Virtual air station quality control data generation method based on spatiotemporal bayesian neural field

Abstract

The application discloses a virtual air station quality control data generation method based on a space-time Bayesian neural field, relates to the technical field of data processing, and can capture and enhance long-range space-time dynamics in discontinuous observation data, directly generate virtual station air quality results, namely, a target pollutant concentration set, under various missing data modes, and provide a confidence interval of a corresponding space-time point, so that the accuracy of air station quality control data can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a virtual air station quality control data generation method based on a spatiotemporal Bayesian neural field. BACKGROUND

[0002] Air quality data is a basic information resource for environmental regulation, urban governance, public health warning and scientific research. In order to ensure the reliability of management decision-making and public service, high spatiotemporal resolution and quality-controlled observation data must be obtained. The current ground observation system is mainly based on fixed monitoring stations supplemented by low-cost sensor vehicle-mounted mobile observation means. Fixed stations have high data quality and traceability, but are relatively sparse in spatial coverage; mobile observation can significantly improve spatial coverage density and fill in the gaps in fixed station observation and micro-environment observation, but the sensors of mobile observation equipment often face problems such as sensor drift, inconsistent calibration, disconnection and discontinuous reporting in actual operation, which makes the original observation unable to be directly used as compliant observation data.

[0003] In order to ensure the usability of mobile observation data, the industry often uses the comparison of mobile observation vehicles and fixed stations when the intersection event occurs to carry out on-site calibration, so as to adjust the bias of mobile measurement. This method can effectively correct local bias when there are enough intersection events, but its inherent limitations are also very obvious: intersection events are relatively rare in space and time, and many mobile observation points are outside the buffer zone of fixed stations, so a large amount of mobile observation data cannot be directly calibrated by intersection. In addition, the calibration based on intersection usually only gives point estimate correction value, and lacks quantification of observation uncertainty after correction, which is difficult to meet the industry's requirements for data traceability, confidence statement and direct reporting compatibility. SUMMARY

[0004] Therefore, the embodiments of the present application provide a virtual air station quality control data generation method, device, electronic equipment and storage medium based on a spatiotemporal Bayesian neural field, so as to improve the accuracy of air station quality control data.

[0005] An aspect of the embodiments of the present application provides a virtual air station quality control data generation method based on a spatiotemporal Bayesian neural field, which comprises the following steps:

[0006] The original data of the target pollutant concentration and the corresponding original spatiotemporal features are brought into a data preprocessing module, and are respectively subjected to sample filtering, time discretization and feature normalization to obtain a normalized original feature matrix;

[0007] The original feature matrix is input into a multi-layer spatio-temporal feature encoding module, and is subjected to spatio-temporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation respectively to obtain spatio-temporal interaction feature, air-air interaction feature, time seasonal feature, spatial Fourier feature and spatial aggregation feature, and then the obtained features are spliced with the original feature matrix to obtain a multi-layer spatio-temporal feature matrix;

[0008] The multi-layer spatio-temporal feature matrix is input into an adaptive scaling and channel attention process in the multi-layer spatio-temporal feature encoding module for processing to form a high-dimensional feature matrix with weight redistribution;

[0009] The high-dimensional feature matrix is input into a feature enhancement and mean prediction module, and is subjected to feature enhancement by three layers of channel gate learning units respectively, and then is input into a fully connected layer network to generate a conditional mean set of the target pollutant concentration prediction;

[0010] The conditional mean set is input into a Bayesian probability prediction module, and the original spatio-temporal features, the original data of the target pollutant concentration and the conditional mean set are used together to restore the spatio-temporal distribution of the target pollutant, and a target pollutant concentration set and a corresponding variance set are generated according to the spatio-temporal distribution of the target pollutant.

[0011] In some embodiments, the spatio-temporal interaction feature is obtained by the following steps:

[0012] The original feature matrix is multiplied by a timestamp set and an observation site longitude set and an observation site latitude set respectively to obtain the spatio-temporal interaction feature :

[0013] ;

[0014] wherein, represents an element-wise product, represents vector or matrix splicing;

[0015] The air-air interaction feature is obtained by the following steps:

[0016] The observation site longitude set and the observation site latitude set in the original feature matrix are multiplied to obtain air-air interaction :

[0017] ;

[0018] The time seasonality term feature is obtained by the following steps:

[0019] The time seasonality term feature is obtained by the following steps:

[0020] The space Fourier term feature is obtained by the following steps:

[0021] The space Fourier term feature is obtained by the following steps: The space Fourier term feature is obtained by the following steps:

[0022] The space aggregation term feature is obtained by the following steps:

[0023] The space aggregation term feature is obtained by the following steps:

[0024] In some embodiments, the adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps:

[0025] The adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps: The adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps:

[0026] The adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps: The adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps:

[0027]

[0028] is a set of learnable scaling coefficients;

[0029] The adaptive scaling and channel attention process is used to process the multi-level spatio-temporal feature matrix in the multi-level spatio-temporal feature encoding module, forming a high-dimensional feature matrix with weight redistribution, including the following steps:

[0030] ​​​​​​​​​​

[0031] wherein, is the current batch index; is the total number of batches; is the feature matrix after global average pooling;

[0032] The channel attention gate vector is generated by using a two-layer fully connected network to capture the inter-channel nonlinear dependence through the following expression :

[0033] ;

[0034] wherein, and are the weight matrices of the two-layer fully connected network in the channel attention layer, and are the bias vectors of the two-layer fully connected network in the channel attention layer;

[0035] The channel attention gate vector is multiplied by the scaled feature matrix to obtain the high-dimensional feature matrix with weight redistribution:

[0036] .

[0037] In some embodiments, the high-dimensional feature matrix sequentially passes through the gate residual network, the learnable activation function, and the channel attention in the channel gate learning unit of each layer;

[0038] The gate residual network refers to robustly regulating the information flow of the feature matrix through the gate residual network, and specifically includes:

[0039] First, set as the input feature matrix of the th layer channel gate learning unit;

[0040] Then, the intermediate feature matrix of the th gate unit is calculated by linear transformation according to the following expression:

[0041] ;

[0042] wherein, and represent the linear transformation weight matrix of the th layer network, respectively; and are the bias vectors of the linear projection of the th layer network, respectively; is the hierarchical index of the channel gate learning unit.

