A multi-source heterogeneous-based atmospheric pollutant early warning method

Abstract

This invention provides a method for early warning of air pollutants based on multi-source heterogeneous data, belonging to the field of data processing technology. Specifically, it includes: Step 1, encrypted access and transmission of multi-source heterogeneous data; Step 2, data quality control and missing data processing; Step 3, spatiotemporal alignment and scale transformation; Step 4, feature engineering and variable construction; Step 5, fusion modeling and pollutant prediction; Step 6, dynamic threshold and risk scoring; Step 7, online learning and drift detection; Step 8, uncertainty assessment and interpretable output; Step 9, early warning issuance and closed-loop iteration. This invention improves response capability and early warning accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an early warning method for air pollutants based on multi-source heterogeneity. Background Technology

[0002] Currently, the development of technologies such as sensor networks, remote sensing, and mobile sensing has led to environmental monitoring data exhibiting characteristics of high dimensionality, multi-source nature, and streaming updates. This provides richer information support for pollution early warning, but it also brings new technical challenges: the spatiotemporal alignment of multi-source data is becoming more difficult, and data collected by different devices may have timestamp discrepancies or spatial resolution differences; data quality is inconsistent, with sensors in some remote areas susceptible to environmental interference and producing outliers, and missing data also affecting the continuity of analysis; and the demand for real-time processing is increasing, with high-frequency updated data requiring systems to have minute-level response capabilities, which traditional batch processing modes cannot meet.

[0003] It is evident that there is an urgent need for an air pollutant early warning method based on multi-source heterogeneity, which has high responsiveness and accuracy. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide an air pollutant early warning method based on multi-source heterogeneity, which at least partially solves the problems of poor response capability and early warning accuracy in the prior art.

[0005] This invention provides an early warning method for air pollutants based on multi-source heterogeneity, comprising:

[0006] Step 1, Encrypted Access and Transmission of Multi-Source Heterogeneous Data: Access multi-source data from ground air quality monitoring stations, remote sensing satellites, regional meteorological stations, and pollution source emission inventories. Use encryption algorithms to encrypt the transmitted data throughout the process and transmit it based on an optimized transmission protocol. At the same time, automatically select the storage medium according to the importance and frequency of use of the data.

[0007] Step 2, Data Quality Control and Missing Detection Handling: Outlier removal and missing detection value completion are performed on the multi-source data after encryption transmission. Outlier removal adopts statistical principles, and missing detection value completion adopts the sliding window Kalman filter algorithm.

[0008] Step 3, Spatiotemporal alignment and scaling transformation: The quality-controlled data is unified to a grid carrier with a preset spatial resolution and a time carrier with a preset time step. The spatial scaling transformation adopts an interpolation algorithm.

[0009] Step 4, Feature Engineering and Variable Construction: Construct multi-dimensional features based on the spatiotemporally aligned data, and fuse them into a feature matrix. Use the variance thresholding method to remove redundant features.

[0010] Step 5, Fusion Modeling and Pollutant Prediction: Construct a spatiotemporal graph convolutional network as a fusion model. Use the feature matrix as input to construct a spatial graph. Through spatial convolution and temporal convolution, output the hourly pollutant concentration prediction values ​​for the future preset time period.

[0011] Step 6, Dynamic Threshold and Risk Score: Based on the national air quality standards, a baseline threshold is determined, and the warning threshold is dynamically adjusted using the entropy weight method. The risk score is calculated based on the deviation between the predicted concentration and the dynamic threshold, and then mapped to the corresponding warning level.

[0012] Step 7, Online Learning and Drift Detection: The data distribution drift is detected using the distribution divergence detection method. When drift is detected, online learning is triggered to fine-tune the parameters of the fusion model.

[0013] Step 8, Uncertainty Assessment and Interpretability Output: The prediction uncertainty is assessed by sampling methods, and the core driving factors of pollution are identified by characteristic attribution analysis. The uncertainty level and the heatmap of the contribution of driving factors are output.

[0014] Step 9, Early Warning Issuance and Closed-Loop Iteration: Early warnings are issued in a tiered manner based on risk scores and interpretability output results. Actual monitoring data and feedback data are collected after the early warning, and the parameters of the data layer, early warning layer and model layer are optimized at different cycles to form an early warning closed loop.

[0015] According to a specific implementation of an embodiment of the present invention, the step of encrypting the transmitted data throughout the entire process using an encryption algorithm includes:

[0016] The key is dynamically updated during the encrypted data transmission process. The expression for the key dynamic update period is as follows:

[0017] ;

[0018] in, The basic update cycle is denoted by k, the attenuation coefficient is denoted by D, and the total amount of data transmitted is denoted by D.

[0019] According to a specific implementation of an embodiment of the present invention, the step of transmitting based on an optimized transmission protocol includes:

[0020] The bandwidth allocation for each data source is calculated and optimized accordingly. The expression for the bandwidth allocation for each data source is as follows:

[0021] ;

[0022] in, Total available bandwidth; This represents the priority coefficient of the current data source: 1.0 for high-importance, high-frequency data, 0.6 for medium-importance, medium-frequency data, and 0.3 for low-importance, low-frequency data; n is the total number of data sources. t represents the priority coefficient of the i-th data source; t represents the transmission time.

[0023] According to a specific implementation of an embodiment of the present invention, the step of automatically selecting a storage medium based on data importance and usage frequency includes:

[0024] Data storage priority scores are calculated based on data importance and usage frequency, and storage media are selected accordingly. The expression for the data storage priority score is as follows:

[0025] ;

[0026] in, Importance weight; Data importance scores are assigned as follows: 10 for high importance, 6 for medium importance, and 3 for low importance. Frequency weights; Scoring is done based on data usage frequency, with high frequency scored as 10, mid-frequency as 6, and low frequency as 3; according to Choose the storage medium. ≥8 select SSD, 5≤ <8. Select SAS, <5. Select HDD.

[0027] According to a specific implementation of the present invention, the spatial convolution weights of the spatiotemporal graph convolutional network are optimized using a physical consistency regularization term, the formula for which the physical consistency regularization term is calculated is:

[0028] ;

[0029] in, λ is the physical consistency regularization term; λ is the regularization coefficient. Total number of grid cells; The spatial convolution weights for the i-th grid; Let be the diffusion coefficient of the i-th grid.

