Geographical meteorological multi-modal fusion air pollutant monitoring data completion method and device

Abstract

The application discloses a geographical meteorological multi-modal fusion air pollutant monitoring data completion method and device, and belongs to the technical field of deep learning and data processing. The method comprises the following steps: acquiring geographical space information, meteorological information and air pollutant monitoring data of a target region; performing time-space alignment and cleaning on multi-modal data; constructing a time-space graph structure, extracting spatial features by using a graph convolution network, and combining a long short-term memory network and an attention mechanism to enhance time features; capturing site spatial correlation and time dynamics by fusing multi-scale features such as space, time, meteorology and time embedding through a multilayer perception machine; and constructing a conditional diffusion model conforming to multi-modal information feature constraints, and realizing high-precision completion of missing data through a forward noise adding and reverse denoising iterative process under the guidance of fused features. The application fully fuses geographical and meteorological multi-modal information, and significantly improves the time-space consistency and physical rationality of the completion result.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and data processing technology, specifically relating to a method and device for supplementing air pollutant monitoring data through geographic and meteorological multimodal fusion. Background Technology

[0002] Air pollutant monitoring data is crucial for assessing atmospheric environmental quality, formulating environmental protection policies, and safeguarding public health. It typically includes concentration monitoring values ​​of major pollutants such as PM2.5, PM10, SO2, NO2, O3, and CO, and is comprehensively reflected in the Air Quality Index (AQI) to indicate the degree of pollution. This data is mainly collected continuously through fixed ground-based monitoring stations. However, in actual monitoring, data gaps often occur due to factors such as sensor malfunctions, communication interruptions, equipment maintenance, and adverse weather conditions, resulting in incomplete time series or spatial distribution gaps, severely limiting the completeness and application value of the data.

[0003] Currently, common methods for imputing missing values ​​in air pollutant monitoring data mainly include spatial interpolation, time series forecasting, and statistical imputation methods. Spatial interpolation methods (such as inverse distance weighting and Kriging interpolation) primarily estimate values ​​based on the spatial correlation between monitoring stations, but neglect the dynamic characteristics of pollutant changes over time. Time series forecasting methods (such as ARIMA models, exponential smoothing, and recurrent neural networks) focus on learning time dependencies from historical data, but do not fully consider spatial correlations and the influence of the geographical environment. Statistical methods (such as mean imputation and linear interpolation), while simple and easy to implement, struggle to capture complex spatiotemporal nonlinear relationships, resulting in limited accuracy in imputation.

[0004] In summary, existing technologies generally suffer from the following shortcomings: First, most methods only model from a single spatial or temporal dimension, failing to fully integrate the spatiotemporal characteristics of air pollutant monitoring data; second, they do not make sufficient use of multimodal environmental information (such as geospatial features like topography, vegetation cover, and building height, as well as meteorological elements like temperature, humidity, and wind speed), making it difficult to comprehensively reflect the diffusion patterns of pollutants under different geographical and meteorological conditions; and third, they lack implicit constraints on the physical diffusion process of pollutants, resulting in limitations in the spatiotemporal continuity and physical rationality of the completed results.

[0005] Therefore, there is an urgent need to develop a high-precision air pollutant data completion method that can deeply integrate geographic and meteorological multimodal information, take into account spatiotemporal characteristics, and introduce physical consistency constraints, so as to improve the accuracy and reliability of missing data completion and provide higher quality data support for environmental management, public health and scientific research. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides a method and apparatus for supplementing air pollutant monitoring data through multimodal fusion of geographic and meteorological data. The aim is to integrate multimodal geographic and meteorological data, comprehensively extracting multimodal information such as geospatial and meteorological conditions related to pollutant monitoring stations, achieving high-precision data supplementation for missing air pollutant monitoring station data. This ensures the continuity and accuracy of pollutant monitoring data across spatial and temporal scales, providing a reliable data source and research foundation for refining pollutant distribution maps, analyzing pollutant dynamic processes, environmental assessments, and policy formulation.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for supplementing air pollutant monitoring data through geographic and meteorological multimodal fusion, the method comprising:

[0009] Step 1: Acquire multimodal data of the target area and preprocess it. The multimodal data includes geospatial information, air pollutant monitoring data, and meteorological element data.