[0043] Next, calculate the first according to the following expression. Layer gate vector Used for dynamically fusing original and enhanced features:

[0044] ;

[0045] in, and These represent the weight matrix and bias vector that generate the gate vector, respectively;

[0046] Finally, by utilizing the gating mechanism and the hybrid activation function in combination according to the following expression, the first... Layer-gated residual output vector :

[0047] ;

[0048] The learnable activation function refers to replacing the single activation function in the original gated residual network with a trainable Elu–Tanh convex hybrid activation function. Elu retains negative responses and accelerates convergence, while Tanh provides smooth and bounded outputs, thereby improving the fitting ability and uncertainty quantification capability of complex spatiotemporal patterns. Specifically, this includes:

[0049] The learnable activation function is defined by the following expression:

[0050] ;

[0051] Among them, the Learnable Hybrid Activation Functions of Layers Through the first Learnable activation blending coefficients of layers The ratio of the two activations is adaptively adjusted during training. is the input to the hybrid activation function, and Tanh is the activation function.

[0052] In some embodiments, the process of accessing the fully connected layer network to generate the conditional mean set of the target pollutant concentration prediction includes the following steps:

[0053] The conditional mean set is generated using a lightweight fully connected network via the following expression:

[0054] ;

[0055] in, Represents the set of conditional means. Indicates after the first The layer can learn the output features after processing by the activation function. and These represent the weight matrix and bias vector of the mean prediction network, respectively.

[0056] In some embodiments, the step of inputting the conditional mean set into the Bayesian probability prediction module and using the original spatiotemporal features, the original data of the target pollutant concentration, and the conditional mean set to jointly reconstruct the spatiotemporal distribution of the target pollutant includes the following steps:

[0057] The concentration distribution model of the target pollutant is defined as a Gaussian likelihood distribution, and a Logistic(0,1) prior distribution is independently set for each parameter in the parameter space.

[0058] The parameters of the Gaussian likelihood distribution are optimized by repeatedly using the maximum probability posterior method starting from random initialization points to obtain multiple local optimal solutions for the parameters;

[0059] The predicted distribution components are constructed using the local optimal solutions of each parameter, and all the predicted distribution components are mixed with equal weights to obtain the posterior predicted distribution as the spatiotemporal distribution of the target pollutant.

[0060] The Gaussian likelihood distribution is expressed as follows:

[0061] ;

[0062] in, The standard deviation of the observed noise, To collect all network parameters, It is the number of entries observed in the training loss. for The identity matrix;

[0063] The optimization objective of the maximum probability posterior method is expressed as:

[0064] ;

[0065] in, Represents the likelihood function; This is the posterior optimal solution for the parameters; This represents the estimated standard deviation of the observed noise in the output distribution during the prediction phase;

[0066] The posterior prediction distribution is represented as:

[0067] ;

[0068] in, Optimize the number of iterations to maximize the posterior probability; and The first a predictive mean of a posteriori modal and an observation noise variance; a likelihood model representing the observation data is Gaussian normal distribution; an input feature vector representing a prediction point; a predictive target output of a model to an input ; a training data set.

[0069] In some embodiments, the generating a target pollutant concentration set and a corresponding variance set according to the spatiotemporal distribution of the target pollutant comprises the following steps:

[0070] The generating the target pollutant concentration set and the variance set according to the spatiotemporal distribution of the target pollutant respectively comprises:

[0071] ;

[0072] wherein, the target pollutant concentration set is represented by the variance set is represented by

[0073] Another aspect of the embodiments of the present application also provides a virtual air station quality control data generation device based on a spatiotemporal Bayesian neural field, comprising:

[0074] a data preprocessing unit configured to input raw data of a target pollutant concentration and corresponding raw spatiotemporal features into a data preprocessing module, and obtain a normalized raw feature matrix after sample filtering, time discretization and feature normalization respectively;

[0075] a feature encoding unit configured to input the raw feature matrix into a multi-layer spatiotemporal feature encoding module, and obtain spatiotemporal interaction term features, space-space interaction term features, time seasonality term features, space Fourier term features and space aggregation term features after spatiotemporal interaction, time harmonic expansion, space harmonic expansion and space correlation respectively, and then splice the obtained features with the raw feature matrix to obtain a multi-layer spatiotemporal feature matrix;

[0076] a weight distribution unit configured to input the multi-layer spatiotemporal feature matrix into an adaptive scaling and channel attention process in the multi-layer spatiotemporal feature encoding module for processing to form a high-dimensional feature matrix with weight distribution;

[0077] a mean prediction unit configured to input the high-dimensional feature matrix into a feature enhancement and mean prediction module, and perform feature enhancement through three-layer channel gate learning units respectively, and then input into a fully connected layer network to generate a conditional mean set of the target pollutant concentration prediction;

[0078] A distribution restoring unit is configured to bring the conditional mean set into a Bayesian probability prediction module, and restore the spatial-temporal distribution of the target pollutant by using the original spatial-temporal feature, the original data of the target pollutant concentration, and the conditional mean set, and generate a target pollutant concentration set and a corresponding variance set according to the spatial-temporal distribution of the target pollutant.

[0079] Another aspect of the embodiments of the present application further provides an electronic device, comprising a processor and a memory;

[0080] The memory is configured to store a program;

[0081] The processor executes the program to implement the method according to any one of the preceding aspects.

[0082] Another aspect of the embodiments of the present application further provides a computer readable storage medium, which stores a program, and the program is executed by a processor to implement the method according to any one of the preceding aspects.

[0083] The present application at least has the following beneficial effects:

[0084] The present application captures multi-scale spatial dependence and seasonal time dynamics through a multi-layer spatial-temporal feature encoding module, filters and enhances information features adaptively by using a feature enhancement and mean prediction module, and jointly optimizes data generation distribution and parameter uncertainty by using a Bayesian probability prediction module, so as to capture and enhance long-range spatial-temporal dynamics in discontinuous observation data, directly generate virtual station air quality results, i.e., a target pollutant concentration set, under multiple missing data modes, and provide a confidence interval of a corresponding spatial-temporal point, thereby improving the accuracy of air station quality control data. BRIEF DESCRIPTION OF DRAWINGS

[0085] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0086] Figure 1 A flowchart of a virtual air station quality control data generation method based on a spatial-temporal Bayesian neural field according to an embodiment of the present application is shown in FIG. 1;

[0087] Figure 2 An example flowchart of a virtual air station quality control data generation method based on a spatial-temporal Bayesian neural field according to an embodiment of the present application is shown in FIG. 2;

[0088] Figure 3 A technical framework diagram of a virtual air station quality control data generation method based on a spatial-temporal Bayesian neural field according to an embodiment of the present application is shown in FIG. 3.