[0030] According to a specific implementation of the present invention, the detection threshold used in the distribution divergence detection method employs an adaptive adjustment strategy, and the adjustment formula is as follows:

[0031] ;

[0032] in, The adjusted KL divergence threshold; Basic threshold; For adjustment coefficients; This represents the number of drift events that have occurred in the past 7 days.

[0033] According to a specific implementation of the present invention, the weight calculation of the feedback data in the closed-loop iteration adopts a time decay strategy, and the weight calculation formula is as follows:

[0034] ;

[0035] Where η is the attenuation coefficient; The current time; The time when the feedback data was generated.

[0036] The multi-source heterogeneous air pollutant early warning scheme in this embodiment of the invention includes: Step 1, encrypted access and transmission of multi-source heterogeneous data: accessing multi-source data from ground air quality monitoring stations, remote sensing satellites, regional meteorological stations, and pollution source emission inventories; encrypting the transmitted data throughout the process using an encryption algorithm; transmitting the data based on an optimized transmission protocol; and automatically selecting the storage medium according to the importance and frequency of use of the data; Step 2, data quality control and missing data processing: performing outlier removal and missing data completion on the encrypted multi-source data; outlier removal using statistical principles and missing data completion using a sliding window Kalman filter algorithm; Step 3, spatiotemporal alignment and scale transformation: unifying the quality-controlled data to a grid carrier with a preset spatial resolution and a time carrier with a preset time step; using an interpolation algorithm for spatial scale transformation; Step 4, feature engineering and variable construction: constructing multi-dimensional features based on the spatiotemporally aligned data and fusing them into a feature matrix; and removing redundant features using a variance thresholding method; Step 5, fusion modeling and pollutant prediction: constructing... The spatial graph convolutional network, used as a fusion model, constructs a spatial graph with the feature matrix as input. Through spatial and temporal convolution, it outputs hourly pollutant concentration predictions for a preset future time period. Step 6: Dynamic threshold and risk score: Based on national air quality standards, a benchmark threshold is determined, and the warning threshold is dynamically adjusted using the entropy weight method. The risk score is calculated based on the deviation between the predicted concentration and the dynamic threshold to map to the corresponding warning level. Step 7: Online learning and drift detection: The data distribution drift is detected using the distribution divergence detection method. When drift is detected, online learning is triggered to fine-tune the parameters of the fusion model. Step 8: Uncertainty assessment and interpretability output: The prediction uncertainty is assessed using a sampling method, and the core driving factors of pollution are identified using feature attribution analysis. The uncertainty level and the contribution heatmap of the driving factors are output. Step 9: Warning issuance and closed-loop iteration: Warnings are issued in stages based on the risk score and interpretability output results. Actual monitoring data and feedback data after the warning are collected, and the parameters of the data layer, warning layer, and model layer are optimized at different cycles to form a warning closed loop.

[0037] The beneficial effects of the embodiments of the present invention are as follows:

[0038] At the data processing level, this invention overcomes the technical bottleneck of multi-source heterogeneous data fusion. Through a unified standardization mechanism and spatiotemporal alignment strategy, it integrates data from different sources and formats, such as ground monitoring, remote sensing, meteorology, and emissions, into a consistent analytical foundation, effectively solving the problems of single data source and chaotic format in traditional methods. Simultaneously, through quality control and missing data completion techniques, it significantly improves data integrity and accuracy, providing high-quality data support for subsequent modeling and prediction, and avoiding early warning biases caused by data defects.

[0039] At the modeling and prediction level, this invention employs a spatiotemporal graph convolutional network to construct a fusion model that considers both the temporal evolution patterns and spatial diffusion characteristics of pollutants. Compared to traditional static models or single-dimensional modeling methods, it can more accurately depict pollution diffusion trends, especially significantly improving the predictive ability for regional and mobile pollution events. Furthermore, online learning and drift detection mechanisms enable the model to continuously adapt to changes in pollution patterns. When new pollution sources appear or seasons change, parameters can be automatically updated, avoiding the decline in prediction accuracy caused by the "static fixation" of traditional models and maintaining stable early warning performance over the long term.

[0040] At the early warning and judgment level, the adaptive threshold adjustment strategy breaks through the limitations of traditional static thresholds. It can dynamically optimize early warning standards based on real-time meteorological conditions, historical pollution patterns, and regional emission intensity, making early warning triggers more closely aligned with the actual pollution situation. This avoids response delays caused by excessively high thresholds and reduces false alarms caused by excessively low thresholds, thereby improving the flexibility and practicality of early warnings. The risk scoring mechanism transforms abstract pollution concentrations into intuitive risk levels, enabling managers to quickly grasp the severity of pollution and formulate targeted prevention and control measures.

[0041] At the level of result reliability and system operation, this invention addresses the difficulty in understanding traditional "black box models" through uncertainty assessment and interpretability output modules. Managers can not only obtain early warning results but also identify the core driving factors and sources of pollution, improving the scientific rigor and regulatory oversight of decisions. The edge-cloud collaborative streaming computing architecture enables minute-level processing of high-frequency data, with the edge ensuring real-time local response and the cloud handling global optimization. This not only meets real-time early warning needs but also supports cross-regional, large-scale deployment, significantly improving system stability and scalability, and providing reliable technical support for ecological environment supervision and urban air quality management. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating an air pollutant early warning method based on multi-source heterogeneity provided in an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the system structure corresponding to an air pollutant early warning method based on multi-source heterogeneity provided in an embodiment of the present invention;

[0045] Figure 3 A bar chart comparing the accuracy of early warnings for different pollutants provided in this embodiment of the invention;

[0046] Figure 4 A line graph showing the variation of model prediction accuracy over time, provided in an embodiment of the present invention;

[0047] Figure 5 A multi-dimensional performance evaluation radar chart provided for embodiments of the present invention. Detailed Implementation

[0048] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0051] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0052] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0053] This invention provides an early warning method for air pollutants based on multi-source heterogeneity, which can be applied to the prevention and control of air pollution in environmental management scenarios.