[0010] Step 2: Encode the time series of air pollutant monitoring data and meteorological element data through time embedding coding, extract station-level spatial features from geospatial information data, and construct a spatiotemporal map structure;

[0011] Step 3: Extract spatial features using graph convolutional networks, extract temporal features by combining long short-term memory networks and attention mechanisms, and fuse spatial features, temporal features, meteorological features and temporal embedding features to generate a unified fused feature representation;

[0012] Step 4: Using the fusion features as conditions, construct a conditional diffusion model based on the denoising diffusion probability model framework. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, generate the completed air pollutant monitoring data.

[0013] Furthermore, in step 1, the preprocessing includes:

[0014] The multimodal data is spatiotemporally aligned to establish a unified spatiotemporal reference framework;

[0015] Perform missing value detection and labeling;

[0016] And feature standardization processing is performed on the multimodal data.

[0017] Furthermore, in step 2, the time embedding encoding adopts the sinusoidal position encoding method; the spatiotemporal graph structure uses pollutant monitoring stations as nodes, and the node attributes include time features and spatial features; the spatial proximity relationship between stations is used as edges, and connections are constructed based on a preset distance threshold; the weight of the edges is calculated based on the distance between stations using a distance decay function.

[0018] Furthermore, in step 3, the graph convolutional network extracts spatial dependencies between stations through multi-layer spatial propagation; the long short-term memory network and the attention mechanism work together to extract long-term dependencies of time series and dynamically allocate time step weights; and spatial features, temporal features, meteorological features and temporal embedding features are fused through a multi-layer perceptron.

[0019] Furthermore, in step 4, the training of the conditional diffusion model adopts a composite loss function, which includes: the core noise prediction loss of the diffusion model, the temporal consistency loss, the spatial consistency loss, and the temporal smoothing loss.

[0020] Furthermore, step 4 is followed by a post-processing step:

[0021] Physical constraints are imposed on the completion results, including non-negativity constraints on pollutant concentration and concentration range constraints;

[0022] Spatiotemporal smoothing is performed, and abnormal fluctuations in the completion results are eliminated by moving average filtering.

[0023] Furthermore, in the composite loss function, the weights of noise prediction loss, temporal consistency loss, spatial consistency loss, and temporal smoothing loss are set to 1.0, 0.5, 0.3, and 0.05, respectively.

[0024] On the other hand, the present invention provides a geographic meteorological multimodal fusion air pollutant monitoring data completion device, characterized in that it includes:

[0025] A multimodal data preprocessing module is used to acquire and preprocess multimodal data of the target area, including geospatial information, air pollutant monitoring data, and meteorological element data.

[0026] The spatiotemporal modeling and coding module is used to encode the time series of air pollutant monitoring data and meteorological element data through time embedding coding, extract station-level spatial features from geospatial information data, and construct a spatiotemporal map structure.

[0027] The multi-scale feature extraction and fusion module is used to extract spatial features using graph convolutional networks, extract temporal features by combining long short-term memory networks and attention mechanisms, and fuse spatial features, temporal features, meteorological features and temporal embedding features to generate a unified fused feature representation.

[0028] The conditional diffusion completion module is used to construct a conditional diffusion model based on the denoising diffusion probability model framework, using the fused features as conditions. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, it generates completed air pollutant monitoring data.

[0029] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for supplementing geographic and meteorological multimodal fusion air pollutant monitoring data.

[0030] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for completing geographic and meteorological multimodal fusion air pollutant monitoring data.

