Atmospheric pollution source tracking and early warning system based on deep learning

Abstract

The invention relates to the technical field of atmospheric environment monitoring, in particular to an atmospheric pollution source tracking and early warning system based on deep learning, which comprises a multi-source data acquisition layer for acquiring multi-dimensional data of atmospheric pollutant concentration, weather, geography and the like through multi-element equipment such as satellite remote sensing and ground sensors; the data preprocessing layer completes data cleaning by using a statistical method and a machine learning algorithm, and realizes heterogeneous data standardization processing through feature extraction, multi-source integration and space-time alignment, and the deep learning model layer realizes heterogeneous data standardization processing by means of a space-time convolutional neural network, a graph neural network and a long-short term memory network in combination with a space-time attention mechanism. The traceability analysis layer is combined with a back diffusion algorithm and a geographic information system to position a pollution source and generate a thermodynamic diagram. And the early warning decision-making layer sets threshold early warning through transfer learning of the adaptive area, and generates an optimal pollution control scheme based on reinforcement learning.

Description

Technical Field

[0001] The present invention relates to the technical field of atmospheric environment monitoring, and specifically to an atmospheric pollution source tracking and early warning system based on deep learning. Background Art

[0002] Air pollution has become a global environmental challenge. Pollutants like PM2.5 and ozone cause haze, photochemical smog, and other phenomena that not only seriously endanger public health but also irreversibly damage ecosystems. In this context, accurately tracking pollution sources and providing timely warnings are crucial for environmental governance. However, traditional methods for monitoring and tracing air pollution sources have significant shortcomings.

[0003] In terms of monitoring coverage, traditional monitoring systems rely on fixed ground stations with low density and irrational layout, making it difficult to effectively cover mountainous areas, remote regions, and urban areas with complex terrain. For example, in parts of central and western my country, monitoring stations are spaced over 100 kilometers apart, resulting in significant loss of pollution data and large errors in locating pollution sources, making it impossible to meet the needs of refined pollution control.

[0004] Traditional data processing relies heavily on manual analysis. This makes manual statistics, screening, and analysis extremely inefficient when faced with massive amounts of monitoring data on pollutant concentrations, meteorological conditions, and other issues. For example, a provincial capital city generates over 10,000 pieces of air quality data daily. Manual processing can take hours, and subjective judgments can easily lead to missed or misjudged data, making it impossible to provide timely and effective data support for pollution prevention and control.

[0005] At the source tracing model level, traditional physical diffusion models (such as the Gaussian diffusion model) are based on the assumption of ideal atmospheric conditions and are difficult to adapt to the complex and changing real-world environment. In areas with large terrain and complex meteorological conditions, it is impossible to accurately simulate the diffusion path of pollutants under special meteorological conditions such as turbulence and inversion layers. As a result, the source tracing results deviate significantly from the actual situation and cannot provide a reliable basis for governance decision-making.

[0006] In terms of early warning response, traditional systems often rely on threshold-triggered warnings, which lack forward-looking predictions of pollution trends. Warnings are only issued when monitoring data reaches a preset threshold, often missing the optimal time for prevention and control, making it difficult to meet the needs of real-time dynamic governance.

[0007] While existing technologies attempt to improve monitoring systems by introducing machine learning algorithms, they still face numerous limitations. Most models are trained on a single data source and lack the ability to integrate heterogeneous data from multiple sources, such as satellite remote sensing, ground monitoring, and meteorological data, effectively failing to fully explore potential connections between them. Furthermore, these models suffer from weak generalization and poor adaptability to diverse regions and meteorological conditions, making it difficult to achieve high-precision pollution source tracking and pollution trend warnings. These limitations hinder the urgent need for timely, accurate, and intelligent air pollution control. Summary of the Invention

[0008] (1) Technical problems solved

[0009] In response to the shortcomings of the existing technology, the present invention provides an atmospheric pollution source tracking and early warning system based on deep learning.