[0054] See Figure 1 This is a flowchart illustrating an air pollutant early warning method based on multi-source heterogeneity, provided by an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps:

[0055] Step 1, Encrypted Access and Transmission of Multi-Source Heterogeneous Data: Access multi-source data from ground air quality monitoring stations, remote sensing satellites, regional meteorological stations, and pollution source emission inventories. Use encryption algorithms to encrypt the transmitted data throughout the process and transmit it based on an optimized transmission protocol. At the same time, automatically select the storage medium according to the importance and frequency of use of the data.

[0056] In specific implementation, the method of the present invention can also be applied to Figure 2The system shown features encrypted access and transmission of multi-source heterogeneous data. It accesses data from ground-based air quality monitoring stations, remote sensing satellites, regional meteorological stations, and pollution source emission inventories. The system employs the AES-256-GCM encryption algorithm to encrypt the transmitted data throughout the process, generating a 128-bit authentication tag for data integrity verification. The transmission protocol uses an optimized MQTT-SN protocol with an adaptive heartbeat interval that dynamically adjusts from 1 to 5 seconds based on network latency. Simultaneously, the system automatically selects the storage medium based on data importance and usage frequency. Pollution source emission inventories and ground monitoring data are considered high-importance data, remote sensing satellite data are considered medium-importance data, meteorological data is considered high-frequency data (15 minutes / time) and emission inventories are considered low-frequency data (1 day / time). High-importance high-frequency data is stored on SSDs, medium-importance medium-frequency data is stored on SAS, and low-importance low-frequency data is stored on HDDs.

[0057] In the key dynamic update step of data encryption transmission, the formula is used. Calculate the key update cycle, where This refers to the key update cycle, expressed in hours. The base update cycle is 24 hours; k is the attenuation coefficient with a value of 0.05; D is the total amount of data transmitted in GB. This cycle calculation dynamically adjusts the key update frequency based on the amount of data transmitted. The larger the amount of data transmitted, the shorter the update cycle, reducing the security risks caused by using the same key for a long time, while balancing key updates and system overhead.

[0058] In the data transmission efficiency optimization step, the formula is used... Calculate the bandwidth allocation for each data source, where Bandwidth allocation for a single data source, in Mbps; Total available bandwidth, in Mbps; This represents the priority coefficient of the current data source: 1.0 for high-importance, high-frequency data, 0.6 for medium-importance, medium-frequency data, and 0.3 for low-importance, low-frequency data; n is the total number of data sources. t represents the priority coefficient of the i-th data source; t represents the transmission time in seconds. This integral calculation dynamically allocates bandwidth based on the priority of each data source, prioritizing high-importance, high-frequency data to obtain sufficient bandwidth, reducing data transmission latency, and improving overall transmission efficiency.

[0059] In the intelligent storage strategy optimization step, through the formula Calculate the data storage priority score, where This is a storage priority score, ranging from 0 to 10; This is the importance weight, with a value of 0.6; Data importance scores are assigned as follows: 10 for high importance, 6 for medium importance, and 3 for low importance. This is the frequency weight, with a value of 0.4; Scoring is done based on data usage frequency, with high frequency scored as 10, mid-frequency as 6, and low frequency as 3; according to Choose the storage medium. ≥8 select SSD, 5≤ <8. Select SAS, A score of <5 indicates the selection of HDD. This score calculation enables automatic matching of data storage media, balancing storage performance and cost, and improving data access efficiency.

[0060] Step 2, Data Quality Control and Missing Detection Handling: Outlier removal and missing detection value completion are performed on the multi-source data after encryption transmission. Outlier removal adopts statistical principles, and missing detection value completion adopts the sliding window Kalman filter algorithm.

[0061] In practice, data quality control and missing value handling involve outlier removal and missing value completion for multi-source data after encryption transmission. Outlier removal adopts the 3σ principle, calculating the mean μ and standard deviation σ of the data sequence. When the absolute value of the deviation between the data value and the mean exceeds 3σ, it is judged as an outlier and replaced with the linear interpolation result of the adjacent normal value. Missing value completion adopts sliding window Kalman filtering, initializing the initial state and covariance of the missing sequence, and calculating the completion value through prediction step and update step. After completion, the missing value rate of the data is controlled below 5%, and the outlier misclassification rate is controlled below 0.3%.

[0062] Step 3, Spatiotemporal alignment and scaling transformation: The quality-controlled data is unified to a grid carrier with a preset spatial resolution and a time carrier with a preset time step. The spatial scaling transformation adopts an interpolation algorithm.

[0063] In practice, spatiotemporal alignment and scale transformation unify the quality-controlled data into a 500m×500m grid spatial carrier and a 15-minute / hour time carrier. Time alignment uses the hour as the anchor point. Data with a deviation of less than 3 minutes is directly aligned, while data with a deviation of more than 5 minutes is retransmitted. Spatial scale transformation uses inverse distance weighted interpolation to convert point data into grid data and bilinear interpolation to convert a 1km coarse grid into a 500m fine grid, ensuring that the spatiotemporal resolution of different data sources remains consistent.

[0064] Step 4, Feature Engineering and Variable Construction: Construct multi-dimensional features based on the spatiotemporally aligned data, and fuse them into a feature matrix. Use the variance thresholding method to remove redundant features.

[0065] In practice, feature engineering and variable construction are carried out. Based on the spatiotemporally aligned data, multi-dimensional features of time, space, meteorology, and emissions are constructed. The time features include the concentration sequence of the previous 24 hours and the time decay factor. The spatial features include the concentration difference between adjacent grids and the mean emission intensity of the region. The meteorological features include the wind direction trigonometric function and the wind speed-diffusion coefficient. The emission features include the emission-time matching factor. The multi-dimensional features are fused into a feature matrix, and redundant features are removed by the variance thresholding method.

[0066] Step 5, Fusion Modeling and Pollutant Prediction: Construct a spatiotemporal graph convolutional network as a fusion model. Use the feature matrix as input to construct a spatial graph. Through spatial convolution and temporal convolution, output the hourly pollutant concentration prediction values ​​for the future preset time period.