[0031] The beneficial effects of this invention are as follows:

[0032] Significantly improved completion accuracy: By deeply integrating multimodal information such as geospatial features and meteorological elements, and introducing a conditional diffusion generation model, high-precision completion of missing parts of air pollutant monitoring data was achieved. Experiments show that the overall completion correlation coefficient R... 2 With a mean squared error (MSE) of 0.96 or higher, its key indicators, such as mean squared error (MSE) and mean absolute error (MAE), are significantly better than traditional Kriging interpolation methods and random forest-based time series forecasting methods.

[0033] Good spatiotemporal consistency and physical plausibility: The spatial dependencies between monitoring stations are effectively modeled using a graph convolutional network (GCN), and long short-term memory (LSTM) network with attention mechanism is combined to capture long-term temporal dependency patterns. The completed data generated by this method is spatially smooth and temporally continuous, and a spatiotemporal consistency constraint is introduced through a composite loss function, making the completed results more consistent with the physical laws of pollutant diffusion.

[0034] High adaptability to complex missing data patterns: This method does not rely on complex physicochemical models, but learns data distribution through a data-driven approach. Whether the missing data is random or continuous, the method demonstrates good adaptability and robustness, making it highly valuable for practical applications. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method for supplementing air pollutant monitoring data through geographic meteorological multimodal fusion according to the present invention;

[0036] Figure 2 This provides multimodal geographic and meteorological data related to air pollutants.

[0037] Figure 3 Complete the scatter plot for the effects of different pollutants;

[0038] Figure 4Comparison of accuracy for completing missing values ​​for different air pollutants. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] Reference Figure 1 This invention provides a method for completing air pollutant monitoring data using a multimodal fusion of geographic and meteorological data. Considering the combined effects of geospatial information and meteorological elements on air pollutants under different geographic conditions, and integrating multimodal data such as topography, vegetation cover, building height, and meteorological elements, the method utilizes graph neural networks and LSTM models to extract spatial and temporal multi-scale features from the multimodal data of pollutant monitoring stations. An air pollutant training set is constructed, and a data completion model for pollutant monitoring stations based on geographic space and meteorological elements is established. This model is suitable for high-precision completion of missing information from air pollutant monitoring stations under different spatiotemporal conditions. The method includes the following four steps:

[0041] Step 1, Multimodal Data Acquisition and Preprocessing: Acquire geospatial information, air pollutant monitoring data and meteorological element data of the target area, perform spatiotemporal alignment and cleaning on the multimodal data, construct a unified spatiotemporal reference framework and mark missing values.

[0042] In step 1, the data source of the present invention (refer to...) Figure 2 The dataset includes geospatial information data with multi-scale spatial raster features (topographic raster data, vegetation cover raster data, building height vector data), vector coordinate data of air pollutant monitoring stations, hourly continuous meteorological monitoring data (precipitation PRE, temperature TMP, air pressure PRS, humidity SHU, wind speed WIN, and wind direction GST structured text data), and 24-hour dynamic monitoring data of air pollutants. To complete the pollutant data, the daily air pollutant monitoring data is divided into training data, test data, and validation data. The training data is used to train the completion model, while the test and validation data are used for completion testing and accuracy verification. The geographical and temporal range of the data covers the Guangxi Zhuang Autonomous Region for 24 hours up to October 1, 2022.

[0043] The process of multimodal data multi-scale spatiotemporal alignment includes data loading and validation, spatial and dimensional consistency checks, and time series integrity checks. Multimodal data loading and validation for air pollutant monitoring stations is completed through file path verification and existence checks, data format consistency verification, and coordinate range rationality checks. The spatial consistency check involves accurately mapping the geographic coordinates of pollutant monitoring stations to raster coordinates for DEM (Digital Elevation Model) and NDVI raster values, checking whether station information is within the valid range of the raster data, and using the neighborhood mean to measure spatial characteristics to avoid errors caused by anomalies and missing information at single points in geospatial data. Dimensional consistency checks include checking the feature dimensions and time dimension of the geospatial information data in which the station is located. The time series integrity check processes meteorological elements, pollutant types, and time series separately, detecting and marking the locations of missing values ​​to maintain the continuity of the time series.

