A method, system and device for quantifying the reliability of vertical atmospheric predictions

Abstract

The present application belongs to the technical field of atmospheric detection and prediction, and particularly relates to a credibility quantification method, system and device based on vertical atmospheric prediction. The method dynamically constructs a vertical correlation graph based on the vertical correlation between a single detection height and the target height for any number of target heights, and then learns the vertical correlation between the detection height and each target height in parallel through a vertical graph attention network to output meteorological variable prediction results. The uncertainty of each target height prediction result is quantified based on the vertical correlation strength and decomposed into two types of uncertainty, namely inherent and cognitive uncertainty. Finally, a unified credibility index is output. The training and verification process is combined to ensure the applicability of the method. The present method can be flexibly adapted to different target height number scenarios, aiming to achieve accurate prediction from a single detection height to any number of target heights, and to provide quantitative credibility evaluation for the prediction results, thereby providing a reliable basis for decision-making in the field of atmospheric science.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atmospheric exploration and prediction, and in particular to a credibility quantification method, system and device based on vertical atmospheric prediction. BACKGROUND

[0002] In the fields of weather forecasting, aviation scheduling, environmental monitoring, etc., obtaining accurate prediction results of atmospheric parameters (such as temperature, humidity, air pressure, wind speed, etc.) at different altitudes is the key to ensuring the orderly development of related work. Traditional atmospheric prediction technology is often limited by the number of fixed target altitudes and specific symbol identifiers. When the number of target altitudes to be predicted changes or the altitude identifier changes in actual application, the model needs to be redesigned and trained, which has poor adaptability and cannot meet the needs of diversified scenarios.

[0003] More importantly, most existing technologies can only output atmospheric parameter prediction values and cannot effectively quantify the credibility of the prediction results. Users cannot judge the reliability of the prediction data. In scenarios where the accuracy of atmospheric parameters is extremely high (such as aviation flight path planning and extreme weather warning), over-reliance on low-credibility prediction results or neglect of high-credibility prediction information may lead to decision-making errors, causing safety risks and economic losses. Some technologies that attempt to analyze credibility have the problems of non-uniform evaluation logic and non-universal indicators. The credibility evaluation methods for different target altitudes differ greatly, and cannot achieve horizontal comparison and comprehensive judgment, which limits their practicality.

[0004] Therefore, there is an urgent need for a method that can flexibly adapt to different numbers of target altitudes and systematically and uniformly quantify the credibility of vertical atmospheric prediction results to overcome the shortcomings of existing technologies and provide more reliable decision support for atmospheric-related fields. SUMMARY

[0005] To overcome the shortcomings of existing technologies, the present application provides a credibility quantification method, system and device based on vertical atmospheric prediction. This method can break free from the limitations of fixed target altitude numbers and specific symbol identifiers, accurately predict atmospheric parameters at any number of target altitudes, and comprehensively and uniformly quantify the credibility of the prediction results, providing users with clear credibility information and interpretation basis.

[0006] To achieve the above-mentioned purposes, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a credibility quantification method based on vertical atmospheric prediction, which comprises:

[0008] For each target altitude to be predicted, a vertical correlation graph is dynamically constructed based on the vertical correlation between a single detection altitude and the target altitude.

[0009] a vertical graph attention network is configured for each target height, the vertical graph attention network is used to learn the vertical correlation between a single detection height and a corresponding target height, to calculate the attention coefficient of a target node to its neighbor nodes, to aggregate the projection features of the neighbor nodes based on the attention coefficient to update the target prediction node feature, and to output the meteorological variable prediction result of the target height in combination with a preset physical constraint function;

[0010] Based on the vertical correlation strength between the detection height and each target height, the number of Monte Carlo Dropout sampling is determined, the uncertainty of the meteorological variable prediction result of each target height is calculated, and the uncertainty is decomposed into inherent uncertainty and cognitive uncertainty to generate a general reliability index at the target height;

[0011] A historical data sample set containing different detection heights and different numbers of target height combinations is used, the error between the meteorological variable prediction result and the true value is used as the optimization target, all vertical graph attention networks are trained and parameter optimized in parallel, and the trained vertical graph attention networks are verified.

[0012] Optionally, the vertical correlation graph includes a node set and an edge set;

[0013] The node set includes core detection nodes and neighbor nodes, and attribute data of the core detection nodes and the neighbor nodes;

[0014] The neighbor nodes include target prediction nodes and auxiliary transition nodes;

[0015] The edge set is composed of connection edges between adjacent nodes in the vertical direction, and the weight of the connection edge is determined based on the vertical distance between the two nodes it connects and / or the historical meteorological gradient stability;

[0016] The attribute data of the core detection node includes the measured values of temperature, relative humidity, air pressure, wind speed and their respective sensor error values;

[0017] The attribute data of the target prediction node includes the historical vertical correlation features between the detection height and the target height;

[0018] The historical vertical correlation features between the detection height and the target height include the average difference of temperature, relative humidity, air pressure and wind speed in the historical period, and the historical standard deviation of temperature difference and air pressure difference;

[0019] The attribute data of the auxiliary transition node includes the historical average data and / or the measured data of the corresponding height.

[0020] Optionally, the number of auxiliary transition nodes is determined based on a predefined number determination rule, which includes:

[0021] obtaining an absolute vertical distance between the probe height and the current target height;

[0022] generating an auxiliary transition node when the absolute vertical distance is not greater than a first preset threshold;

[0023] generating two auxiliary transition nodes when the absolute vertical distance is greater than the first preset threshold and not greater than a second preset threshold;

[0024] generating at least three auxiliary transition nodes when the absolute vertical distance is greater than the second preset threshold.