[0067] In practice, modeling and pollutant prediction are integrated. A spatiotemporal graph convolutional network is constructed as the fusion model. The feature matrix is ​​used as input to construct a spatial graph. The nodes are grids and the edges are grids connected by a distance of less than 5km. The features of neighboring nodes are aggregated through spatial convolution and hourly concentration fluctuations are captured through temporal convolution. The predicted hourly pollutant concentration values ​​for the next 1-24 hours are output through a fully connected layer. The model training adopts the MAE+L2 regularized loss function and Adam optimizer. The 24-hour prediction accuracy of PM2.5 is maintained above 85%.

[0068] Furthermore, in the fusion modeling and pollutant prediction steps, the spatial convolution weights of the spatiotemporal graph convolutional network are optimized through a physical consistency regularization term, the formula of which is:

[0069] ;

[0070] in λ is the physical consistency regularization term; λ is the regularization coefficient, with a value of 0.01. Total number of grid cells; The spatial convolution weights for the i-th grid; The diffusion coefficient of the i-th grid is calculated based on wind speed. This regularization term makes the model parameters conform to the physical laws of pollutant diffusion, reduces the physical bias of the purely data-driven model, and improves the rationality and reliability of the prediction results.

[0071] Step 6, Dynamic Threshold and Risk Score: Based on the national air quality standards, a baseline threshold is determined, and the warning threshold is dynamically adjusted using the entropy weight method. The risk score is calculated based on the deviation between the predicted concentration and the dynamic threshold, and then mapped to the corresponding warning level.

[0072] In practice, dynamic thresholds and risk scores are determined based on national air quality standards. The weights of wind speed, historical pollution, and emission intensity are calculated using the entropy weight method. The warning thresholds for each grid and time period are dynamically adjusted. A risk score of 0-100 is calculated based on the deviation between the predicted concentration and the dynamic threshold. The risk score corresponds to five levels of warning: no warning, blue, yellow, orange, and red.

[0073] Step 7, Online Learning and Drift Detection: The data distribution drift is detected using the distribution divergence detection method. When drift is detected, online learning is triggered to fine-tune the parameters of the fusion model.

[0074] In practice, online learning and drift detection are carried out using the KL divergence method to detect data distribution drift. The KL divergence between the reference distribution and the current distribution is calculated. The reference distribution is the data distribution of the previous 30 days, and the current distribution is the data distribution of the most recent day. When the divergence value exceeds 0.1, it is determined to be drift and online learning is triggered. Online learning uses the incremental gradient descent method to fine-tune the model parameters, and the model adapts to changes in pollution patterns.

[0075] Furthermore, in the online learning and drift detection steps, the KL divergence threshold for drift detection adopts an adaptive adjustment strategy, and the threshold adjustment formula is:

[0076] ;

[0077] in The adjusted KL divergence threshold; The base threshold is set to 0.1. The adjustment factor is set to 0.02. The threshold is the number of drift events that have occurred in the past 7 days. This adjustment dynamically optimizes the threshold based on the frequency of historical drift events. When drift events occur frequently, the threshold is appropriately increased to reduce the frequent triggering of online learning. When drift events are rare, the threshold is decreased to improve the sensitivity of drift detection.

[0078] Step 8, Uncertainty Assessment and Interpretability Output: The prediction uncertainty is assessed by sampling methods, and the core driving factors of pollution are identified by characteristic attribution analysis. The uncertainty level and the heatmap of the contribution of driving factors are output.

[0079] In practice, uncertainty assessment and interpretability output are achieved by evaluating and predicting uncertainty through Monte Carlo sampling, sampling 1000 times for error and calculating 95% confidence intervals and relative uncertainty, using SHAP value analysis to identify core pollution driving factors, and outputting the uncertainty level and driving factor contribution heatmap for each grid.

[0080] Step 9, Early Warning Issuance and Closed-Loop Iteration: Early warnings are issued in a tiered manner based on risk scores and interpretability output results. Actual monitoring data and feedback data are collected after the early warning, and the parameters of the data layer, early warning layer and model layer are optimized at different cycles to form an early warning closed loop.

[0081] In practice, the early warning system is updated and closed-loop iterated. Early warnings are issued in a tiered manner based on risk scores and interpretability results. Local early warnings are issued to regional ecological and environmental departments and communities at the edge, while global early warnings are issued to cross-regional platforms and public service platforms in the cloud. Actual monitoring data and feedback data are collected after the early warning. Data layer parameters are optimized daily, early warning layer parameters are optimized weekly, and model layer parameters are optimized monthly to form an early warning closed loop.

[0082] Furthermore, in the early warning issuance and closed-loop iteration steps, the weight calculation of the feedback data adopts a time decay strategy, and the weight formula is as follows:

[0083] ;

[0084] in The weight for the feedback data is 0-1; η is the attenuation coefficient, with a value of 0.1. The current time is in days. The time of data generation is given in days; this weighting calculation gives higher weight to recent feedback data, making iterative optimization more in line with the current pollution situation and improving the timeliness and accuracy of closed-loop optimization.

[0085] The multi-source heterogeneous air pollutant early warning method provided in this embodiment overcomes the technical bottleneck of multi-source heterogeneous data fusion at the data processing level. Through a unified standardization mechanism and spatiotemporal alignment strategy, it integrates data from different sources and formats, such as ground monitoring, remote sensing, meteorology, and emissions, into a consistent analytical basis, effectively solving the problems of single data source and chaotic formats in traditional methods. Simultaneously, through quality control and missing data completion techniques, it significantly improves data integrity and accuracy, providing high-quality data support for subsequent modeling and prediction, and avoiding early warning deviations caused by data defects.

[0086] At the modeling and prediction level, this invention employs a spatiotemporal graph convolutional network to construct a fusion model that considers both the temporal evolution patterns and spatial diffusion characteristics of pollutants. Compared to traditional static models or single-dimensional modeling methods, it can more accurately depict pollution diffusion trends, especially significantly improving the predictive ability for regional and mobile pollution events. Furthermore, online learning and drift detection mechanisms enable the model to continuously adapt to changes in pollution patterns. When new pollution sources appear or seasons change, parameters can be automatically updated, avoiding the decline in prediction accuracy caused by the "static fixation" of traditional models and maintaining stable early warning performance over the long term.