[0044] A unified spatiotemporal reference framework was established through spatiotemporal alignment of multimodal data; a multi-level missing value detection and labeling system was established; a temporal feature dimension consistency mechanism was established, and dimension verification and automatic adjustment mechanisms were added to avoid the problem of spatiotemporal feature dimension mismatch during training and inference.

[0045] Step 2, Multi-scale spatiotemporal modeling and coding of data: Sinusoidal position coding is used to encode the position of each time step of the time series of meteorological and air pollutant monitoring data through time embedding coding, which provides station time sequence information and enhances the learning ability of time series patterns, and is conducive to capturing spatial and temporal dependency relationship features; station-level spatial features are extracted from geospatial raster data to construct a spatiotemporal map structure.

[0046] In step 2, the multi-modal data multi-scale spatiotemporal features are constructed. For spatial features, they are divided into station-level features and regional features: (1) Station-level features are obtained based on station information, including station geospatial features and station observations; (2) Regional features include spatial distance relationships between stations, which are used to model spatial dependencies through graph structures. For station features in the time series domain, hourly pollutant concentration changes and meteorological element fluctuation features are established on the time scale. The feature type scale includes original observation features, namely pollutant concentration measurements, meteorological element observations, original pixel values ​​of raster data, and standardized feature values, time series differential features, and spatial domain statistical features.

[0047] Data standardization for hourly time-series pollutant concentration changes and the temporal characteristics of meteorological element fluctuations includes feature standardization and time standardization.

[0048] Feature standardization employs Z-score standardization to normalize the standard deviation of the original feature information, preserving numerical variation characteristics over time. Since air pollutant data typically ranges from 0 to 300, the original values ​​and variation ranges (dimensions) of meteorological data such as temperature (TMP, usually in degrees Fahrenheit) and air pressure (PRS, in hectopascals) are completely different. Z-score standardization is necessary to unify these values ​​to the same dimension, facilitating feature comparison and fusion, and subsequent training of the pollutant completion model, thereby improving model training stability and convergence speed. For the temporal standardization of pollutant concentration changes and meteorological elements, the original time information (such as year, month, day, hour, etc.) is transformed into a series of standardized, normalized feature vectors containing periodic and ordinal information, enabling the model to effectively understand and learn the dynamic patterns of the time series.

[0049] The temporal standardization of pollutant concentration changes and meteorological elements mainly comprises two core components: sequence length standardization and temporal embedding standardization. Sequence length standardization provides the model with a fixed-length, continuous, and aligned temporal input window. This invention uniformly uses a 12-hour sequence length, which can capture short-periodic changes in pollutant and meteorological element information within a day (such as morning and evening rush hours), while also keeping the computational burden manageable. Dynamic adjustment adapts to data volume limitations, ensuring data dimensionality consistency. Temporal embedding standardization transforms the original timestamps of pollutant concentration changes and meteorological elements into continuous low-dimensional vectors implying time periods. This method employs sine / cosine positional encoding.

[0050] ,

[0051] In the formula, It is the position code value with position pos and dimension 2i (even-numbered dimensions); It is an absolute time position; 2i is the total dimension of the embedding vector; 2i is the dimension index of the positional encoding in the embedding vector.

[0052] Spatial feature modeling transforms continuous raster data (DEM, vegetation index, and building height) covering the entire study area into feature values ​​that correspond one-to-one with discrete monitoring stations. By extracting raw values ​​or the mean of a 5km radius neighborhood from the raster data at each station location, a precise mapping from geographic coordinates to raster coordinates is achieved. This process extracts the DEM features, NDVI features, and building height features for each station, reflecting the impact of geospatial features on the dispersion of air pollutants.

[0053] Graph structure construction transforms the spatial information related to scattered and heterogeneous monitoring stations into a structured and unified data graph, enabling graph neural networks to effectively perform message passing and information aggregation on the graph. The graph structure construction process includes coordinate system 1 and graph structure standardization. Coordinate system 1 refers to adopting the WGS84 coordinate system, a unified distance calculation standard, and a consistent neighborhood search radius. Then, graph structure standardization transforms the standardized coordinate and distance information into a data structure that can be directly used by the subsequent graph neural network, employing a unified edge connection threshold, standardized edge weight calculation, and a consistent graph neural network input format.