[0089] Figure 4 PM provided for the embodiments of the present application 10 Inference result timing display diagram;

[0090] Figure 5 The structural block diagram of the virtual air station quality control data generation device based on the spatio-temporal Bayesian neural field provided for the embodiments of the present application. DETAILED DESCRIPTION

[0091] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0092] Before the embodiments of the present application are described in detail, first, some related technologies involved in the embodiments of the present application are described as follows:

[0093] Based on the status and deficiencies of the prior art, the idea of constructing a "virtual ground air station" on the spatio-temporal point of mobile observation is proposed: that is, inferring an air quality value of a virtual fixed station at the spatio-temporal position of the mobile observation, and using the inferred value of the virtual station as a "pseudo true value" for calibrating the mobile observation data, especially for a large number of observation points located outside the buffer area of the fixed station. In order to realize this idea, there is an urgent need for a virtual ground air station quality control data generation scheme, which can give the corresponding air quality inference value for any virtual station under the scene of discontinuous air quality observation data, and quantize the confidence interval of the generation result.

[0094] In order to meet the above needs, two-stage framework and end-to-end framework are proposed. The two-stage framework first completes the missing data, and then trains the generation model based on the complete data set. However, this direct coupling has limitations: existing interpolation methods are difficult to learn the fine spatio-temporal features required for high-precision air quality prediction, and systematic research on how interpolation accuracy affects generation performance is still relatively scarce. The end-to-end framework optimizes feature extraction, data completion and target generation in a single framework. Gaussian process method is the most used end-to-end method, which combines the expression ability of deep network with the closed-form interpolation and uncertainty quantification of probability method. However, there are two difficulties in applying this method to air quality generation. First, the posterior inference needs to bear a huge computational cost. Second, the selection of key parameters (such as covariance kernel function and mean function) is highly dependent on domain expert knowledge. Therefore, under the condition of discontinuous observation data, how to design a generation method that can ensure high-precision reliable uncertainty quantification and flexibly handle different degrees of missing data is still an important research topic to be tackled.

[0095] ​A search of existing technologies revealed that Chinese patent document CN116542551A (application date 2023-04-13) discloses an air quality prediction method based on incomplete data. This method addresses scenarios with incomplete observations, first categorizing historical monitoring data by "monitoring point..." × Time × Types of pollutants Organized into a three-dimensional tensor and using an observation index set Mark observable elements; for predicting the future Heaven, introducing the time sliding window dimension Expand the data into a four-dimensional tensor, and according to A corresponding weight tensor is constructed to handle the imbalance between missing and known values. Then, CP decomposition is performed on the four-dimensional tensor to model the imbalance and establish an objective function with a regularization term. The method captures the correlations between spatial (site), temporal (including periodic sliding window), and categorical (multi-pollutant) factors by alternately minimizing the factor matrix; finally, it reconstructs the unknown elements of the tensor using the estimated factor matrix, and outputs the result. to This method provides predicted values ​​for each pollutant at each monitoring station. It addresses the problem of air quality forecasting based on incomplete data.

[0096] However, compared with this application, the method has the following technical limitations: (1) It only gives deterministic point inference values ​​and lacks the uncertainty characterization of the generated results. This method constructs a weight tensor and solves the factor matrix by alternating minimization, and directly gives the point estimate of "the predicted value of the m-th pollutant at the p-th monitoring point on day t". It does not provide quantitative results such as interval, variance or confidence level, and cannot assess the reliability and risk propagation of the generated results. In contrast, this application synchronously outputs the generated air quality values ​​of the points and the confidence intervals of their corresponding generated results in an end-to-end framework, and supports the evaluation of the generated results by the average interval width and the relative average interval width index. (2) It lacks the adaptive ability of the model under different data missing modes. This method uses the observation index set The generation of the weight tensor and the fitting of the factor matrix under the Frobenius norm essentially focus on the "complement-fitting" paradigm. When the missing presents node-level, timestamp-level or spatio-temporal block-level structured missing, the method does not provide additional robustness or uncertainty characterization means. Without prior interpolation, the present application directly learns and infers under any missing pattern (random missing, node missing, timestamp missing, spatio-temporal block missing), and through Bayesianized parameters, noise modeling and structure prior, explicitly incorporates the information incompleteness caused by missing into uncertainty propagation. The time modeling is constrained by a fixed window, and the periodicity representation ability is not strong. The method fixes the time dimension as a multiple of 7, and accordingly maps and expands, which is equivalent to imposing a prior "weekly period" structure on the time dependence. The representation of complex periodicity with multiple periodic scales coexisting, drifting or misaligning is not flexible enough. The present application captures the superimposed periods of hours, days, weeks, months, quarters and even years through the harmonic and Fourier mechanism, and through the gating and channel attention structure, adaptively reweights the different time scale components, breaks through the limitation of the fixed week window, and improves the generalization ability under the condition of non-stationary and misaligned periods.

[0097] Referring to Figure 1 , the embodiment of the present application provides a virtual air station quality control data generation method based on a spatio-temporal Bayesian neural field, which specifically comprises the following steps S100-S140:

[0098] S100: The original data of the target pollutant concentration and the corresponding original spatio-temporal features are brought into a data preprocessing module, and are respectively subjected to sample filtering, time discretization and feature normalization to obtain a normalized original feature matrix;

[0099] S110: The original feature matrix is brought into a multi-layer spatio-temporal feature coding module, and is respectively subjected to spatio-temporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation to obtain spatio-temporal interaction item features, space-space interaction item features, time seasonality item features, spatial Fourier item features and spatial aggregation item features. Then, the obtained various features are spliced with the original feature matrix to obtain a multi-layer spatio-temporal feature matrix;

[0100] S120: The multi-level spatio-temporal feature matrix is input into an adaptive scaling and channel attention process in the multi-level spatio-temporal feature encoding module for processing to form a high-dimensional feature matrix with weight redistribution;

[0101] S130: The high-dimensional feature matrix is input into a feature enhancement and mean prediction module, and is subjected to feature enhancement through three layers of channel gate learning units, and then is input into a fully connected layer network to generate a conditional mean set for prediction of the target pollutant concentration;

[0102] S140: The conditional mean set is input into a Bayesian probability prediction module, and the original spatio-temporal features, the original data of the target pollutant concentration, and the conditional mean set are used together to restore the spatio-temporal distribution of the target pollutant, and a target pollutant concentration set and a corresponding variance set are generated according to the spatio-temporal distribution of the target pollutant.

[0103] Next, the scheme of the embodiments of the present application will be described in detail with reference to specific application examples.