[0087] At the early warning and judgment level, the adaptive threshold adjustment strategy breaks through the limitations of traditional static thresholds. It can dynamically optimize early warning standards based on real-time meteorological conditions, historical pollution patterns, and regional emission intensity, making early warning triggers more closely aligned with the actual pollution situation. This avoids response delays caused by excessively high thresholds and reduces false alarms caused by excessively low thresholds, thereby improving the flexibility and practicality of early warnings. The risk scoring mechanism transforms abstract pollution concentrations into intuitive risk levels, enabling managers to quickly grasp the severity of pollution and formulate targeted prevention and control measures.

[0088] At the level of result reliability and system operation, this invention addresses the difficulty in understanding traditional "black box models" through uncertainty assessment and interpretability output modules. Managers can not only obtain early warning results but also identify the core driving factors and sources of pollution, improving the scientific rigor and regulatory oversight of decisions. The edge-cloud collaborative streaming computing architecture enables minute-level processing of high-frequency data, with the edge ensuring real-time local response and the cloud handling global optimization. This not only meets real-time early warning needs but also supports cross-regional, large-scale deployment, significantly improving system stability and scalability, and providing reliable technical support for ecological environment supervision and urban air quality management.

[0089] The method of the present invention will be further described below with reference to two specific embodiments:

[0090] Example 1: Application of Air Pollutant Early Warning in a Provincial Capital City in Central China

[0091] I. Implementation Environment Configuration

[0092] This embodiment is applied to a provincial capital city in central China (built-up area 450 km²), deploying 12 ground-based air quality monitoring stations (including 6 national-level stations and 6 municipal-level stations, distributed in industrial areas, residential areas, and transportation hubs). It accesses satellite (1 km × 1 km grid resolution) remote sensing data, links with 8 regional meteorological stations (covering the urban area and surrounding counties, with an observation frequency of 15 minutes / time), and synchronizes with the pollution source emission inventory (including industrial, transportation, and agricultural pollution sources, updated daily) on the municipal ecological and environmental bureau's government affairs platform. In terms of hardware, edge computing nodes are used at the edge (deployed next to each monitoring station, CPU is ARM Cortex-A55, 8GB memory), and ECS servers (4 servers with 8 cores and 16GB memory, built on an Apache Flink 1.17.0 streaming platform) are used in the cloud. Data transmission uses AES-256-GCM encryption, and the key is dynamically updated every 8 hours (based on a formula). , =24 hours, k=0.05, daily average transfer D=100GB, (≈8.4 hours).

[0093] II. Detailed Implementation of Technical Solutions

[0094] 1. Data Access and Standardization

[0095] Multi-source data access: Ground stations transmit CSV format data (including PM2.5, O3, and NO2 concentrations, once per hour) in real time via 4G modules, which is then parsed into JSON format at the edge. MODIS data obtains TIFF files through the National Satellite Meteorological Center's push stream interface (once every 4 hours) and is synchronized to the cloud. Meteorological stations transmit binary streams (including wind direction, wind speed, temperature, and humidity, once every 15 minutes) via wired dedicated lines (telecom SDH dedicated lines, bandwidth 100Mbps), with priority given to edge access. Emission inventory is synchronized to Excel spreadsheets via the government platform API (HTTPS protocol) (updated daily at 00:00) and stored in the cloud HDD.

[0096] Data standardization: Pollutant concentrations were standardized to μg / m³, and O3 1-hour averages were converted to 8-hour averages. Wind direction and wind speed are orthogonally decomposed, such as when wind speed W = 3 m / s and wind direction θ = 90°. , Time is aligned to a 15-minute step, e.g., meteorological data at 08:05 is aligned to 08:15, and data at 08:20 is retransmitted due to a 10-minute deviation; time gaps are filled using linear interpolation, e.g., at ti=08:00, concentration xi=50μg / m³, at ti+1=08:15, concentration xi+1=60μg / m³, and at tk=08:07... .

[0097] 2. Quality control and handling of missing measurements

[0098] Outlier removal (3σ principle): Taking the PM2.5 time series (T=24 hours, Z_t is the hourly concentration) of a certain national monitoring station as an example, the mean is calculated. Standard deviation If at a certain moment Zt = 105 μg / m³, If it is determined to be an outlier, use the adjacent time interval ( Linear interpolation replacement: .

[0099] Missing value completion (Kalman filtering): Initializing the initial state of the missing value sequence (Last normal value before missing data), covariance p0=0.1; prediction step A=1, B=0. , , The update step takes H=1 and R=0.05. If the adjacent grid Zk=67μg / m³, calculate the Kalman gain. , Pk = (1 - 0.6875 × 1) × 0.11 = 0.034375, and the completed missing value is 66.4 μg / m³. The data after processing has a missing value rate of 3.2% and an outlier misclassification rate of 0.2%.

[0100] 3. Spatiotemporal alignment and scale transformation

[0101] Time alignment: Using the hour as the anchor point, the edge verifies the timestamps of each data source in real time. For example, if the data from the ground station at 09:15 deviates from the data from the meteorological station at 09:15 by 2 minutes, it is directly aligned; if it deviates from the data from the remote sensing data at 09:15 by 6 minutes, the remote sensing data is retransmitted to the cloud.

[0102] Spatial scale transformation: The warning area is divided into 1800 500m×500m grids (450km²÷0.25km² / grid); point data is converted to grids using inverse distance weighted interpolation, taking m=5 nearest monitoring stations. For example, at the grid center (x=113.7°E, y=28.2°N), the Z_i of the surrounding 5 monitoring stations are 58, 62, 65, 59, and 63 μg / m³, respectively, and the Euclidean distances di are 0.8, 1.2, 1.5, 1.0, and 1.3km, respectively. p=2, calculate... The 1km remote sensing coarse grid to fine grid conversion uses bilinear interpolation. The parent grid has four vertices: Z11=55, Z12=58, Z21=57, and Z22=60. The child grid is located in the lower right 1 / 4 region of the parent grid, with weights a=0.2, b=0.3, c=0.2, and d=0.3 (total 1). .