[0025] Optionally, the weight of the connecting edge is determined based on the vertical distance between the heights represented by the two nodes connected by the edge and / or the historical meteorological gradient stability, and the weight of the connecting edge comprises:

[0026] for an edge connecting the core probe node and any auxiliary transition node, the weight of the edge is determined by the vertical distance between the heights represented by the two nodes;

[0027] for an edge connecting any auxiliary transition node and the target prediction node, the weight of the edge is determined by the vertical distance between the heights represented by the two nodes and the historical standard deviation of the temperature difference between the probe height and the target height.

[0028] Optionally, the attention coefficient of the target node to its neighbor nodes comprises:

[0029] performing linear transformation on the projection feature of the target prediction node and the projection feature of any neighbor node thereof, splicing the linearly transformed features;

[0030] performing dot product operation on the spliced features and the globally shared attention vector, and processing through an activation function;

[0031] multiplying the processing result by the edge weight from the neighbor node to the target prediction node to obtain the original attention score;

[0032] performing exponential operation on the original attention scores of all neighbor nodes, summing up the exponentialized scores, and dividing each neighbor node's exponentialized score by the sum to obtain the normalized attention coefficient.

[0033] Optionally, the output of the meteorological variable prediction result of the target height in combination with the preset physical constraint function comprises: weighting and summing the projection features of each neighbor node according to the normalized attention coefficient to obtain a comprehensive feature fused with multi-node information; performing linear transformation on the comprehensive feature through globally shared decoding weights and biases to output the meteorological variable prediction value of the target height; introducing a physical constraint term; the physical constraint term comprises a temperature vertical gradient constraint and a pressure vertical decreasing constraint.

[0034] The meteorological variable prediction value is corrected to obtain a final meteorological variable prediction result.

[0035] Optionally, the determination of the number of Monte Carlo Dropout samplings based on the vertical correlation strength between the detection height and each target height comprises:

[0036] The average value of all edge weights in the vertical correlation graph is calculated as the vertical correlation strength.

[0037] The number of samplings required for the target height is determined by linear interpolation between a preset minimum value and a maximum value according to the vertical correlation strength.

[0038] Optionally, the inherent uncertainty is calculated based on the sensor error of the detection height and the standard deviation of the historical meteorological variable difference between the detection height and the target height.

[0039] The cognitive uncertainty is obtained by taking the non-negative value of the result of subtracting the uncertainty variance from the inherent uncertainty variance.

[0040] In a second aspect, the present application provides a credibility quantification system based on vertical atmospheric prediction, which comprises:

[0041] A dynamic construction module is configured to dynamically construct a vertical correlation graph based on the vertical correlation relationship between a single detection height and each target height to be predicted.

[0042] A meteorological variable prediction module is configured to configure a vertical graph attention network for each target height, learn the vertical correlation relationship between a single detection height and the corresponding target height through the vertical graph attention network, calculate the attention coefficient of the target node to its neighbor nodes, aggregate the projection features of the neighbor nodes based on the attention coefficient to update the target prediction node features, and output the meteorological variable prediction result of the target height in combination with a preset physical constraint function.

[0043] An uncertainty quantification module is configured to determine the number of Monte Carlo Dropout samplings based on the vertical correlation strength between the detection height and each target height, calculate the uncertainty of the meteorological variable prediction result of each target height, decompose the uncertainty into inherent uncertainty and cognitive uncertainty, and generate a general credibility index for the target height.

[0044] A training and verification module is configured to use a historical data sample set comprising different detection heights and different numbers of target height combinations, use the error between the meteorological variable prediction result and the true value as an optimization target, perform parallel training and parameter optimization on all vertical graph attention networks, and verify the trained vertical graph attention networks.

[0045] In a third aspect, the present application provides an electronic device, comprising:

[0046] at least one processor; and

[0047] a memory connected to the at least one processor in communication; wherein,

[0048] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the method of any one of the first aspect.

[0049] Compared with the closest prior art, the present application has the beneficial effects of:

[0050] The present application proposes a credibility quantification method, system and device based on vertical atmospheric prediction, which completely gets rid of the restrictions of fixed target height quantity and specific symbol identification, can automatically adapt to any number of target heights, does not need to redesign the method due to changes in target height quantity or symbol difference, is suitable for various scenes such as weather forecasting, aviation scheduling and environmental monitoring, and reduces the application threshold and adjustment cost of the method.

[0051] The present application accurately depicts the vertical correlation characteristics of the detection height and the target height by dynamically constructing a vertical correlation graph from the detection height to the target height; combined with the parallel learning of the vertical graph attention network and the general physical constraint function, it can not only capture complex vertical correlation rules, but also ensure that the prediction results comply with atmospheric physical laws, avoid the occurrence of prediction values that violate common sense, and improve the accuracy and reliability of the prediction results. Moreover, in the training process, the training sample data collected by the present application covers different combinations of detection height and target height quantity, and the verification index is unbiased, ensuring that the method can stably operate in various scenarios and has good generalization ability, and can adapt to atmospheric prediction needs in different regions and different time periods.

[0052] The present application carries out exclusive credibility quantification for the prediction results of each target height, decomposes uncertainty into inherent and cognitive types, outputs quantification indicators, probability confidence bands and qualitative labels in a unified format, and users can fully master the prediction credibility details and uncertainty sources, providing more accurate reference for decision-making and reducing the decision-making risks caused by unclear information.