[0104] The purpose of the present embodiment is to overcome the shortcomings of the prior art and propose a virtual air station quality control data generation method based on a spatio-temporal Bayesian neural field. The method captures multi-scale spatial dependence and seasonal time dynamics through a multi-level spatio-temporal feature encoding layer, adaptively filters and enhances information features using a feature enhancement and mean prediction layer, and jointly optimizes data generation distribution and parameter uncertainty using a Bayesian probability prediction layer. The present embodiment can capture and enhance long-range spatio-temporal dynamics in discontinuous observation data, directly generate virtual station air quality results under various missing data patterns, and provide confidence intervals for corresponding spatio-temporal points.

[0105] To achieve the above, the present embodiment adopts the quality control data generation process and specific technical scheme as shown in Figure 2 and Figure 3 The present embodiment method includes four core parts: (1) a data preprocessing module. This module performs time discretization and feature normalization on discontinuous data to form a standardized data input form. (2) A multi-level spatio-temporal feature encoding module. This module integrates time harmonic, spatial Fourier embedding, and graph attention mechanism technology to capture seasonal cycles and cross-site dependence under discontinuous spatio-temporal observation data. (3) A feature enhancement and mean prediction module. This module uses multiple layers of channel gate units to adaptively filter and enhance spatio-temporal features and map them to conditional means. (4) A Bayesian probability prediction module. This module estimates data distribution and generates virtual station quality control data generation results with confidence intervals through particle-based posterior maximum likelihood inference.

[0106] Specifically, the present embodiment includes the following steps:

[0107] Step 1: The collected target pollutant concentration raw data and its corresponding raw spatiotemporal features are input into the data preprocessing module, and after sample filtering, time discretization, and feature normalization, the normalized original feature matrix is obtained . The information of the raw spatiotemporal features includes the spatiotemporal information (time, station longitude, and station latitude) of all observation stations and the exogenous variable data (the concentration data of other pollutant species collected at the same time and same place).

[0108] Step 2: The original feature matrix obtained in Step 1 is input into the multi-level spatiotemporal feature encoding module, and after spatiotemporal interaction, time harmonic expansion, spatial harmonic expansion, and spatial correlation construction, the obtained spatiotemporal interaction feature , air-air interaction feature , time seasonal feature , spatial Fourier feature , and spatial aggregation feature are concatenated with the original feature matrix to form a multi-level spatiotemporal feature matrix . Then, the obtained multi-level spatiotemporal feature matrix is input into the adaptive scaling and channel attention process, and finally a high-dimensional feature matrix with weight redistribution is formed .

[0109] Step 3: The high-dimensional feature matrix obtained in Step 2 is input into the feature enhancement and mean prediction module, and after three layers of channel gate learning units for feature enhancement, a fully connected layer network is connected to generate a set of conditional mean values for pollutant concentration prediction . In each layer of channel learning gate unit, the feature matrix sequentially passes through a gated residual network, a learnable activation function, and a channel attention part to maximize the enhancement of key feature information.

[0110] Step 4: The set of conditional mean values predicted in Step 3 is input into the Bayesian probability prediction module, and the raw spatiotemporal features , the target pollutant concentration raw data , and the set of predicted conditional mean values are used to restore the spatiotemporal distribution trend of the pollutant, and based on the restored spatiotemporal distribution, a final set of pollutant concentration and a corresponding set of variances Specifically, firstly, the pollutant concentration distribution model is defined as a Gaussian likelihood distribution form, and a Logistic(0, 1) prior distribution is independently set for each parameter in the parameter space. Then, the defined conditional prediction probability model parameters are optimized by the maximum posterior probability method starting from multiple random initialization points, and multiple local optimal solutions of the parameters are obtained. Next, the prediction distribution components are respectively constructed by using each parameter solution, and all components are equally mixed to obtain the posterior prediction distribution. Finally, based on the obtained posterior prediction distribution, the pollutant concentration set and the corresponding variance set are generated.

[0111] Further, the target pollutant concentration original data shown in step 1 refers to the target pollutant concentration data collected at all observation sites within a certain historical time period. The site space information refers to the longitude and latitude coordinate information of all observation sites and all recorded time nodes corresponding to the observation site when collecting the pollutant concentration. The exogenous variable data refers to the concentration data of the remaining pollutant species collected at the same time and the same place as the observation site when collecting the target pollutant concentration.

[0112] Further, the sample filtering shown in step 1 refers to eliminating invalid sample records lacking target values. Time discretization refers to mapping an arbitrary timestamp to an integer index by calculating the offset relative to the reference time , and offsetting the minimum value to zero, as shown in formula 1. Feature normalization refers to normalizing the remaining feature columns except time by using the Z-score method. Wherein, index() represents the "discretization index mapping function", which maps time from the actual time axis to the integer axis.

[0113] (1)

[0114] Further, the space-time interaction process shown in step 2 refers to multiplying the timestamp set in the normalized original feature matrix with the observation site longitude set and the observation site latitude set respectively, so as to obtain the space-time interaction term , as shown in formula 2. Wherein, represents the element-wise product, represents vector or matrix splicing.

[0115] (2)

[0116] Further, the air-to-air interaction flow shown in step 2 refers to multiplying the longitude set and the latitude set after normalization, so as to obtain an air-to-air interaction term , as shown in formula 3.

[0117] (3)

[0118] Further, the time harmonic expansion flow shown in step 2 refers to using an orthogonal sine-cosine pair as a fixed basis function, which is smoothly integrated into gradient training, so as to capture multiple cycles (supporting minute, hour, day, week, month, quarter and year cycles) and form a time seasonal term feature . Specifically, let be a base cycle set, and for any time cycle in the base cycle set, define its corresponding harmonic basis function set (satisfying ). Wherein, is each cycle element in the base cycle set ; is the set of all harmonic basis functions generated under the cycle ; is the maximum harmonic order corresponding to the time cycle .

[0119] Under the remapped integer index , the time harmonic feature vector generated for the harmonic component of the time cycle can be obtained by formula 4. Subsequently, all harmonics of all base cycles are concatenated by channel to form a seasonal feature term . Wherein, is the harmonic order index.

[0120] (4)

[0121] Further, the spatial harmonic expansion flow shown in step 2 refers to mapping the normalized longitude and latitude using Fourier features to form a spatial Fourier term feature . Specifically, for a spatial coordinate channel index , define the maximum harmonic order of the axis, and concatenate the normalized axis coordinates​ Fourier feature vector mapped to the th spatial axis , as shown in equation 5. Subsequently, all spatial mappings are concatenated by channel to form the spatial Fourier term . Here, is the spatial harmonic order index.