[0103] 4. Feature Engineering and Fusion Modeling

[0104] Feature engineering: Constructing temporal features (concentration sequence of 96 points in 15-minute intervals over the past 24 hours, time decay factor (weight of the last hour is 0.6), seasonal factor (2 in summer)), spatial features (concentration difference between 8 adjacent grids (maximum difference 5 μg / m³), mean regional emission intensity (80 t / day in industrial areas)), meteorological features (u=0, v=3, diffusion coefficient K=0.8×3^0.5≈1.39), emission features (emission-time matching factor (14:00 is the industrial peak, taken as 1)); feature fusion according to... (α=0.5, β=0.3, γ=0.2), Dimension 96 Dimension 4 Dimension 2, after fusion 102 dimensions, using variance thresholding to remove redundant features (retaining 64 dimensions).

[0105] Fusion Modeling (ST-GCN): The spatial graph G contains 1800 nodes, with edges forming a grid connection where d_ij < 5km (edge ​​weight A_ij = 1 / d_ij); the spatial convolutional layer has 32 3×3 convolutional kernels, and W_s has a dimension of 64×64. (N(i) takes 8 neighbors); the temporal convolutional layer uses 1D convolution (k=3), Wt dimension is 3×64×64; the fully connected layer output ( Dimensions 64×1, =0.1); loss function The Adam optimizer has a learning rate of 0.001 and is trained for 50 epochs.

[0106] 5. Dynamic thresholds and online learning

[0107] Dynamic threshold: PM2.5 baseline Th0 = 75 μg / m³, obtained using entropy weight method. ;

[0108] A certain grid =6 hours ;

[0109] ;

[0110] ( );

[0111] ;

[0112] Risk score , (Yellow alert).

[0113] Online learning: Given a reference distribution P containing data from the previous 30 days and a current distribution Q containing data from the most recent day, calculate the KL divergence. This triggers online learning; batch data B=1000 samples. The updated edge verification accuracy has been improved by 4.2%, and new parameters have been enabled.

[0114] 6. Uncertainty Assessment and Early Warning Issuance

[0115] Uncertainty assessment: Monte Carlo sampling M=1000 times, error ε~N(0.5,2²), sampled predicted value After sorting, Q2.5% = 65, Q97.5% = 73, CI95% = [65, 73]. (Reliable); SHAP value analysis yields core driving factors: regional emission intensity ( ), wind speed (0.72), concentration difference between adjacent grids (0.68), historical concentration (0.61), time period factor (0.55).

[0116] Warning issuance: The edge terminal pushes a yellow warning to the Municipal Ecology and Environment Bureau and residential area bulletin boards (content includes the affected area (industrial zone and 3 downwind grids), predicted concentration of 78 μg / m³, core driving factors (industrial emissions + low wind speed), and recommended measures (sensitive populations should reduce outdoor activities)); the cloud terminal synchronizes data with the Provincial Department of Ecology and Environment, optimizing the Kalman filter Q=0.015 daily (originally 0.01) and adjusting it weekly. =0.35 (originally 0.4), ST-GCN parameters are updated monthly.

[0117] III. Effect Verification

[0118] Table 1

[0119]

[0120] Table 1 shows that traditional methods suffer from limited data completeness (only 72%) due to a single data source, low prediction accuracy (68%) relying on static statistical models, 45-minute delay requiring manual verification, and low strategy adaptability due to the inability to identify driving factors. This invention improves data completeness to 96.8% through multi-source fusion (ground + remote sensing + meteorology + emissions), achieves prediction accuracy of 87.5% by capturing spatiotemporal features using ST-GCN, reduces latency to 8.2 minutes through real-time edge processing and cloud-based global optimization, achieves 89% accuracy in identifying driving factors through SHAP value analysis, and increases strategy adaptability to 92% through closed-loop iterative optimization. This fully demonstrates the technical advantages in single-city scenarios and meets the needs for precise early warning and decision support.

[0121] Example 2: Application of Air Pollutant Early Warning in a City Cluster in the Yangtze River Delta

[0122] I. Implementation Environment Configuration

[0123] This embodiment is applied to a city cluster in the Yangtze River Delta (including 4 prefecture-level cities, with a total area of ​​1200 km²), deploying 28 ground-based air quality monitoring stations (7 per city, including national / provincial control stations). It accesses remote sensing data from satellites (1 km × 1 km) and the Gaofen-6 satellite (500 m × 500 m, passing twice daily), and links with the regional meteorological center (12 meteorological stations, observing every 15 minutes). It also synchronizes cross-city emission inventories (including industrial clusters, intercity transportation, and regional agriculture, updated daily). In terms of hardware, the edge devices consist of 2 CloudLinkEdge100 servers per city (deployed at the municipal ecological and environmental bureau, with Intel Core i5 CPUs and 16GB of memory), and the cloud is the Yangtze River Delta Ecological and Environmental Collaborative Center server. Data transmission uses AES-256-GCM encryption, with the key configured according to... Update (when D=300GB per day, Tupdate≈3.6 hours). The software environment is the same as in Example 1, with the addition of consortium blockchain nodes (1 endorsement node per city, 1 ledger node in the cloud).

[0124] II. Detailed Implementation of Technical Solutions

[0125] 1. Data Access and Standardization

[0126] Multi-source data access: Ground stations in each city transmit JSON data via 5G modules (once per hour), which is parsed at the edge and synchronized to the cloud; MODIS and Gaofen-6 data are pushed via regional satellite receiving stations (once every 4 hours / 2 hours) and stored in the cloud SSD; meteorological data are transmitted as binary streams (once every 15 minutes) via a dedicated line (200Mbps bandwidth) of the Yangtze River Delta Meteorological Data Sharing Platform, with edge stations in each city receiving them first; cross-city emission inventories are synchronized to database tables via government collaboration API (updated at 1:00 AM daily) and stored in the cloud SAS.

[0127] Data standardization: pollutant concentrations are standardized to μg / m³, and O3 is converted to an 8-hour average; wind direction and speed are orthogonally decomposed, such as when W=4m / s and θ=180°. , Time alignment to 15-minute steps; cross-city data deviation exceeding 5 minutes triggers cloud-based coordinated retransmission; linear interpolation completion, such as... =10:00 a certain city =45μg / m³, =10:15 Neighboring City =50μg / m³, =10:09, .