[0053] After completing the steps of the present application, three general factors of vertical distance, gradient stability and data quantity can be set to quantitatively explain the credibility differences of different target heights, clearly present the reasons for the credibility differences, improve the transparency of the method, enhance the trust of users on the method, and provide a clear direction for method optimization. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0055] Figure 1 is a flow chart of a credibility quantification method based on vertical atmospheric prediction provided by the present application;

[0056] Figure 2 is a schematic structural diagram of a credibility quantification system based on vertical atmospheric prediction provided by the present application;

[0057] Figure 3 is an internal structural diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0058] The embodiments of the technical solutions of the present application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0059] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present application should be understood as the usual meanings understood by the skilled in the art to which the present application belongs.

[0060] The embodiments of the present application provide a credibility quantification method, system and device based on vertical atmospheric prediction, which sequentially includes the following processes: dynamically constructing a vertical correlation graph from a detection height to a target height, performing vertical correlation relationship learning and meteorological variable prediction in parallel, quantifying the credibility of the target height prediction result, training and verifying the method, etc. The dynamic and accurate prediction from a single detection height to any number of target heights is realized, and the credibility quantification index and reasonable explanation of each prediction result are output at the same time. The constraints of fixed target height number and specific symbols in vertical atmospheric prediction are completely eliminated.

[0061] Embodiment 1: Reference Figure 1 , Figure 1 A credibility quantification method based on vertical atmospheric prediction is provided for the embodiment 1 of the present application, which specifically includes the following steps:

[0062] S101, for each target height to be predicted, a vertical correlation graph is dynamically constructed based on the vertical correlation relationship between a single detection height and the target height;

[0063] S102, for each of the target heights, configuring a vertical graph attention network, learning a vertical correlation between a single detection height and a corresponding target height through the vertical graph attention network, calculating an attention coefficient of a target node to a neighbor node of the target node, aggregating projection features of the neighbor nodes based on the attention coefficient to update a target prediction node feature, and outputting a meteorological variable prediction result of the target height in combination with a preset physical constraint function;

[0064] S103, based on vertical correlation strengths between the detection height and the target heights, determining a number of Monte Carlo Dropout sampling times, calculating an uncertainty of the meteorological variable prediction result of each target height, decomposing the uncertainty into an inherent uncertainty and a cognitive uncertainty, and generating a general reliability index at the target height;

[0065] S104, using a historical data sample set containing combinations of different detection heights and different numbers of target heights, taking an error between the meteorological variable prediction result and an actual value as an optimization target, performing parallel training and parameter optimization on all vertical graph attention networks, and verifying the trained vertical graph attention networks.

[0066] In the above embodiment, the vertical correlation graph dynamically constructed in step S101 includes a node set and an edge set;

[0067] The node set includes: core detection nodes and neighbor nodes; and attribute data of the core detection nodes and the neighbor nodes; the neighbor nodes include: target prediction nodes and auxiliary transition nodes;

[0068] The attribute data of the core detection nodes includes: measured values of temperature, relative humidity, air pressure, and wind speed, and respective sensor error values of the measured values;

[0069] The attribute data of the target prediction nodes includes: historical vertical correlation features between the detection height and the target height;

[0070] The historical vertical correlation features between the detection height and the target height include: average difference values of temperature, relative humidity, air pressure, and wind speed in a historical period, and historical standard deviations of temperature difference and air pressure difference;

[0071] The attribute data of the auxiliary transition nodes includes: historical average data and / or measured data of corresponding heights;

[0072] In the above embodiment, the edge set is composed of connection edges between adjacent nodes in the vertical direction, and weights of the connection edges are determined based on vertical distances between heights represented by the two nodes connected by the connection edges and / or historical meteorological gradient stability.

[0073] Specifically, for the edge connecting the core probe node and any auxiliary transition node, the weight is determined by the vertical distance between the heights represented by the two nodes.

[0074] For the edge connecting any auxiliary transition node and the target prediction node, the weight is determined by the vertical distance between the heights represented by the two nodes, and the historical standard deviation of the temperature difference between the probe height and the target height.

[0075] Further, in step S101, based on the vertical correlation between the single probe height and the target height, the dynamic construction of the vertical correlation graph specifically includes: for any number of target heights, automatically generating a dedicated vertical correlation graph for each target height, and by dynamically adjusting the graph structure, ensuring that the vertical correlation between the probe height and the target height is fully and accurately reflected, while all nodes included in the vertical correlation graph are processed in a unified dimension, laying a foundation for subsequent learning and prediction.

[0076] In the above embodiments, the node set of the vertical correlation graph includes three types of core probe nodes, target prediction nodes and auxiliary transition nodes, and the number and attributes of the nodes dynamically change with the characteristics of the probe height and the target height, rather than being fixedly set.

[0077] Among them, the core probe node corresponds to the probe height one-to-one, and is the only fixed node in the vertical correlation graph, which is used to provide basic measured data support. Its attributes include measured meteorological data of the probe height and sensor errors; among them, the measured meteorological data includes temperature, relative humidity, air pressure, wind speed and other core atmospheric parameters, and the sensor errors include the inherent errors of the corresponding meteorological parameter sensors. These data directly reflect the actual state of the atmosphere at the probe height and the data acquisition accuracy.

[0078] Secondly, the target prediction node corresponds to each target height one-to-one, and is used to carry the correlation information between the probe height and the target height. Its attribute is the historical vertical correlation feature of the probe height and the target height, including the historical average difference of temperature, relative humidity, air pressure and wind speed between the two (reflecting the long-term correlation trend), and the historical standard deviation of the temperature difference and the air pressure difference (reflecting the correlation stability), which provides a historical basis for learning the vertical correlation.