[0010] (2) Technical solution

[0011] To achieve the above objectives, the present invention provides the following technical solutions: The deep learning-based atmospheric pollution source tracking and early warning system of the present invention comprises:

[0012] Multi-source data collection layer, used to collect atmospheric pollutant concentration data, meteorological data and geographic information;

[0013] The data preprocessing layer is used to clean the collected data, fuse the data, and align the time and space;

[0014] The deep learning model layer, including spatiotemporal convolutional neural networks, graph neural networks, and long short-term memory networks, is used to extract the spatiotemporal characteristics of pollutants and predict concentration trends;

[0015] The source tracing analysis layer locates the pollution source based on the output of the deep learning model and the reverse diffusion algorithm;

[0016] The early warning decision-making layer is used to set pollution thresholds, generate early warning information and provide prevention and control recommendations.

[0017] Preferably, the input data of the multi-source data collection layer includes meteorological data, pollution concentration, geographic information, real-time traffic and industrial emission data.

[0018] Further preferably, the multi-source data acquisition layer includes a satellite remote sensing image acquisition unit, a ground sensor network unit, drone monitoring, weather station data acquisition and an industrial emission database interface.

[0019] Again preferably, the data cleaning comprises the following steps:

[0020] Denoising: Identify and remove outliers using statistical methods or machine learning algorithms.

[0021] Missing value processing: For data points with missing values, interpolation, forward filling, backward filling or model-based methods are used to fill them;

[0022] Format standardization: Convert data from different sources into a unified standard format to facilitate subsequent processing.

[0023] Preferably, the data fusion comprises the following steps:

[0024] Feature selection and extraction: Select relevant feature variables based on the research objectives and extract these features from the original data;

[0025] Multi-source data integration: integrating information from multiple data sources including satellite remote sensing, ground sensor networks, drone monitoring, and weather stations;

[0026] Consistency verification: Verify the consistency between data from different sources through cross-validation and other methods to ensure data quality.

[0027] Further preferably, the spatiotemporal alignment comprises the following steps:

[0028] Time synchronization: The time axis of all data is aligned to the same reference time interval by using resampling technology;

[0029] Spatial matching: For observation points with different geographical locations, spatial interpolation and smoothing are performed using geographic information system tools so that all data can be represented on the same geographic grid;

[0030] Resolution adjustment: downscaling high-resolution data or upscaling low-resolution data.

[0031] Again preferably, the deep learning model layer adopts a spatiotemporal attention mechanism to adaptively assign spatiotemporal feature weights.

[0032] Preferably, the source tracing analysis layer is combined with a geographic information system to generate a heat map of the spatiotemporal distribution of pollution sources.

[0033] Further preferably, the early warning decision-making layer quickly adapts to the air pollution monitoring needs of different regions through transfer learning technology

[0034] Preferably again, the early warning decision layer generates an optimal pollution control plan based on a reinforcement learning algorithm.

[0035] (3) Beneficial effects

[0036] Compared with the existing technology, the present invention provides an air pollution source tracking and early warning system based on deep learning, which has the following beneficial effects:

[0037] Comprehensive and Accurate Data: The multi-source data collection layer integrates meteorological, pollution concentration, geographic, real-time traffic, and industrial emissions data, leveraging a variety of acquisition units to comprehensively capture information. The data preprocessing layer enhances data quality through noise reduction, missing value processing, and format standardization. Data fusion is achieved through feature selection, multi-source integration, and consistency verification. Time synchronization, spatial matching, and resolution adjustment are also implemented, ensuring data accuracy and adaptability to model analysis, laying a solid foundation for subsequent steps.

[0038] Accurate prediction and tracing: The deep learning model layer's spatiotemporal convolutional neural networks, graph neural networks, and long-short-term memory networks work together, combined with a spatiotemporal attention mechanism, to effectively extract the spatiotemporal characteristics of pollutants and accurately predict concentration trends. The source analysis layer, based on model results, combines a back-diffusion algorithm with a geographic information system to generate a heat map of the spatiotemporal distribution of pollution sources. This allows for rapid and accurate location of pollution sources, clarifying their scope and intensity.