[0054] The construction of a graph structure is the creation of a graph structure. In the formula, V represents a node, signifying each spatial pollutant monitoring station. Each node possesses a series of attributes, such as temporal and spatial characteristics. E indicates the interaction or influence relationship between stations, defined using spatial proximity. W represents the intensity of the spatial influence; the closer the distance, the greater the intensity. When constructing the graph structure, the temporal and spatial characteristics of the stations are assigned to the corresponding nodes. Based on empirical values ​​from meteorology and atmospheric physics, a local neighborhood range of 50 km is determined as an empirical value for edge calculation. An undirected graph is constructed based on the interactive influence of air pollutant diffusion. Edge weights... The formula for quantifying the influence between nodes is as follows:

[0055] ,

[0056] In the formula, This represents the distance between nodes m and n, calculated precisely using the Haversine formula; It is a decay coefficient that controls the rate at which the weights decay with distance. The standardized calculation results through the graph structure are organized into the input format of the graph neural network (GNN) to extract the spatial context information of the monitoring station, so that the subsequent model can effectively use this spatial context information to complete and predict air pollutant data.

[0057] The aforementioned spatiotemporal modeling includes site-level features and region-level graph structures. Spatially, the graph structures establish spatial distribution features between sites, while temporally, they provide short-term sequence and trend features, as well as original observations and derived features. This ensures the robustness and spatiotemporal consistency of feature information extraction.

[0058] Step 3, Multi-scale feature extraction and fusion: Spatial features are extracted using graph convolutional networks, and temporal features are extracted by combining long short-term memory networks and attention mechanisms. Spatial features, temporal features, meteorological features and temporal embedding features are fused through multilayer perceptrons to generate a unified fused feature representation.

[0059] In step 3, an air pollutant completion model that integrates geospatial and temporal multi-scale features is constructed. The completion model extracts the spatial dependencies between stations and the long-term temporal patterns of the sequence through a spatial graph structure convolutional neural network and an enhanced temporal model, respectively. These are then aggregated with contextual information such as meteorological and temporal embeddings in a spatiotemporal fusion network to form a high-dimensional fusion feature representation.

[0060] Feature extraction employs a multi-level feature extraction strategy to construct a multi-dimensional vector of fused spatiotemporal features from the original multimodal data, including three core steps: spatial feature extraction, temporal feature extraction, and feature fusion.

[0061] Spatial feature extraction is achieved by constructing a graph structure based on a site network and utilizing a graph convolutional neural network (GCN) to capture spatial dependencies in the data. The spatial propagation formula for each layer is as follows:

[0062] ,

[0063] In the formula, The node features of the l-th layer; This is the normalized adjacency matrix; Let L be the learnable weight matrix of the l-th layer; It is a non-linear activation function.

[0064] Temporal feature extraction, by combining LSTM and an attention mechanism, constructs a collaborative architecture with strong temporal modeling capabilities and dynamic information selection capabilities, forming an enhanced temporal model. LSTM utilizes a gating mechanism—input gate, forget gate, and output gate—to capture the long-term dependencies of time series data by combining 24-hour changes in air pollutants at training data stations with meteorological elements to learn the trends and historical patterns of pollution concentration changes. The attention mechanism, on the other hand, dynamically assigns importance weights to different time steps by calculating the similarity between query, key, and value matrices, highlighting the historical time points with the greatest impact on the current moment. Both work collaboratively in the temporal feature encoding module: LSTM outputs a global temporal representation of the sequence, while the attention mechanism weights the time steps, ultimately fusing the two to form a rich temporal feature representation. This provides more accurate temporal context conditions for subsequent data completion, and improves the model's ability to understand and model the laws of temporal evolution.