[0079] The auxiliary transition node is used to connect the core probe node and the target prediction node, and ensures that the vertical correlation is not interrupted. The number of auxiliary transition nodes is dynamically adjusted according to the vertical distance between the probe height and the target height. When the vertical distance is small, fewer auxiliary transition nodes are set; when the vertical distance is large, the number of auxiliary transition nodes is increased to more finely depict the changes of atmospheric parameters in the vertical direction. The attribute of the auxiliary transition node preferably uses the measured data of the corresponding height, and if there is no measured data, the historical average data is used to ensure the accuracy and availability of the node attribute.

[0080] In the above embodiment, the construction method of the edge set of the vertical association graph includes that the edges of the vertical association graph exist only between nodes adjacent in the vertical direction; and the edge weight is used to quantify the vertical association strength between nodes, and the calculation logic is unified and does not depend on the number of target heights. For example, the edge weights of the core detection node and the adjacent auxiliary transition node, and the auxiliary transition node and the target prediction node.

[0081] Specifically, the edge weight of the core detection node and the auxiliary transition node is determined only by the vertical distance between the two nodes. The closer the distance, the greater the edge weight, which directly reflects the influence of the distance on the association between the nodes, and ensures that the strong association between the nodes in close proximity is accurately represented.

[0082] The edge weight of the auxiliary transition node and the target prediction node is determined by the vertical distance between the two nodes and the historical gradient stability of the detection height and the target height. The closer the distance, the more stable the historical gradient (i.e., the more fixed the change rule of the atmospheric parameter in the vertical direction), and the greater the edge weight, which more comprehensively reflects the vertical association characteristics between the nodes and provides a reasonable basis for subsequent attention allocation.

[0083] In the above embodiment, the number of auxiliary transition nodes is determined based on a pre-defined number determination rule, which includes:

[0084] Obtaining the absolute vertical distance between the detection height and the current target height;

[0085] When the absolute vertical distance is not greater than a first preset threshold, one auxiliary transition node is generated;

[0086] When the absolute vertical distance is greater than the first preset threshold and not greater than a second preset threshold, two auxiliary transition nodes are generated;

[0087] When the absolute vertical distance is greater than the second preset threshold, at least three auxiliary transition nodes are generated.

[0088] In the above embodiment, the calculation of the attention coefficient of the target node to its neighbor nodes in step S102 includes:

[0089] Performing linear transformation on the projection feature of the target prediction node and the projection feature of any neighbor node thereof, and splicing the linearly transformed features;

[0090] Performing dot product operation on the spliced features and the globally shared attention vector, and processing through an activation function;

[0091] Multiplying the processing result by the edge weight of the neighbor node to the target prediction node to obtain the original attention score;

[0092] The normalized attention coefficients are obtained by exponentiating and summing the original attention scores of all neighbor nodes, and dividing the exponentiated score of each neighbor node by the sum.

[0093] In step S102, a vertical graph attention network is configured for each target height, and the vertical correlation between a single detection height and the corresponding target height is learned through the vertical graph attention network. Specifically, the vertical correlation includes:

[0094] For any number of target heights, an independent vertical graph attention network (Vertical-GAT) is configured for each target height. Through parallel learning, the vertical correlation between the detection height and each target height is accurately captured, and a self-defined physical constraint function is combined to output meteorological variable prediction results that conform to atmospheric physical laws.

[0095] It is worth noting that in the design of the vertical graph attention network structure, it is necessary to ensure that the structures of all vertical graph attention networks configured for target heights are completely consistent and include both attention calculation and prediction output, ensuring the generality and consistency of the method.

[0096] Among them, the attention calculation includes two steps of attention score calculation and attention coefficient normalization, which are used to reasonably allocate the contribution weight of each neighbor node to the target prediction node.

[0097] The implementation process of attention score calculation includes: taking the projection features of the neighbor nodes (core detection nodes or auxiliary transition nodes) and the target prediction node as input, combining the globally shared attention vector and the attention weight matrix, performing linear transformation and concatenation on the features, then introducing the edge weight from the neighbor node to the target prediction node, and generating the original attention score through the activation function. The original attention score intuitively reflects the importance of the neighbor node to the target prediction node, providing a basis for subsequent weight allocation.

[0098] The implementation process of attention coefficient normalization includes: exponentiating the original attention score (to ensure that the score is non-negative), summing the exponentiated scores of all neighbor nodes, and finally dividing the exponentiated score of each neighbor node by the sum to obtain the normalized attention coefficient. The sum of the attention coefficients of all neighbor nodes is 1, which realizes the reasonable allocation of the contribution weight of each neighbor node and ensures that the features of important nodes are fully valued.

[0099] In the step S102, the meteorological variable prediction result of the target height output by the preset physical constraint function includes: weighting and summing the projection features of each neighbor node to obtain a comprehensive feature of fused multi-node information according to the normalized attention coefficient; performing linear transformation on the comprehensive feature through the globally shared decoding weight and bias to preliminarily generate the meteorological variable prediction value of the target height; introducing a physical constraint term; the physical constraint term includes temperature vertical gradient constraint and pressure vertical decrease constraint, and the preliminary prediction value is modified to obtain the final meteorological variable prediction result; the meteorological variable prediction result includes temperature, relative humidity, pressure, wind speed, wind direction and the like, and the prediction result is ensured to comply with the atmospheric vertical physical law.

