifiable delay function timestamp are completed, and then the warning message is written into a blockchain, the warning message is distributed in a named data network according to a reinforcement learning replication strategy, and the replication level, the action amount threshold value and the regularization coefficient are adjusted in real time according to the network time delay and the hit rate.

3. The method of claim 1, wherein, The static satellite or low-orbit satellite remote sensing data is composed of infrared brightness temperature data and lightning imaging data, the infrared brightness temperature data are used for extracting a cloud top brightness temperature field, and the lightning imaging data are used for positioning instantaneous discharge activities to supplement convection triggering information.

4. The method of claim 1, wherein, Radiation calibration is performed by matching the radiation transmission model output on the clear sky pixel to establish a brightness temperature baseline, space-time registration is completed by performing an exterior orientation angle correction on the orbit attitude data and combining an optical interpolation method, and integrity processing is performed by establishing a Gaussian process prior to spatially inferring the missing pixels.

5. The method of claim 1, wherein, The observation nodes of the weighted multi-modal graph are composed of satellite pixel nodes, millimeter wave precipitation rate slice nodes and ground cloud body optical flow nodes, the node feature vectors are spliced by brightness temperature, precipitation rate, optical flow velocity and electric field amplitude, and the edge weight is determined according to an exponential decay function of the Euclidean distance of the node feature vectors.

6. The method of claim 1, wherein, The source density and sink density are discretized into a binary variable optimization model, and the optimization model is embedded into a quantum bit network by using a quantum annealing processor, an optimal transmission mapping is obtained by single pulse annealing, and a unique minimum free energy solution is returned.

7. The method of claim 1, wherein, The mass flow of the optimal transmission mapping is used as a source term, a mass conservation partial differential equation is solved by using a finite difference method, and a divergence-free constraint is maintained in the iteration process, and a three-dimensional wind speed field, a specific humidity field and a temperature field are obtained.

8. The method of claim 1, wherein, The space-time high potential energy function is composed of specific humidity, isentropic potential temperature, millimeter wave precipitation rate and potential energy tensor through linear weighted combination, the specific humidity and isentropic potential temperature are used for describing thermal state, the millimeter wave precipitation rate is used for describing liquid water content, and the potential energy tensor is used for describing kinetic energy and potential energy distribution.

9. The method of claim 1, wherein, The variation sampling of the candidate air particle path adopts a Hamilton Monte Carlo algorithm, the phase space volume is kept unchanged by using symplectic integration, the minimum action amount is used as the path screening criterion, and a single minimum action path is output. The warning message is embedded into a verifiable delay function timestamp after being signed by an elliptic curve digital signature, and is written into a blockchain maintained by using a proof of stake consensus mechanism to ensure the integrity of the data chain.

10. The method of claim 1, wherein, The cache replication strategy of named data networking is determined by a reinforcement learning model trained with network latency and hit rate as reward function in real time, and the replication level is changed to update the action threshold and the regularization coefficient of optimal transmission mapping in real time to maintain the balance between the overall warning timeliness and accuracy.

Images