[0039] Early Warning and Decision-Making Intelligence: The early warning decision-making layer sets pollution thresholds and generates timely warning information. Using transfer learning technology, it can quickly adapt to the monitoring needs of different regions, improving system versatility. The optimal pollution control plan generated by the reinforcement learning algorithm provides scientific decision-making for relevant departments, improving the relevance and effectiveness of pollution prevention and control, thereby reducing the harm of air pollution to the environment and human health, and has important environmental and social value. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the system workflow of the present invention;

[0041] Figure 2 This is a schematic diagram of the structure of the multi-source data acquisition layer of the present invention;

[0042] Figure 3 This is a schematic diagram of the data preprocessing layer structure of the present invention; DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] See also Figure 1-3 The air pollution source tracking and early warning system based on deep learning of the present invention includes:

[0045] Multi-source data collection layer, used to collect atmospheric pollutant concentration data, meteorological data and geographic information;

[0046] The data preprocessing layer is used to clean the collected data, fuse the data, and align the time and space;

[0047] The deep learning model layer, including spatiotemporal convolutional neural networks, graph neural networks, and long short-term memory networks, is used to extract the spatiotemporal characteristics of pollutants and predict concentration trends;

[0048] The source tracing analysis layer locates the pollution source based on the output of the deep learning model and the reverse diffusion algorithm;

[0049] The early warning decision-making layer is used to set pollution thresholds, generate early warning information and provide prevention and control recommendations.

[0050] Overall architecture synergy principle

[0051] The deep learning-based air pollution source tracking and early warning system achieves tracking and early warning of air pollution sources through the collaborative work of a multi-source data acquisition layer, a data preprocessing layer, a deep learning model layer, a source tracing analysis layer, and an early warning decision-making layer. The multi-source data acquisition layer collects various types of data, providing a foundation for subsequent analysis; the data preprocessing layer cleans, integrates, and aligns data in time and space to make it usable; the deep learning model layer mines data features and predicts trends; the source tracing analysis layer locates pollution sources based on model results; and the early warning decision-making layer sets thresholds and generates early warnings and prevention and control recommendations. These layers work closely together to form a complete monitoring, analysis, and decision-making system.

[0052] Working principle of each layer

[0053] The multi-source data collection layer utilizes satellite remote sensing imagery to acquire large-scale atmospheric pollution and geographic information. Ground-based sensor networks provide precise localized monitoring. Unmanned aerial vehicle monitoring flexibly supplements data on specific areas. Weather station data collection provides meteorological elements, and an industrial emissions database interface captures industrial emissions information. This multi-source data reflects atmospheric pollution from diverse perspectives, comprehensively collecting meteorological data, pollution concentrations, geographic information, real-time traffic data, and industrial emissions data, providing rich information for subsequent analysis.

[0054] The satellite remote sensing image acquisition unit, ground sensor network unit, drone monitoring, weather station data acquisition and industrial emission database interface can adopt the existing application technology structure:

[0055] Satellite remote sensing image acquisition unit: For example, when monitoring pollution in industrially dense areas of Europe, the TROPOMI sensor on the Sentinel-5P satellite can obtain atmospheric composition data at high resolution;

[0056] Ground sensor network units: For example, Beijing has built a vast ground-based air quality monitoring network with numerous monitoring stations distributed throughout the city. These stations are equipped with sensors that monitor the concentrations of pollutants such as PM2.5, PM10, and ozone (O3), as well as equipment that measures meteorological parameters such as temperature, humidity, wind speed, and wind direction.

[0057] Drone monitoring, for example, uses drones equipped with high-resolution cameras and gas sensors to fly close to factory areas. During monitoring, drones can capture factory activities and emissions, while also measuring pollutant concentrations in the atmosphere surrounding the factory.

[0058] Weather station data collection, for example, when monitoring photochemical smog pollution in the Los Angeles area, weather stations provide data such as temperature, humidity, sunshine duration, and wind speed and direction;

[0059] Industrial emissions database interface. For example, Volkswagen Group transmits exhaust emission data from its factories, including information on the amount, time, and method of emission of various pollutants, through a database interface. Environmental protection departments can access this data through this database interface.

[0060] Data preprocessing layer

[0061] Data cleaning: Statistical methods use reasonable thresholds, such as calculating the mean and standard deviation based on historical data, to identify and remove data that deviates from the normal range as outliers. Machine learning algorithms use models such as the isolation forest to automatically learn data distribution patterns and detect outliers. For missing values, interpolation methods use trends in adjacent data to fill in missing values. Forward filling and backward filling use the previous or next valid data point, respectively. Model-based methods use deep learning models to predict missing values. Format standardization unifies data formats, such as time format and data units, to facilitate subsequent processing.