[0100] Specifically, the preset physical constraint function is introduced to avoid the prediction result from violating the atmospheric physical common sense, and the function includes temperature vertical gradient constraint and pressure vertical decrease constraint. The temperature vertical gradient constraint is based on the physical characteristics of the atmospheric temperature vertical decrease rate (the inherent change law of the atmospheric temperature with the increase of height), takes the measured temperature at the detection height as the reference, combines the vertical distance between the detection height and the target height, and calculates the theoretical temperature at the target height. The constraint ensures that the change of the predicted temperature in the vertical direction complies with the actual atmospheric law, and prevents the temperature from abnormally increasing or decreasing with the height. The pressure vertical decrease constraint is based on the Laplace equation (a classic physical equation describing the change of atmospheric pressure with height), takes the measured pressure at the detection height as the basis, introduces the gravitational acceleration, the gas constant (a physical parameter representing the characteristics of atmospheric gas), and the average temperature between the detection height and the target height (calculated by the measured temperature at the detection height and the theoretical temperature calculated by the temperature vertical gradient constraint), and calculates the theoretical pressure at the target height. The constraint ensures that the distribution of the predicted pressure in the vertical direction complies with the physical principle, and avoids the prediction result of abnormal change of pressure with height.

[0101] In the embodiment, in order to improve the training efficiency and the applicability of the method, batch parallel training is adopted for all vertical graph attention networks. In each training iteration, multiple vertical correlation graph data of target heights are simultaneously input, and each vertical graph attention network independently calculates two parts of loss: one is the error loss of the prediction value and the measured value of the target height, to measure the prediction accuracy; the other is the error loss of the prediction value and the output value of the preset physical constraint function, to measure the physical rationality of the prediction result; the two types of loss of all vertical graph attention networks are summed to obtain total error loss and total physical constraint loss, and the two losses are combined to form the total training loss; the shared parameters of the vertical graph attention network, including the attention vector, the attention weight matrix, the decoding weight, the bias and the like, are updated based on the total loss through the gradient descent algorithm; the synchronous optimization of all vertical graph attention networks is realized, and it is ensured that each vertical graph attention network can accurately learn the vertical correlation of the corresponding target height.

[0102] The core purpose of the above step S103 is to quantify the reliability of the target height prediction result. To achieve this purpose, step S103 determines the number of Monte Carlo Dropout (MC Dropout) samplings based on the vertical correlation strength between the detection height and the target height for each meteorological variable prediction result, carries out exclusive reliability quantification, clearly defines the uncertainty source, divides the uncertainty by type, and finally outputs a unified format reliability index to help users comprehensively understand the reliability of the prediction result.

[0103] In the above embodiment, the calculation process of the vertical correlation strength of the target height is as follows: the average of the edge weights of all neighbor nodes to the target prediction node in the vertical correlation graph is taken as the quantitative indicator of the vertical correlation strength. The larger the average edge weight, the closer the vertical correlation between the detection height and the target height, and the stronger the basic reliability of the prediction result; otherwise, the correlation is weaker, and the uncertainty of the prediction result may be higher, which needs to be reduced by more means to reduce the uncertainty evaluation error.

[0104] Further, based on the vertical correlation strength between the detection height and each target height, the number of Monte Carlo Dropout samplings is determined, which specifically includes: calculating the average value of all edge weights in the vertical correlation graph as the vertical correlation strength; determining the number of samplings required for the target height by linear interpolation between a preset minimum value and a maximum value according to the vertical correlation strength.

[0105] Further, after determining the number of Monte Carlo Dropout samplings, it further includes: quantifying the uncertainty by using the Monte Carlo Dropout method, and then reflecting the uncertainty of the prediction result.

[0106] According to the vertical correlation strength, the number of Monte Carlo Dropout samplings is dynamically adjusted: the higher the vertical correlation strength, the fewer the number of samplings; otherwise, the lower the vertical correlation strength, the more the number of samplings (by increasing the sampling amount to reduce the influence of random error on uncertainty evaluation). The target height meteorological variable prediction values obtained by each sampling are counted, and the variance of the prediction values is calculated as the quantification result of the uncertainty. The larger the variance, the greater the fluctuation range of the prediction result, and the lower the reliability.

[0107] In addition, the intrinsic uncertainty and cognitive uncertainty decomposed in the above step S103 are the uncertainty divided by type. The above two types of uncertainty correspond to different improvement directions, providing more detailed basis for method optimization and user decision-making.

[0108] The inherent uncertainty is derived from the noise characteristics of the data itself, mainly including sensor errors of the detection height and historical data noise of the detection height and target height. Based on the sensor errors of the detection height of each meteorological variable, combined with the historical average difference of the corresponding meteorological variables of the detection height and the target height and the measured value of the detection height, the influence degree of the sensor error in the vertical correlation transmission process is calculated; at the same time, the standard deviation of the historical difference of the corresponding meteorological variables of the detection height and the target height is introduced, and the inherent uncertainty is comprehensively obtained. This type of uncertainty is determined by data acquisition and historical data characteristics, and it is difficult to completely eliminate through method optimization, and it needs to be improved through means such as improving sensor accuracy and increasing historical data volume.

[0109] The cognitive uncertainty is derived from the insufficient learning of the method for the vertical correlation law, for example, insufficient learning of the correlation law due to insufficient historical data, and the model structure cannot fully capture the complex correlation. The cognitive uncertainty is obtained by taking the non-negative value of the result of subtracting the uncertainty variance from the inherent uncertainty variance; if the calculation result is negative, it is taken as 0, so as to ensure that the cognitive uncertainty is non-negative. This type of uncertainty can be reduced through optimization methods such as increasing the amount of training data and adjusting the network model structure.

[0110] In the above step S103, generating the general reliability index at the target height includes: in order to ensure that the reliability information of different target heights is comparable and easy to understand, the general reliability index of each target height is output in a unified format, including three types of information: quantitative core index, probability confidence band and qualitative reliability label. Among them, the quantitative core index includes the prediction mean (i.e. the final meteorological variable prediction value, which is the core data for user reference), the total standard deviation (the square root of the uncertainty, the smaller the value, the higher the total reliability), the inherent uncertainty ratio (the ratio of the inherent uncertainty to the uncertainty, reflecting the influence degree of data noise on the reliability), and the cognitive uncertainty ratio (the ratio of the cognitive uncertainty to the uncertainty, reflecting the influence degree of the method learning deficiency on the reliability). Through these quantitative indexes, the user can accurately master the reliability details of the prediction result.