[0122] (5)

[0123] Further, the spatial correlation construction process shown in step 2 refers to learning the field-specific non-stationary spatial correlation relationship from the data using a single-layer multi-head graph attention network to form the spatial aggregation term feature . Specifically, for any observation site index , according to the historical pollutant concentration data collected by the site, the average value, the 25th percentile and the 75th percentile statistical indicators of the pollutant concentration of the site are obtained, and an initial node feature vector of a static site is constructed using the statistical indicators. The graph attention network with heads only propagates information along the non-zero edges in the adjacency matrix , and its attention weight is modulated by a Gaussian connection prior. This attention mechanism can be represented by equation 6.

[0124] (6)

[0125] where, is the number of attention heads; and represent any two adjacent site indices; is the attention kernel vector of the th attention head; is the linear projection matrix of the th attention head; is the input feature vector of the adjacent node; is the unnormalized attention coefficient under the th attention head; is the Euclidean distance between node and node ; is a scale parameter to control the distance decay; is the weighted attention score; is the normalized attention weight; is the set of observation sites; ReLU, softmax, Elu are different activation functions.

[0126] (7)

[0127] The output result of attention is concatenated and linearly projected according to formula 7 to form the node aggregated feature output vector of the node . Wherein, represents the output projection matrix, and concat represents the feature matrix concatenation. Finally, stack along the time dimension to obtain the spatial aggregated feature .

[0128] Further, the adaptive scaling and channel attention process shown in step 2 refers to automatic adjustment of the amplitude of each feature, amplification of key information and suppression of redundancy of the multi-level spatio-temporal feature matrix , and finally forms a high-dimensional feature matrix with weight redistribution . Specifically, for the multi-level spatio-temporal feature matrix , first, according to formula 8, each channel is scaled by using in each batch to realize automatic rescaling of feature scale. Wherein, is a set of learnable scaling coefficients.

[0129] (8)

[0130] Then, according to formula 9, the scaled feature matrix is globally averaged to extract channel-level statistics and eliminate spatial position interference. Wherein, is the current batch index; is the total number of batches; is the feature matrix after global average pooling.

[0131] (9)

[0132] Next, according to formula 10, a two-layer fully connected network is used to capture the nonlinear dependence between channels to generate a channel attention gating vector . Wherein, and are the weight matrices of the two-layer fully connected network in the channel attention layer, and are the bias vectors of the two-layer fully connected network in the channel attention layer.

[0133] (10)

[0134] Finally, according to formula 11, the channel attention gating vector is multiplied by the scaled feature to obtain a high-dimensional feature matrix with weight redistribution :

[0135] (11)

[0136] Further, the gated residual network module shown in step 3 refers to the robust regulation of information flow of the feature matrix through the gated residual network. Specifically, first set the input feature matrix of the first layer channel gating learning unit. Then, according to formula 12, the intermediate feature matrix of the first layer gating unit is calculated by linear transformation . Wherein, and represent the linear transformation weight matrix of the first layer network respectively; and are the bias vectors of the linear projection of the first layer network respectively;

[0137] (12)

[0138] Next, according to formula 13, the first layer gating vector is calculated, which is used to dynamically fuse the original and enhanced features. Wherein, and represent the weight matrix and bias vector of the generated gating vector respectively.

[0139] (13)

[0140] Finally, according to formula 14, the first layer gated residual output vector is generated by using the gating mechanism and the mixed activation function together.

[0141] (14)

[0142] Further, the learnable activation function shown in step 3 refers to replacing the single activation function in the original gated residual network with a trainable Elu-Tanh convex combination mixed activation function, i.e. Elu retains negative response and accelerates convergence, and Tanh provides smooth and bounded output, thereby improving the fitting of complex spatiotemporal patterns and the ability to quantify uncertainty. Specifically, the learnable activation function can be defined by formula 15. Wherein, the learnable mixed activation function of the first layer is calculated by the learnable activation mixing coefficient of the first layer. The proportion of the two activations is adjusted adaptively in training, is the input of the mixed activation function module, and Tanh is the activation function.

[0143] (15)

[0144] Further, the condition mean value shown in step 3 is generated by a lightweight fully connected network. Specifically, denotes the set of concentration condition mean values, denotes the output feature processed by the layer learnable activation function, denotes the output feature processed by the layer learnable activation function, and denote the weight matrix and bias vector of the mean prediction network, respectively.

[0145] (16)

[0146] Further, the Gaussian likelihood distribution form shown in step 4 can be represented by formula 17. Wherein, denotes the standard deviation of the observation noise, is the collection of all network parameters (scaling factor, network weight, bias, and activation combination coefficient), is the number of observed entries in the training loss, is the identity matrix of .

[0147] (17)

[0148] Further, the parameter estimation method based on maximum a posteriori probability shown in step 4, the optimization objective can be represented by formula 18. Wherein, denotes the likelihood function; is the posterior optimal solution of the parameter; denotes the estimated standard deviation of the observation noise in the output distribution in the prediction stage.

[0149] (18)

[0150] Further, the posterior prediction distribution form shown in step 4 can be represented by formula 19. Wherein, is the number of maximum a posteriori probability optimization times; and are the prediction mean and observation noise variance of the th posterior mode, respectively; denotes that the likelihood model of the observation data is Gaussian normal distribution; denotes the input feature vector of the prediction point; denotes the predicted target output of the model to the input ; is the training data set.

[0151] (19)

[0152] Further, the final generated output mean and variance shown in step 4 can be obtained by formula 20.

[0153] (20)

[0154] To sum up, the embodiment includes the following key technical solutions:

[0155] 1. A virtual air station quality control data generation method based on a spatiotemporal Bayesian neural field is provided, which includes the following four core steps: (1) a data preprocessing module. This module performs time discretization and feature normalization processing on discontinuous data to form a standardized data input form. (2) A multi-layer spatiotemporal feature encoding module. This module integrates time harmonic, spatial Fourier embedding and graph attention mechanism technology to capture seasonal periodicity and cross-site dependence under discontinuous spatiotemporal observation data. (3) A feature enhancement and mean prediction module. This module uses a multi-layer channel gating unit to perform adaptive filtering and enhancement on spatiotemporal features and map them to conditional means. (4) A Bayesian probability prediction module. This module estimates the data distribution and generates results with confidence intervals through particle-based posterior maximum likelihood inference.