[0128] 2. Quality control and spatiotemporal alignment

[0129] Outlier removal (3σ principle): Taking the NO2 sequence (T=24 hours) of a monitoring station along an intercity transportation route as an example, μ=52μg / m³, σ=10μg / m³, at a certain time... It was determined to be an outlier, and was used Interpolation replacement: .

[0130] Missing value completion (Kalman filter): Initialization , Prediction Step , Update steps are like those of neighboring cities. , After completion, the missing data rate for the region was 2.8%, and the outlier misclassification rate was 0.18%.

[0131] Spatiotemporal alignment: Spatial subdivision is divided into 4800 500m×500m grids (1200km²÷0.25km² / grid); point data is converted to grids by taking m=6 nearest monitoring stations (including cross-city monitoring stations), and Z(x,y) is calculated by inverse distance weighted interpolation; 1km remote sensing grid is converted to fine grids using bilinear interpolation, and Gaofen-6 500m data is directly associated with the grid; temporal alignment is checked hourly in the cloud to ensure that the data anchor points of the four cities are consistent.

[0132] 3. Feature Engineering and Fusion Modeling

[0133] Feature engineering: temporal features (concentration sequence of 96 points in the previous 24 hours, time decay factor (weight of the last hour is 0.6), seasonal factor (3 in autumn)), spatial features (concentration difference between 8 adjacent grids (maximum difference of 8 μg / m³), average cross-city emission intensity (95 t / day for industrial clusters)), meteorological features (u=-4, v=0, diffusion coefficient K=0.8×4^0.5≈1.6), emission features (emission-time matching factor (1 for 8:00 intercity traffic peak)); feature fusion (α=0.45, β=0.35, γ=0.2), the fusion dimension is 102, and the variance thresholding method retains 64 dimensions.

[0134] Fusion Modeling (ST-GCN): Spatial graph G contains 4800 nodes, with edges forming cross-city grid connections where d_ij < 8km (edge ​​weight A_ij = 1 / d_ij); Spatial convolutional layers have 32 3×3 convolutional kernels, W_s dimension is 64×64, and N(i) takes 10 neighbors; Temporal convolutional layers have k=3, W_t dimension is 3×64×64; Fully connected layers. The loss function is the same as in Example 1, and the training is performed for 60 epochs (1.5 years of historical data across cities).

[0135] 4. Dynamic thresholds and online learning

[0136] Dynamic threshold:

[0137] O3 baseline Th0 = 160 μg / m³, entropy weight method ;

[0138] A cross-city grid =12 hours ;

[0139] ;

[0140] ;

[0141] ;

[0142] Risk score (Orange alert).

[0143] Online learning:

[0144] The reference distribution P represents cross-city data from the previous 30 days, and the current distribution Q represents data from the most recent day. This triggered online learning; the cloud aggregated 1000 samples / city from the edge of 4 cities, totaling 4000 samples. After the update, the verification accuracy at the edge of each city has improved by 3.8%-4.5%, and new parameters have been enabled.

[0145] 5. Uncertainty Assessment and Early Warning Issuance

[0146] Uncertainty assessment:

[0147] Monte Carlo sampling M=1000 times, error ε~N(0.8,2.5²), CI95%=[148,162], (High reliability); Core driver of SHAP value: Intercity industrial emissions Intercity traffic flow (0.81), regional wind speed (0.75), remote sensing inversion (0.69), History Concentration (0.63).

[0148] Warning issued:

[0149] The edge device pushes orange alerts to the municipal ecological and environmental bureaus and intercity transportation hubs (including cross-city affected areas (120 grids in 3 cities), predicted concentration of 170 μg / m³, core driving factors (industry + transportation), and recommended measures (industrial production reduction of 30%, intercity truck restrictions)); the cloud device shares the alert information through the consortium blockchain (after desensitization, the information is uploaded to the blockchain and stored in hash value), and synchronizes it to the Yangtze River Delta Ecological and Environmental Collaborative Center. The model layer parameters are optimized monthly, and the cross-city emission inventory features are updated quarterly.

[0150] III. Effect Verification

[0151] Table 2

[0152]

[0153] Table 2 shows that traditional city-level early warning systems suffer from low accuracy (only 65-70% for PM2.5) due to data incompatibility and independent models. Cross-city responses require manual coordination and are delayed by 90-120 minutes. Source tracing relies on single-city data with an accuracy of only 35-40%, model updates are slow, and public reach is low. This invention improves accuracy to 85-89% (PM2.5) through cross-city multi-source fusion. The edge terminals of each city and the regional cloud shorten the collaborative response time to 15-22 minutes. Consortium blockchain-shared intelligence and SHAP value analysis improve source tracing accuracy to 88-92%. Daily global model updates and closed-loop iterative optimization improve reach to 90-93%, fully verifying the collaborative advantages in cross-regional scenarios and meeting the integrated pollution prevention and control needs of urban agglomerations.

[0154] refer to Figure 3 This bar chart visually demonstrates the superior early warning accuracy of this invention after multi-source data fusion. Traditional methods rely solely on single-source data from ground monitoring stations. Due to limited spatial coverage of monitoring points and a single data dimension, the early warning accuracy for various pollutants is generally below 75%, with ozone early warning accuracy at only 65% ​​because it does not incorporate meteorological diffusion conditions. This invention integrates ground monitoring, remote sensing satellite data, meteorological data, and emission inventories to achieve multi-dimensional information complementarity. The 24-hour early warning accuracy for PM2.5 is increased to 89%, and the accuracy for nitrogen dioxide and carbon monoxide exceeds 90%, fully validating the role of multi-source heterogeneous data fusion in improving early warning accuracy.

[0155] refer to Figure 4 The line graph illustrates the improvement in model accuracy and stability brought about by the online learning mechanism of this invention. Traditional static models, failing to consider changes in pollution patterns, show a gradual decline in accuracy from 82% after 10 days, dropping to 73% after 30 days due to data drift, and only 65% ​​after 50 days, failing to adapt to dynamic pollution situations. This invention detects data distribution drift using KL divergence. While accuracy slightly decreases after 30 days due to environmental changes, online learning is triggered after 40 days, updating parameters through incremental gradient descent, resulting in a recovery and stabilization above 85%. This dynamic optimization capability provides technical support for continuous and reliable pollution early warning.