[0062] The above-mentioned machine learning algorithm uses the isolation forest model and is often used in industrial production fault detection or network security intrusion detection. The principle is: the construction process of the isolation forest model is similar to that of a random forest, which consists of multiple isolated trees. When constructing each isolated tree, a sample subset is randomly selected from the original data set, and this subset is recursively divided. Each time a partition is performed, a feature and a split value on the feature are randomly selected to divide the data into two child nodes. For a data point, the shorter its path length in the isolation tree, the easier it is to be isolated, and the more likely it is to be an anomaly. For example, in a data set containing temperature data, the normal temperature range is 10℃-30℃. If there is a data point that is 50℃, then its path in the isolation tree will be very short, and the model will consider it an anomaly. By calculating the average path length of all data points on multiple isolated trees, the anomaly score of each data point is obtained. The closer the anomaly score is to 1, the more likely the data point is to be an anomaly; the closer the score is to 0, the more likely it is to be a normal point.

[0063] Data fusion: Feature selection and extraction, based on an understanding of air pollution and research objectives, selects features that are important for pollution source tracking and early warning, such as specific pollutant concentrations, wind direction and speed, and extracts these features from the raw data. Multi-source data integration combines data from diverse sources, such as satellite remote sensing and ground-based sensor networks, leveraging the strengths of each. Consistency verification, through methods such as cross-validation, compares data from different sources describing the same area or phenomenon, verifying data consistency and ensuring data quality.

[0064] Spatiotemporal alignment: Time synchronization uses resampling technology to unify data collected at different frequencies to the same baseline time interval, facilitating time series analysis. Spatial matching utilizes GIS tools to perform spatial interpolation and smoothing on data from observation points at different locations, allowing the data to be represented on a common geographic grid for easy spatial analysis. Resolution adjustment involves downscaling high-resolution data or upscaling low-resolution data based on actual needs, balancing data accuracy and computational complexity.

[0065] This resampling technology effectively resolves the issue of data time inconsistency by unifying data collected at different frequencies to the same baseline time interval. It has widespread and important applications in fields such as meteorological monitoring, intelligent transportation, and financial transactions. For example, in meteorological monitoring, satellite remote sensing data can be resampled to a 15-minute baseline time interval. If a satellite collects data on the hour, to meet the 15-minute interval requirement, the data for the remaining 15-minute intervals must be estimated using methods such as linear interpolation based on two adjacent satellite data points. For ground-based meteorological stations with high data collection frequencies, statistical methods such as mean, maximum, or minimum values ​​can be used for downsampling. For example, the average wind speed over every 15 minutes can be calculated as the wind speed data for that period. For weather balloon data, which is collected less frequently, if no balloon data is available within a particular 15-minute interval, a reasonable estimate can be made based on the preceding and following data. This unified time interval allows for better analysis of relationships between meteorological elements and improves the accuracy of weather forecasts.

[0066] Deep Learning Model Layer: The Spatiotemporal Convolutional Neural Network (STCNN) uses convolution kernels sliding across time and space to extract the spatiotemporal characteristics of atmospheric pollutants and capture local spatiotemporal patterns. The Graph Neural Network (GNN) treats different regions as nodes and the pollution transmission relationships between regions as edges, building a graph structure to learn inter-regional pollution correlations. The Long Short-Term Memory Network (LSTM) can process long-term dependencies in time series data and predict trends in pollutant concentrations. Using a spatiotemporal attention mechanism, the model adaptively assigns weights to spatiotemporal features based on data importance, highlighting key information.

[0067] In complex spatiotemporal data modeling scenarios, the collaborative architecture of spatiotemporal convolutional neural networks (STCNN), graph neural networks (GNN), and long short-term memory networks (LSTM) can fully leverage their respective strengths to achieve in-depth mining of spatiotemporal dependencies, spatial topological structures, and temporal dynamic features. The following further analyzes the technical principles, collaborative mechanisms, and typical application scenarios:

[0068] STCNN: Capturing Local Spatiotemporal Patterns

[0069] Core capabilities: It uses spatiotemporal convolution kernels to simultaneously extract features within spatial neighborhoods and time windows, and is applicable to regular grid structures (such as meteorological grids and urban partitions).