[0111] Probability confidence band: based on the prediction mean and the total standard deviation, a confidence band with a confidence level of 95% (i.e. prediction mean ± 1.96 × total standard deviation) is set according to the statistical principle. The confidence band represents the range in which the true value of the target height atmospheric parameter may fall under the confidence probability of 95%, and the narrower the bandwidth, the higher the accuracy of the prediction result and the stronger the reliability.

[0112] Qualitative confidence label: According to the total standard deviation and the type of meteorological variable, a unified classification rule is formulated to qualitatively divide the confidence of the prediction result. For example, for temperature variables, when the total standard deviation is less than a small threshold, it is marked as "high confidence", between two thresholds, it is marked as "general confidence", and greater than a larger threshold, it is marked as "low confidence"; for other variables such as pressure and wind speed, according to the accuracy requirements of their actual application, corresponding threshold standards are formulated to ensure that the qualitative label can intuitively reflect the confidence of the prediction result and meet the rapid judgment needs of different users.

[0113] In the above step S104, the historical data sample set containing different detection heights and different number of target height combinations is used to train and optimize all vertical graph attention networks in parallel with the error between the meteorological variable prediction result and the true value as the optimization target, and the trained vertical graph attention network is verified. In fact, it is a general training and verification of the method of steps S101-S103. To ensure the stability of the method in different detection heights and different target height number scenarios, a general training and verification system is constructed, which follows unified logic and standards from data preparation, training process to performance evaluation, ensuring good generalization ability of the method.

[0114] In the above step S104, the historical data sample set containing different detection heights and different number of target height combinations: Collecting sounding data covering different regions and different seasons in the past 1-2 years, dividing data samples according to the combination of "detection height→any target height". The sample size of each combination needs to meet the requirements of reliable statistical correlation characteristics, and the sample format is unified as "detection height measured data + target height measured data + historical correlation characteristics", providing sufficient and standardized input data for the training process to ensure that the method can learn the vertical correlation rules in different scenarios.

[0115] The training method includes: all vertical graph attention networks share the initial weights of the projection layer, attention layer and decoding layer, set a fixed learning rate during the training process, and use batch update method (batch size equal to the number of target heights) to optimize the parameters;

[0116] Carrying out generalization verification: verifying the performance of the method from two dimensions of prediction accuracy and reliability to quantify accuracy. The verification index is general and unbiased, suitable for different number of target height scenarios, and ensures that the verification result can truly reflect the actual performance of the method. Prediction accuracy verification: set index threshold according to the vertical distance between the detection height and the target height. The closer the vertical distance, the flatter the change of atmospheric parameters in the vertical direction, the higher the requirement for prediction accuracy, and the smaller the RMSE (Root Mean Square Error, which measures the deviation between predicted value and measured value) threshold. The farther the vertical distance, the more complex the change of atmospheric parameters, the prediction difficulty increases, and the RMSE threshold is appropriately relaxed. The verification index includes temperature RMSE, pressure RMSE and wind speed RMSE. By comparing the predicted value of the method with the measured value of the target height, it is judged whether the prediction accuracy of the method meets the actual application requirements.

[0117] Reliability quantitative accuracy verification includes calibration error verification and uncertainty error ratio verification, which is used to evaluate the accuracy of the reliability quantitative result. Calibration error verification: statistics the proportion of the measured value of the atmospheric parameter at the target height falling within the 95% probability confidence band. If the proportion is greater than the set threshold (such as 90%), it means that the reliability quantitative result of the method is highly consistent with the actual situation, and the calibration error is small; if the proportion is lower than the threshold, it means that the reliability quantitative result has deviation, and the uncertainty calculation logic needs to be adjusted. Uncertainty error ratio verification: through additional data analysis, determine the "actual error ratio caused by insufficient historical data", calculate the difference between the cognitive uncertainty ratio output by the method and the true reference value. If the difference is less than the set threshold, it means that the method is accurate in analyzing the source of uncertainty; if the difference is larger, the calculation logic of uncertainty decomposition needs to be optimized to improve the accuracy of the judgment of the source of uncertainty.

[0118] In addition, to help users understand the reasons for the differences in the reliability of the prediction results at different target heights, and to increase the transparency and user trust of the method, the embodiment of the present application can also set a reliability explainability analysis step after step S104 is executed, and quantitatively explain the reliability difference through three general influence factors. The value range of the general influence factor is [0, 1], and the closer the value is to 1, the greater the positive influence on the reliability. Among them, the general influence factors include: vertical distance factor, gradient stability factor and data quantity factor.

[0119] Vertical distance factor: calculated based on the vertical distance between the detection height and the target height. The closer the distance, the closer the factor value to 1. This factor reflects the influence of vertical distance on reliability: the closer the distance, the more direct the correlation of atmospheric parameters in the vertical direction, the method is easier to accurately capture the correlation rule, and the reliability of the prediction result is higher; the farther the distance, the more indirect the correlation of atmospheric parameters, and the reliability is relatively reduced.

[0120] Gradient stability factor: according to the historical gradient stability calculation of the detection height and the target height, the more stable the gradient is, the closer the factor value is to 1. This factor reflects the effect of historical gradient stability on credibility: the more stable the gradient is, the more fixed the change rule of the atmospheric parameter in the vertical direction is, the more sufficient the method learns the correlation rule, and the higher the prediction credibility is; the more unstable the gradient is, the more complex and variable the change rule is, the learning difficulty of the method increases, and the credibility may decrease.