[0156] refer to Figure 5This radar chart comprehensively demonstrates the overall performance advantages of this invention. Traditional methods are weak in data integrity (6 points) and real-time performance (5 points), especially in interpretability (only 3 points), failing to explain the basis for early warning, and exhibiting poor system scalability (4 points), making it difficult to adapt to cross-regional scenarios. This invention improves data integrity to 9 points through multi-source data fusion, achieves real-time performance of 9 points through edge-cloud collaborative streaming processing, ensures prediction accuracy of 9 points through online learning and spatiotemporal modeling, enhances interpretability to 8 points through SHAP value analysis and uncertainty assessment, and achieves scalability of 8 points through elastic architecture design. This multi-dimensional balanced improvement proves that this invention not only solves single technical pain points but also constructs an integrated early warning system that is "high-precision, high-real-time, high-reliability, and scalable," meeting the needs of multiple scenarios such as ecological environment supervision and urban air quality management.

[0157] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0158] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-source heterogeneous based atmospheric pollutant early warning method, characterized in that, include: Step 1, Encrypted Access and Transmission of Multi-Source Heterogeneous Data: Access multi-source data from ground air quality monitoring stations, remote sensing satellites, regional meteorological stations, and pollution source emission inventories. Use encryption algorithms to encrypt the transmitted data throughout the process and transmit it based on an optimized transmission protocol. At the same time, automatically select the storage medium according to the importance and frequency of use of the data. Step 2, Data Quality Control and Missing Detection Handling: Outlier removal and missing detection completion are performed on the encrypted multi-source data. Outlier removal adopts statistical principles, and missing detection completion adopts the sliding window Kalman filter algorithm. Step 3, Spatiotemporal alignment and scaling transformation: The quality-controlled data is unified to a grid carrier with a preset spatial resolution and a time carrier with a preset time step. The spatial scaling transformation adopts an interpolation algorithm. Step 4, Feature Engineering and Variable Construction: Construct multi-dimensional features based on the spatiotemporally aligned data and fuse them into a feature matrix. Redundant features are removed using the variance thresholding method. The multi-dimensional features include temporal features, spatial features, meteorological features, and emission features. Temporal features include the concentration sequence of the previous 24 hours and the time decay factor. Spatial features include the concentration difference between adjacent grids and the mean emission intensity of the region. Meteorological features include wind direction trigonometric functions and wind speed-diffusion coefficient. Emission features include emission-time matching factors. Step 5, Fusion Modeling and Pollutant Prediction: Construct a spatiotemporal graph convolutional network as a fusion model. Use the feature matrix as input to construct a spatial graph. Through spatial convolution and temporal convolution, output the hourly pollutant concentration prediction values ​​for the future preset time period. Step 6, Dynamic Threshold and Risk Score: Based on the national air quality standards, a baseline threshold is determined, and the warning threshold is dynamically adjusted using the entropy weight method. The risk score is calculated based on the deviation between the predicted concentration and the dynamic threshold, and then mapped to the corresponding warning level. Step 7, Online Learning and Drift Detection: The data distribution drift is detected using the distribution divergence detection method. When drift is detected, online learning is triggered to fine-tune the parameters of the fusion model. Step 8, Uncertainty Assessment and Interpretability Output: The prediction uncertainty is assessed by sampling methods, and the core driving factors of pollution are identified by characteristic attribution analysis. The uncertainty level and the contribution heatmap of the driving factors are output. Step 9, Early Warning Issuance and Closed-Loop Iteration: Early warnings are issued in a tiered manner based on risk scores and interpretability output results. Actual monitoring data and feedback data are collected after the early warning, and the parameters of the data layer, early warning layer and model layer are optimized at different cycles to form an early warning closed loop.

2. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 1, characterized in that, The step of encrypting the transmitted data throughout the entire process using an encryption algorithm includes: The key is dynamically updated during the encrypted data transmission process. The expression for the key dynamic update period is as follows: in, The basic update cycle is denoted by k, the attenuation coefficient is denoted by D, and the total amount of data transmitted is denoted by D.

3. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 2, characterized in that, The steps for transmitting based on the optimized transmission protocol include: The bandwidth allocation for each data source is calculated and optimized accordingly. The expression for the bandwidth allocation for each data source is as follows: in, Total available bandwidth; This represents the priority coefficient of the current data source: 1.0 for high-importance, high-frequency data, 0.6 for medium-importance, medium-frequency data, and 0.3 for low-importance, low-frequency data; n is the total number of data sources. t represents the priority coefficient of the i-th data source; t represents the transmission time.

4. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 3, characterized in that, The step of automatically selecting the storage medium based on data importance and usage frequency includes: Data storage priority scores are calculated based on data importance and usage frequency, and storage media are selected accordingly. The expression for the data storage priority score is as follows: in, Importance weight; Data importance scores are assigned as follows: 10 for high importance, 6 for medium importance, and 3 for low importance. Frequency weights; Scoring is done based on data usage frequency, with high frequency scored as 10, mid-frequency as 6, and low frequency as 3; according to Choose the storage medium. ≥8 select SSD, 5≤ <8. Select SAS, <5. Select HDD.

5. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 4, characterized in that, The spatial convolution weights of the spatiotemporal graph convolutional network are optimized using a physical consistency regularization term, the formula for which is: in, λ is the physical consistency regularization term; λ is the regularization coefficient. Total number of grid cells; The spatial convolution weights for the i-th grid; Let be the diffusion coefficient of the i-th grid.

6. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 5, characterized in that, The distribution divergence detection method employs an adaptive adjustment strategy for the detection threshold, and the adjustment formula is as follows: in, The adjusted KL divergence threshold; Basic threshold; For adjustment coefficients; This represents the number of drift events that have occurred in the past 7 days.

7. The atmospheric pollutant early warning method based on multi-source heterogeneity according to claim 6, characterized in that, The weight calculation of the feedback data in the closed-loop iteration adopts a time decay strategy, and the weight calculation formula is as follows: Where η is the attenuation coefficient; The current time; The time when the feedback data was generated.

Images