[0070] GNN: Modeling Unstructured Spatial Relationships

[0071] Core capabilities: Modeling entities (nodes) and relationships (edges) as graph structures, and learning dependencies between nodes (such as pollution spread between cities and traffic flow impacts) through message passing mechanisms.

[0072] LSTM: Handling long sequence dependencies

[0073] Core Capabilities: Effectively handle long-term dependencies in sequence data (such as the historical cumulative effects of pollutant concentrations) through gating mechanisms (input gate, forget gate, output gate).

[0074] Typical application scenarios

[0075] Air pollution source tracing and early warning

[0076] Data characteristics: The spatial distribution of monitoring sites is uneven (irregular grid), and the spread of pollutants is affected by meteorological conditions (time series dynamics).

[0077] Collaborative logic:

[0078] STCNN: Analyze local spatiotemporal variation patterns of sites (e.g., diurnal variation patterns of pollution in industrial areas);

[0079] GNN: Constructs pollution propagation maps based on meteorological fields (wind direction and speed) to identify potential source area associations;

[0080] LSTM: Predict pollutant concentration trends and trace the source using the back-diffusion algorithm.

[0081] Smart Grid Load Forecasting

[0082] Data characteristics: Grid nodes (substations, users) form a complex network, and load changes are affected by temporal and spatial factors (such as weather and user behavior).

[0083] Collaborative logic:

[0084] STCNN: Extracts spatiotemporal patterns of regional electricity load (e.g., differences between commercial and residential areas);

[0085] GNN: Modeling the topology of the power grid and analyzing the power transmission relationship between nodes;

[0086] LSTM: Predict future load demand and support scheduling decisions.

[0087] Epidemic spread prediction

[0088] Data characteristics: The distribution of cases is spatially clustered (communities, cities) and temporally diffuse.

[0089] Collaborative logic:

[0090] STCNN: Analyze the growth pattern of cases in the region (such as the rate of community transmission);

[0091] GNN: Constructing a propagation graph based on population mobility data (transportation networks, social relationships);

[0092] LSTM: Predict epidemic trends and evaluate the effectiveness of prevention and control measures.

[0093] Source Analysis Layer: Based on the pollutant concentration trends and spatiotemporal characteristics output by the deep learning model, combined with the reverse diffusion algorithm, this layer works backward from the current pollution distribution to identify the possible source locations of the pollution. Integrating with the Geographic Information System (GIS), this layer combines the location information of pollution sources with geographic data to generate a heat map of the spatiotemporal distribution of pollution sources, visually displaying their distribution across time and space.

[0094] The backdiffusion algorithm is a computational method based on physical models and mathematical principles. It solves practical problems by inversely simulating the diffusion phenomenon in physical processes. It has a wide range of applications in areas such as atmospheric pollution source tracking, groundwater pollution tracing, and medical image processing.

[0095] Early warning decision-making layer: Reasonable pollution thresholds are set. When the pollutant concentration predicted by the deep learning model exceeds the threshold, the early warning mechanism is triggered. Using transfer learning technology, models trained in one region can be quickly adapted to other regions, leveraging existing knowledge to reduce training costs in new areas and improve monitoring efficiency. Based on reinforcement learning algorithms, considering different pollution control measures and their impact on the environment, the optimal pollution control plan is generated through continuous trial and error and optimization.

[0096] The core principle of transfer learning is to leverage knowledge learned in one or more source tasks to accelerate the learning process of a target task. This is based on the fact that different tasks often have certain correlations. Knowledge learned in the source tasks, such as features, model structures, or training experience, can be appropriately transferred to the target task, thereby reducing the target task's reliance on large amounts of training data and improving learning efficiency and model performance.

[0097] Reinforcement learning, as described above, is a machine learning method in which an agent learns optimal behavior strategies based on reward signals through continuous trial and error in an environment. It plays an important role in complex scenarios such as optimizing decision-making and resource allocation.

[0098] Transfer learning and reinforcement learning are common application technologies in the fields of computer vision, natural language processing, and recommendation systems.