[0121] Data volume factor: calculated according to the historical common observation data volume of the detection height and the target height, when the data volume reaches the set empirical threshold, the factor value is 1; when the data volume is less than the threshold, the factor value decreases with the decrease of the data volume. This factor reflects the influence of historical data volume on credibility: the more sufficient the data volume is, the more correlation samples the method can learn, the more accurate the description of the vertical correlation rule is, and the higher the credibility is; when the data volume is insufficient, the sample learned by the method is limited, and the correlation rule may not be fully mastered, and the credibility is affected. Through the combination analysis of the above three factors, the user can clearly understand "why the prediction credibility of A target height is higher than that of B target height"; for example, the vertical distance factor of A target height is 0.8, the gradient stability factor is 0.9, and the corresponding factors of B target height are 0.5 and 0.6 respectively, which indicates that A target height has higher credibility due to "short distance + stable gradient", and provides a clear basis for the user to judge the applicable scene of the prediction result and the improvement direction of the method.

[0122] In one embodiment, the present application embodiment 2 provides a vertical atmospheric prediction credibility quantification system based on the same inventive concept. The implementation scheme of the system for solving the problem is similar to the implementation scheme described in the above embodiment method, so the specific limitations in one or more vertical atmospheric prediction credibility quantification system embodiments provided below can be referred to the limitations of the vertical atmospheric prediction credibility quantification method in the above, which will not be repeated here.

[0123] In one embodiment, the present application embodiment 2 provides a vertical atmospheric prediction credibility quantification system based on the same inventive concept. The implementation scheme of the system for solving the problem is similar to the implementation scheme described in the above embodiment method, so the specific limitations in one or more vertical atmospheric prediction credibility quantification system embodiments provided below can be referred to the limitations of the vertical atmospheric prediction credibility quantification method in the above, which will not be repeated here. Figure 2 As shown in the figure, the system comprises a dynamic construction module 210, a meteorological variable prediction module 220, an uncertainty quantification module 230 and a training and verification module 240, wherein:

[0124] The dynamic construction module 210 is configured to dynamically construct a vertical correlation graph based on the vertical correlation relationship between a single detection height and the target height for each target height to be predicted.

[0125] The meteorological variable prediction module 220 is configured to configure a vertical graph attention network for each target height, learn vertical correlation between a single sounding height and a corresponding target height through the vertical graph attention network, calculate an attention coefficient of a target node to a neighbor node, aggregate projection features of the neighbor node based on the attention coefficient to update a target prediction node feature, and output a meteorological variable prediction result of the target height in combination with a preset physical constraint function.

[0126] The uncertainty quantification module 230 is configured to determine a Monte Carlo Dropout sampling number based on vertical correlation strength between a sounding height and each target height, calculate an uncertainty of the meteorological variable prediction result of each target height, decompose the uncertainty into an inherent uncertainty and a cognitive uncertainty, and generate a general reliability index at the target height.

[0127] The training and verification module 240 is configured to use a historical data sample set containing different sounding heights and different numbers of target height combinations as an optimization target, perform parallel training and parameter optimization on all vertical graph attention networks based on errors between the meteorological variable prediction result and a true value, and verify the trained vertical graph attention networks.

[0128] Embodiment 3: Based on the technical principles of the above-mentioned reliability quantification method and system based on vertical atmospheric prediction, the third embodiment of the present application provides an electronic device. In an embodiment, the electronic device can be a terminal, and its internal structure diagram can be as shown in Figure 3 The electronic device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the electronic device is configured to perform wired or wireless communication with external terminals. The wireless communication can be achieved through WIFI, mobile cellular network, NFC (Near Field Communication), or other technologies. The computer program is executed by the processor to implement the reliability quantification method based on vertical atmospheric prediction in any one of steps S101 to S104. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball, or touchpad arranged on the shell of the electronic device. The input device can also be an external keyboard, touchpad, or mouse, etc.

[0129] Those skilled in the art can understand that Figure 3The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0130] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage etc.) containing computer-usable program code.

[0131] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by one or more blocks.

[0132] These computer program instructions can also be stored in a computer readable memory capable of directing the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction means, which implements the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0134] The above merely illustrates the embodiments of the present application, but should not be taken as limitations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall into the protection scope of the present application.

Claims

1. A reliable quantification method based on vertical atmospheric prediction, characterized in that, The method includes: For each target height to be predicted, a vertical correlation map is dynamically constructed based on the vertical correlation between a single detection height and the target height. A vertical graph attention network is configured for each target height. The vertical graph attention network learns the vertical correlation between a single detection height and the corresponding target height. The attention coefficient of the target node to its neighboring nodes is calculated. Based on the attention coefficient, the projection features of the neighboring nodes are aggregated to update the target prediction node features. The meteorological variable prediction result of the target height is output in combination with the preset physical constraint function. Based on the vertical correlation strength between the detection height and the height of each target, the number of Monte Carlo Dropout sampling times is determined, the uncertainty of the meteorological variable prediction results for each target height is calculated, and the uncertainty is decomposed into inherent uncertainty and cognitive uncertainty to generate a general credibility index at the target height. Using a historical data sample set containing different combinations of detection heights and different numbers of target heights, the error between the predicted results and the actual values ​​of the meteorological variables is used as the optimization objective. All vertical graph attention networks are trained and their parameters are optimized in parallel, and the trained vertical graph attention networks are validated. The meteorological variable prediction result of the target altitude combined with the preset physical constraint function includes: weighted summation of the projection features of each neighboring node according to the normalized attention coefficient to obtain the comprehensive feature that integrates multi-node information; linear transformation of the comprehensive feature through globally shared decoding weights and biases to output the meteorological variable prediction value of the target altitude; and the introduction of a physical constraint term; the physical constraint term includes temperature vertical gradient constraint and air pressure vertical decrease constraint. The predicted values ​​of meteorological variables are corrected to obtain the final predicted results of meteorological variables; The determination of the Monte Carlo Dropout sampling number based on the vertical correlation strength between the detection height and the height of each target includes: Calculate the average weight of all edges in the vertical association graph as the vertical association strength; The number of samplings required to determine the target height is determined by linear interpolation between a preset minimum and maximum value based on the vertical correlation strength.