[0156] 2. In step (1), the collected target pollutant concentration raw data and its corresponding spatiotemporal feature information will be brought into the data preprocessing module, and after sample filtering, time discretization and feature normalization processes, the model standardized feature matrix will be obtained. Among them, the spatiotemporal feature information includes the spatiotemporal information (time, station longitude and station latitude) of all observation stations, and exogenous variable data (other types of pollutant concentrations collected synchronously).

[0157] 3. In step (2), the obtained feature matrix will be brought into the multi-layer spatiotemporal feature encoding module, and after the spatiotemporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation construction processes, the obtained spatiotemporal interaction feature , space-space interaction feature , time seasonality feature , spatial Fourier feature , spatial aggregation feature and the original feature matrix will be spliced to form a multi-level spatiotemporal feature matrix .

[0158] 4. In step (2), the obtained multi-level spatiotemporal feature matrix will be brought into the adaptive scaling and channel attention process to form a high-dimensional feature tensor with weight redistribution .

[0159] 5、Step (3), the obtained high-dimensional feature tensor will be brought into the feature enhancement and mean prediction module, respectively through three layers of channel gate learning units for feature enhancement, and then connected to a fully connected layer network to generate a set of conditional mean values of pollutant concentration prediction .

[0160] 6、Step (3), each layer of channel learning gate unit contains a gated residual network, a learnable activation function and a channel attention. The high-dimensional feature matrix will sequentially pass through the three parts of each layer of channel learning gate unit to maximize the enhancement of key feature information. Among them, the learnable activation function is a convex combination of two separate activation functions Elu and Tanh.

[0161] 7、Step (4), the set of conditional mean values predicted will be brought into the Bayesian probability prediction module, using the original spatiotemporal features , the original data of the target pollutant concentration and the set of conditional mean values predicted to restore the spatiotemporal distribution trend of the pollutant concentration, and based on the restored spatiotemporal distribution to generate the final set of pollutant concentration and the corresponding set of variances .

[0162] 8、Step (4), the specific steps for restoring the spatiotemporal distribution trend of the pollutant concentration are as follows: first, define the pollutant concentration distribution model as a Gaussian likelihood distribution, and independently set the Logistic(0,1) prior distribution for each parameter in the parameter space. Then, use the maximum posterior probability method starting from multiple randomly initialized points to optimize the parameters of the defined conditional prediction probability model, obtaining multiple local optimal solutions for the parameters. Next, use each parameter solution to construct a prediction distribution component, and mix all components with equal weight to obtain the posterior prediction distribution. Finally, based on the obtained posterior prediction distribution, the set of pollutant concentration and the corresponding set of variances are obtained.

[0163] The embodiment provides a virtual air station quality control data generation method based on a spatio-temporal Bayesian neural field, which unifies feature extraction, target generation and uncertainty quantification in the same system, can realize direct inference under four common data missing modes of random missing, node missing, timestamp missing and spatio-temporal block missing, avoids the additional steps of relying on interpolation or filling in the traditional method, and thus improves the processing efficiency and data generation reliability. In addition, the embodiment fuses a graph attention mechanism, a time series harmonic decomposition and a spatial Fourier embedding, constructs a multi-scale spatio-temporal encoder, can effectively capture cross-scale spatio-temporal dependence and correlation under irregular sampling conditions; meanwhile, a channel gating learning unit is introduced, a channel attention mechanism, a gated residual structure and a learnable activation function are combined, dynamic screening and enhancement of key information channels are realized, noise is effectively suppressed and the stability of the training process is maintained. Experimental results show that the embodiment has lower inference error on a real air quality data set, generates narrower inference results, takes into account the inference accuracy and reliability, and can provide reliable and efficient technical support for virtual ground air station quality control data generation and related applications.

[0164] The following is example data for verifying the embodiment:

[0165] The embodiment selects a large-scale discontinuous air quality spatio-temporal data set of a certain region to generate PM 10 concentration. The data set contains collection time, site longitude, site latitude, PM 10 concentration basic information and SO2, NO2, O3, PM 2.5 four kinds of exogenous covariate information.

[0166] Table 1 Basic information of the data set

[0167]

[0168] Table 2 Example data of the data set

[0169]

[0170] To evaluate the robustness of the present embodiment under different degrees of data missing, four common missing data scenarios are simulated: (i) random missing, i.e., intermittent loss of values in the data stream (e.g., data packet loss when the communication quality temporarily decreases); (ii) node missing, i.e., all observations of a single node are missing for a long time (e.g., continuous sensor failure); (iii) timestamp missing, all nodes cannot obtain data at a certain time point at the same time (e.g., local power failure); (iv) spatiotemporal block missing, the missing data forms a continuous spatiotemporal block (e.g., a mobile sensor passing through a tunnel produces a persistent blind area). For each scenario, the missing rate is selected as the classic missing rate of 30%. The target value will be removed according to the specified missing mode. In cross-validation, the observation sites are randomly divided into five mutually exclusive subsets. In each fold, the latest one month of records of the reserved subset sites constitutes the test set. The training set contains all the remaining data: the observation data of other sites throughout the cycle and the non-test period data of the reserved subset sites.

[0171] All experiments are performed on a server equipped with an Intel Xeon Gold 6133 CPU and four NVIDIA GeForce RTX 4090 graphics cards. The generation method of the present embodiment consists of three stacked channel gating learning layers, each containing 512 hidden units and 16 particles. The AdamW optimizer is used for training, with an initial learning rate of , a batch size of 512, and a maximum of 5000 training rounds. The hyperparameters are optimized in preliminary experiments, and the contribution of each module is evaluated through ablation studies. Under the same settings, the present embodiment method and all baseline models are trained and validated on five non-overlapping partition sets of each data set, and the average performance indicators are reported to achieve a fair comparison under the incomplete observation scenario. Among them, the baseline methods select four models: historical average (HA), random forest (RF), spatiotemporal gradient boosting tree (STGBOOST), and spatiotemporal sparse variational Gaussian process method (ST-SVGP): the data generation performance evaluation indicators select six kinds of root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ), symmetric mean absolute percentage error (SMAPE), average interval width (AIW), and average relative interval width (RIWM).

[0172] The present embodiment method is used to infer PM 10 concentration under the original data, random missing, node missing, timestamp missing, and spatiotemporal block missing scenarios, respectively. The virtual station quality control data generation results evaluation indicators are shown in Tables 3-7.