[0099] Detailed workflow

[0100] Data collection stage: Satellite remote sensing image acquisition unit, ground sensor network unit, drone monitoring, weather station data collection and industrial emission database interface work simultaneously to collect meteorological data, pollution concentration, geographic information, real-time traffic and industrial emission data, and transmit the collected data to the data storage center.

[0101] Data preprocessing: The collected data is cleaned, outliers are removed, missing values ​​are filled, and the data format is standardized. Next, data fusion is performed to select and extract relevant features, integrate multi-source data, and verify consistency. Finally, spatiotemporal alignment is performed to synchronize timelines, match spatial locations, and adjust data resolution to obtain usable preprocessed data.

[0102] Model Analysis Phase: Preprocessed data is fed into the deep learning model layer. STCNN, GNN, and LSTM work collaboratively to extract spatiotemporal features and predict pollutant concentration trends. A spatiotemporal attention mechanism is employed to enhance model performance. The Source Analysis Layer uses the model output and a back-diffusion algorithm to locate pollution sources. Using GIS, a heat map of the spatiotemporal distribution of pollution sources is generated.

[0103] Early Warning Decision-Making Stage: The early warning decision-making layer receives the predictions from the deep learning model and compares them with the set pollution threshold. If the threshold is exceeded, an early warning message is generated and disseminated through various channels. Simultaneously, transfer learning technology is used to quickly adapt to regional differences, and a reinforcement learning algorithm is used to generate an optimal pollution control plan, which is then provided to relevant departments for reference and implementation.

[0104] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The deep learning-based air pollution source tracking and early warning system is characterized by: include: Multi-source data collection layer, used to collect atmospheric pollutant concentration data, meteorological data and geographic information; The data preprocessing layer is used to clean the collected data, fuse the data, and align the time and space; The deep learning model layer, including spatiotemporal convolutional neural networks, graph neural networks, and long short-term memory networks, is used to extract the spatiotemporal characteristics of pollutants and predict concentration trends; The source tracing analysis layer locates the pollution source based on the output of the deep learning model and the reverse diffusion algorithm; The early warning decision-making layer is used to set pollution thresholds, generate early warning information and provide prevention and control recommendations.

2. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The input data of the multi-source data collection layer includes meteorological data, pollution concentration, geographic information, real-time traffic and industrial emission data.

3. The deep learning-based air pollution source tracking and early warning system according to claim 2 is characterized in that: The multi-source data acquisition layer includes satellite remote sensing image acquisition unit, ground sensor network unit, drone monitoring, weather station data acquisition and industrial emission database interface.

4. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The data cleaning comprises the following steps: Denoising: Identify and remove outliers using statistical methods or machine learning algorithms. Missing value processing: For data points with missing values, interpolation, forward filling, backward filling or model-based methods are used to fill them; Format standardization: Convert data from different sources into a unified standard format to facilitate subsequent processing.

5. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The data fusion comprises the following steps: Feature selection and extraction: Select relevant feature variables based on the research objectives and extract these features from the original data; Multi-source data integration: integrating information from multiple data sources including satellite remote sensing, ground sensor networks, drone monitoring, and weather stations; Consistency verification: Verify the consistency between data from different sources through cross-validation and other methods to ensure data quality.

6. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The spatiotemporal alignment comprises the following steps: Time synchronization: The time axis of all data is aligned to the same reference time interval by using resampling technology; Spatial matching: For observation points with different geographical locations, spatial interpolation and smoothing are performed using geographic information system tools so that all data can be represented on the same geographic grid; Resolution adjustment: downscaling high-resolution data or upscaling low-resolution data.

7. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The deep learning model layer adopts a spatiotemporal attention mechanism to adaptively assign spatiotemporal feature weights.

8. The deep learning-based air pollution source tracking and early warning system according to claim 1 is characterized in that: The source tracing analysis layer is combined with the geographic information system to generate a heat map of the spatiotemporal distribution of pollution sources.

9. The deep learning-based air pollution source tracking and early warning system according to claim 1, characterized in that: The early warning decision-making layer uses transfer learning technology to quickly adapt to the air pollution monitoring needs of different regions.

10. The deep learning-based air pollution source tracking and early warning system according to claim 1, characterized in that: The early warning decision-making layer generates the optimal pollution control plan based on the reinforcement learning algorithm.

Images