2. The method according to claim 1, characterized in that, The vertical association graph includes a set of nodes and a set of edges; The node set includes: core detection nodes and neighboring nodes; and attribute data of the core detection nodes and neighboring nodes; The neighboring nodes include: target prediction nodes and auxiliary transition nodes; The edge set consists of connecting edges between adjacent nodes in the vertical direction, and the weight of the connecting edge is determined based on the vertical distance between the heights represented by the two nodes it connects and / or the stability of historical meteorological gradients. The attribute data of the core detection node includes: measured values ​​of temperature, relative humidity, air pressure, and wind speed, as well as their respective sensor error values; The attribute data of the target prediction node includes: historical vertical correlation features between the detection height and the target height; The historical vertical correlation features between the detection altitude and the target altitude include: the average differences in temperature, relative humidity, air pressure, and wind speed over historical periods, as well as the historical standard deviations of temperature difference and air pressure difference; The attribute data of the auxiliary transition node includes: historical average data and / or measured data for the corresponding height.

3. The method according to claim 2, characterized in that, The number of auxiliary transition nodes is determined based on predefined number determination rules, which include: Obtain the absolute vertical distance between the detection altitude and the current target altitude; When the absolute vertical distance is not greater than the first preset threshold, an auxiliary transition node is generated; When the absolute vertical distance is greater than the first preset threshold and not greater than the second preset threshold, two auxiliary transition nodes are generated. When the absolute vertical distance is greater than the second preset threshold, at least three auxiliary transition nodes are generated.

4. The method according to claim 2, characterized in that, The weight of the connecting edge is determined based on the vertical distance between the heights represented by the two nodes it connects and / or the stability of historical meteorological gradients, including: The weight of an edge connecting the core detection node and any auxiliary transition node is determined by the vertical distance between the heights represented by the two nodes. The weight of an edge connecting any auxiliary transition node to the target prediction node is determined by the vertical distance between the heights represented by the two nodes and the historical standard deviation of the temperature difference between the detected height and the target height.

5. The method according to claim 1, characterized in that, The attention coefficients of the target node to its neighboring nodes include: A linear transformation is performed on the projection features of the target prediction node and the projection features of any of its neighboring nodes, and the transformed features are then concatenated. The concatenated features are multiplied by the globally shared attention vector and then processed by an activation function. The processing result is multiplied by the edge weights from the neighboring nodes to the target prediction node to obtain the original attention score; The original attention scores of all neighboring nodes are exponentially calculated and summed. Then, the exponentialized score of each neighboring node is divided by this sum to obtain the normalized attention coefficient.

6. The method according to claim 1, characterized in that, The inherent uncertainty is calculated based on the sensor error at the detection altitude and the standard deviation of the historical meteorological variable differences between the detection altitude and the target altitude. The cognitive uncertainty is obtained by subtracting the variance of uncertainty from the variance of inherent uncertainty and taking the non-negative value.

7. A reliable quantification system based on vertical atmospheric prediction, characterized in that, The system includes: The dynamic construction module is used to dynamically construct a vertical correlation map for each target height to be predicted, based on the vertical correlation between a single detection height and the target height. The meteorological variable prediction module is used to configure a vertical graph attention network for each target height, learn the vertical correlation between a single detection height and the corresponding target height through the vertical graph attention network, calculate the attention coefficient of the target node to its neighboring nodes, aggregate the projection features of the neighboring nodes based on the attention coefficients to update the target prediction node features, and output the meteorological variable prediction results of the target height in combination with a preset physical constraint function. The meteorological variable prediction result of the target altitude combined with the preset physical constraint function includes: weighted summation of the projection features of each neighboring node according to the normalized attention coefficient to obtain the comprehensive feature that integrates multi-node information; linear transformation of the comprehensive feature through globally shared decoding weights and biases to output the meteorological variable prediction value of the target altitude; and the introduction of a physical constraint term; the physical constraint term includes temperature vertical gradient constraint and air pressure vertical decrease constraint. The predicted values ​​of meteorological variables are corrected to obtain the final predicted results of meteorological variables; The uncertainty quantification module is used to determine the number of Monte Carlo Dropout samplings based on the vertical correlation strength between the detection height and the height of each target, calculate the uncertainty of the meteorological variable prediction results for each target height, decompose the uncertainty into inherent uncertainty and cognitive uncertainty, and generate a general credibility index at the target height. The determination of the Monte Carlo Dropout sampling number based on the vertical correlation strength between the detection height and the height of each target includes: Calculate the average weight of all edges in the vertical association graph as the vertical association strength; Based on the vertical correlation strength, the number of samplings required to determine the target height is determined by linear interpolation between a preset minimum and maximum value; The training and validation module is used to perform parallel training and parameter optimization on all vertical graph attention networks using a historical data sample set containing different combinations of detection heights and different numbers of target heights, with the error between the predicted results and the actual values ​​of the meteorological variables as the optimization objective, and to validate the trained vertical graph attention networks.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.

Images