[0173] Table 3 PM 10 inference results under the original data scenario

[0174]

[0175] Table 4 PM under random missing scenarios 10 Inference results

[0176]

[0177] Table 5 PM under node missing scenarios 10 Inference results

[0178]

[0179] Table 6 PM under timestamp missing scenarios 10 Inference results

[0180]

[0181] Table 7 PM under spatio-temporal block missing scenarios 10 Inference results

[0182]

[0183] From the data in the above tables, it can be seen that the generation method of the embodiment is significantly better than other air quality generation methods with uncertainty estimation capability under the scenarios of original data, random missing, node missing, timestamp missing and spatio-temporal block missing. In addition, Figure 4 The generation results of a virtual station of the embodiment method under the original data scenario and the confidence interval thereof are shown.

[0184] With reference to Figure 5 The embodiment of the present application provides a virtual air station quality control data generation device based on a spatio-temporal Bayesian neural field, which comprises:

[0185] A data preprocessing unit is configured to input original data of a target pollutant concentration and corresponding original spatio-temporal features into a data preprocessing module, and respectively perform sample filtering, time discretization and feature normalization to obtain a normalized original feature matrix.

[0186] A feature encoding unit is configured to input the original feature matrix into a multi-layer spatio-temporal feature encoding module, and respectively perform spatio-temporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation to obtain spatio-temporal interaction feature, space-space interaction feature, time seasonality feature, spatial Fourier feature and spatial aggregation feature, and then splice each feature obtained and the original feature matrix to obtain a multi-layer spatio-temporal feature matrix.

[0187] A weight distribution unit is configured to input the multi-layer spatio-temporal feature matrix into an adaptive scaling and channel attention process in the multi-layer spatio-temporal feature encoding module for processing to form a high-dimensional feature matrix with weight distribution.

[0188] a mean prediction unit configured to bring the high-dimensional feature matrix into a feature enhancement and mean prediction module, respectively perform feature enhancement through three layers of channel gate learning units, and then access a fully connected layer network to generate a conditional mean set of the target pollutant concentration prediction;

[0189] a distribution restoration unit configured to bring the conditional mean set into a Bayesian probability prediction module, restore the target pollutant spatiotemporal distribution by using the original spatiotemporal feature, the original data of the target pollutant concentration, and the conditional mean set together, and generate a target pollutant concentration set and a corresponding variance set according to the target pollutant spatiotemporal distribution.

[0190] It can be understood that the contents in the above method embodiments are all applicable to the device embodiments, the device embodiments specifically implement the functions of the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0191] In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example, with the purpose of providing a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of a larger operation are independently executed.

[0192] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It can also be understood that a detailed discussion of the actual implementation of each module is unnecessary for an understanding of the present application. Rather, the properties, functions, and internal relationships of the various functional modules disclosed in the devices herein are considered within the ordinary skill of an engineer, and the actual implementation of the module will be appreciated within the context of the present application. Therefore, those skilled in the art can implement the present application as set forth in the claims without undue experimentation using ordinary skill. It can also be understood that the disclosed specific concepts are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0193] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts of the prior art that make contributions or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0194] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other system that can take instructions from an instruction execution system, apparatus, or device, or in conjunction with these instructions execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0195] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpreting, or otherwise processing, if necessary, in other suitable ways, to be electronically obtained, and then stored in the computer memory.

[0196] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0197] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0198] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0199] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for generating quality control data of a virtual air station based on a spatiotemporal Bayesian neural field, characterized in that, The method comprises the following steps: The original data of the target pollutant concentration and the corresponding original spatio-temporal features are brought into a data preprocessing module, and are subjected to sample filtering, time discretization and feature normalization respectively to obtain a normalized original feature matrix; The original feature matrix is brought into a multi-layer spatio-temporal feature coding module, and is subjected to spatio-temporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation to obtain spatio-temporal interaction feature, space-space interaction feature, time seasonality feature, spatial Fourier feature and spatial aggregation feature, and then the obtained features are spliced with the original feature matrix to obtain a multi-layer spatio-temporal feature matrix; The multi-layer spatio-temporal feature matrix is brought into the adaptive scaling and channel attention process in the multi-layer spatio-temporal feature coding module for processing to form a high-dimensional feature matrix with weight redistribution; The high-dimensional feature matrix is brought into a feature enhancement and mean prediction module, and is subjected to feature enhancement by three layers of channel gate learning units, and then is connected to a fully connected layer network to generate a conditional mean set for the target pollutant concentration prediction; The conditional mean set is brought into a Bayesian probability prediction module, and the original spatio-temporal features, the original data of the target pollutant concentration and the conditional mean set are used together to restore the spatio-temporal distribution of the target pollutant, and a target pollutant concentration set and a corresponding variance set are generated according to the spatio-temporal distribution of the target pollutant.

2. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 1, characterized in that, The spatio-temporal interaction feature is obtained by the following steps: The original feature matrix is multiplied by a set of timestamps , a set of observation site longitudes , and a set of observation site latitudes respectively, to obtain the spatio-temporal interaction feature : ； wherein, denotes element-wise multiplication, denotes vector or matrix concatenation; The space-space interaction feature is obtained by the following steps: multiplying the set of observation site longitudes in the original feature matrix by the set of observation site latitudes , thereby obtaining an air-to-air interaction term : ； The time seasonality feature is obtained by the following steps: By the time-harmonic expansion procedure employing orthogonal sine and cosine pairs as fixed basis functions, the fixed basis functions are smoothly integrated into the gradient training to capture multiple cycles and then form time seasonality term features ; The spatial Fourier feature is obtained by the following steps: mapping the normalized observation site longitude and observation site latitude using Fourier features through the spatial harmonic expansion to form the spatial Fourier term features and observation site latitude mapping the normalized observation site longitude and observation site latitude using Fourier features through the spatial harmonic expansion to form the spatial Fourier term features ; The spatial aggregation feature is obtained by the following steps: The space correlation is utilized to learn a field-specific non-stationary spatial correlation from the original feature matrix by using a single-layer multi-head graph attention network to form the spatial aggregation term feature .

3. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 1, characterized in that, The multi-layer spatio-temporal feature matrix is brought into the adaptive scaling and channel attention process in the multi-layer spatio-temporal feature coding module for processing to form a high-dimensional feature matrix with weight redistribution, comprising the following steps: The multi-level spatiotemporal feature matrix is ​​processed through the adaptive scaling and channel attention process. Automatic adjustment of the amplitude of each feature, amplification of key information, and suppression of redundancy are performed to form the high-dimensional feature matrix with weight redistribution. Specifically, it includes: for the multi-level spatio-temporal feature matrix scaling each channel in each batch by the following expression to achieve automatic rescaling of feature scales: ； wherein, is a set of learnable scaling coefficients; The feature matrix obtained after scaling is subjected to global average pooling to extract channel-level statistics and eliminate spatial position interference by the following expression The feature matrix obtained after scaling is subjected to global average pooling to extract channel-level statistics and eliminate spatial position interference by the following expression ； wherein, is the current batch index; is the total number of batches; is the feature matrix after global average pooling; The channel attention gate vector is generated using a two-layer fully connected network to capture the inter-channel nonlinear dependency by the following expression : ； wherein, and are weight matrices of two fully connected networks in the channel attention layer, respectively, and are bias vectors of two fully connected networks in the channel attention layer, respectively. The channel attention gate vector is obtained by multiplying the scaled feature matrix with the scaled feature matrix The high-dimensional feature matrix with weight redistribution is obtained by multiplying the channel attention gate vector : 。 4. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 1, characterized in that, The high-dimensional feature matrix is sequentially subjected to a gate residual network, a learnable activation function and channel attention in each layer of the channel gate learning unit; The gate residual network refers to the robust regulation of information flow of the feature matrix by the gate residual network, and specifically comprises: First, set For the Input feature matrix of the layer channel-gated learning unit; Then, the first intermediate feature matrix X1 is calculated by a linear transformation from the first feature matrix X0 according to the following expression Intermediate feature matrix of layer gate unit : ； wherein, and respectively represent the linear transformation weight matrix of the layer network; and respectively represent the bias vector of the linear projection of the layer network; is the hierarchical index of the channel-gated learning unit; Next, the first layer gate vector is calculated according to the following expression Layer gate vector for dynamically fusing raw and enhanced features: ； wherein, and respectively represent a weight matrix and a bias vector for generating a gating vector. Finally, the gating mechanism is utilized in conjunction with the hybrid activation function to produce the final layer gated residual output vector y according to the following expression layer gated residual output vector : ； The learnable activation function refers to replacing the single activation function in the original gate residual network with a trainable Elu-Tanh convex combination hybrid activation function, i.e. Elu retains negative response and accelerates convergence, and Tanh provides smooth and bounded output, thereby improving the fitting of complex spatio-temporal patterns and the uncertainty quantification capability, and specifically comprises: The learnable activation function is defined by the following expression: ； wherein the first Learnable mixed activation function for a layer By the first Learnable activation mixing coefficients for a layer Adaptively adjusting the proportion of the two activations during training, for the input of the mixed activation function, Tanh for the activation function.

5. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 1, characterized in that, The connection to the fully connected layer network to generate the conditional mean set for the target pollutant concentration prediction comprises the following steps: The conditional mean set is generated by the following expression using a lightweight fully connected network: ； wherein, denotes the set of conditional means, denotes the set of conditional means, layer output feature after the layer learnable activation function processing, and denote the weight matrix and bias vector of the mean prediction network, respectively.

6. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 1, characterized in that, The conditional mean set is brought into the Bayesian probability prediction module, and the original spatio-temporal features, the original data of the target pollutant concentration and the conditional mean set are used together to restore the spatio-temporal distribution of the target pollutant, comprising the following steps: The concentration distribution model of the target pollutant is defined in a Gaussian likelihood distribution form, and a Logistic(0, 1) prior distribution is independently set for each parameter in the parameter space; The parameters in the Gaussian likelihood distribution form are optimized by a maximum posterior probability method starting from randomly initialized points multiple times to obtain multiple local optimal solutions of the parameters; A prediction distribution component is constructed by using each of the local optimal solutions of the parameters, and all the prediction distribution components are equally mixed to obtain a posterior prediction distribution as the spatiotemporal distribution of the target pollutant; The Gaussian likelihood distribution form is represented as: ； wherein, denotes the standard deviation of the observation noise, collecting all network parameters, denotes the set of conditional means, is the number of observed entries in the training loss, is the identity matrix; denotes the raw data of the target pollutant concentration, denotes the corresponding raw spatio-temporal features; The optimization target of the maximum posterior probability method is represented as: ； wherein, represents a likelihood function; is the posterior optimal solution for the parameters; represents the estimated standard deviation of the observation noise in the output distribution at the prediction stage; The posterior prediction distribution is represented as: ； wherein, is the maximum a posteriori optimization number; and are the predicted mean and observation noise variance of the th posterior mode, respectively; indicates that the likelihood model of the observation data is Gaussian normal distribution; indicates the input feature vector of the prediction point; indicates the predicted target output of the model on the input ; is the training data set.

7. The spatiotemporal Bayesian neural field-based virtual air station quality control data generation method according to claim 6, characterized in that, The method for generating a target pollutant concentration set and a corresponding variance set according to the spatiotemporal distribution of the target pollutant comprises the following steps: The target pollutant concentration set and the variance set are generated according to the spatiotemporal distribution of the target pollutant as follows: ； wherein, denotes the set of target pollutant concentrations, denotes the set of variances.

8. A virtual air station quality control data generation device based on a spatiotemporal Bayesian neural field, characterized by, The device comprises: A data preprocessing unit is configured to input original data of target pollutant concentration and corresponding original spatiotemporal features into a data preprocessing module, and obtain a normalized original feature matrix after sample filtering, time discretization and feature normalization; A feature encoding unit is configured to input the original feature matrix into a multi-layer spatiotemporal feature encoding module, and obtain spatiotemporal interaction features, space-space interaction features, time seasonality features, spatial Fourier features and spatial aggregation features after spatiotemporal interaction, time harmonic expansion, spatial harmonic expansion and spatial correlation, and then splice the obtained features with the original feature matrix to obtain a multi-level spatiotemporal feature matrix; A weight distribution unit is configured to input the multi-level spatiotemporal feature matrix into an adaptive scaling and channel attention process in the multi-layer spatiotemporal feature encoding module for processing to form a high-dimensional feature matrix with weight distribution; A mean prediction unit is configured to input the high-dimensional feature matrix into a feature enhancement and mean prediction module, and perform feature enhancement through three layers of channel gate learning units, and then input into a fully connected layer network to generate a conditional mean set of target pollutant concentration prediction; A distribution restoration unit is configured to input the conditional mean set into a Bayesian probability prediction module, and restore the spatiotemporal distribution of the target pollutant by using the original spatiotemporal features, the original data of the target pollutant concentration and the conditional mean set, and generate a target pollutant concentration set and a corresponding variance set according to the spatiotemporal distribution of the target pollutant.

9. An electronic device, comprising: The electronic device comprises a processor and a memory; The memory is configured to store a program; The processor executes the program to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program, and the program is executed by a processor to implement the method of any one of claims 1 to 7. The storage medium stores a program, and the program is executed by a processor to implement the method of any one of claims 1 to 7.

Images