[0065] Feature fusion is based on the MLP fusion module, which combines multiple heterogeneous features (spatial features) Time characteristics Meteorological elements and characteristics Temporal embedding features This process integrates features to generate a unified feature representation with contextual information. This is used in the subsequent model generation process. Spatial features include three raster features: DEM, NDVI, and building height. Temporal pollutant features include six pollutant features over 24 hours, and meteorological element features include six meteorological element features over 24 hours. The temporal embedding vector provides the model with periodic information that is not easily obtained directly from the raw numerical values. For example, pollutant diffusion features have a strong diurnal cycle; directly inputting the raw hour numbers is not conducive to the model and makes it difficult to learn this cyclical relationship. Temporal embedding maps discrete, monotonous time points (such as the t-th hour) into a continuous, low-dimensional vector that reflects its periodicity.

[0066] ,

[0067] The aforementioned feature extraction and feature fusion method is based on the following steps: First, a graph convolutional neural network (GCN) is used as the basic module to construct a graph convolutional neural network to model the spatial dependencies between sites. Second, an augmented temporal model is used to extract long-term dependencies and key time point information of the time series. Finally, a multilayer perceptron (MLP) is used to concatenate and nonlinearly transform multiple features to form a unified feature representation with spatiotemporal semantics, providing accurate contextual conditions for high-quality data completion of the subsequent diffusion model.

[0068] Step 4, Conditional Diffusion Completion: Using the fused features as conditions, a conditional diffusion model is constructed based on the Denoising Diffusion Probability Model (DDPM) framework. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, the completed air pollutant monitoring data is generated.

[0069] The air pollutant incompleteness model, which fuses geospatial and temporal multi-scale features, is based on a conditional diffusion model constructed using the Denoising Diffusion Probability Model (DDPM) framework. Through a reversible iterative process of noise addition (forward process) and noise removal (backward process), it learns the generation process of the data distribution and can gradually "recover" missing data, resulting in good consistency of the incompleteness results both temporally and spatially. The fusion features, as strong "conditional" information, are introduced into the incompleteness network. The network not only receives noisy input data but also senses the diffusion progress through time-step encoding. Guided by the fusion features, it iteratively and selectively removes noise, ultimately generating spatially smooth, temporally continuous, and physically consistent air quality incompleteness data. The forward diffusion process (noise addition) formula is as follows:

[0070] ,

[0071] In the formula, q represents the conditional probability distribution of forward diffusion; This represents a normal distribution (Gaussian distribution). This is the noise data at step t; It is the first Accurate sample data for each step; It is the diffusion coefficient at step t; Let I represent the covariance term of the normal distribution, and let I be the identity matrix. It is used for training through reparameterization.

[0072] , ,

[0073] In the formula, Original, noise-free data; ~N(0, I) is standard Gaussian noise used to simulate random disturbances; The cumulative α value represents the proportion of the original data retained.

[0074] The backward process is crucial for model learning, starting from noisy data. China is gradually "restoring" the original data. Its formula is:

[0075] ,

[0076] In the formula, It is the accurate sample data at step t-1; It is the predicted mean of the model output; It is the variance of the prediction.

[0077] During training, a composite loss function is formed by using mean squared error (MSE) as the loss function and combining it with spatiotemporal consistency constraint loss to prevent overfitting. The formula is as follows:

[0078] ,

[0079] In the formula, The core loss of the diffusion model is the mean square error (MSE) between the noise predicted by the model and the actual noise. This results in a loss of time consistency. This represents a loss of spatial consistency. For time smoothing loss; As weights. Considering the consistency and continuity of pollutant concentration monitoring at various stations over time, and the invariance of spatial scale elements, to avoid abrupt changes in the completion results along the time axis, , and The values ​​were set to 0.5, 0.3, and 0.05, respectively.

[0080] The model training process employs an end-to-end training strategy, optimizing model parameters through multiple iterations to ensure the diffusion model accurately learns the distribution characteristics of air quality data. The training process is divided into four stages: data preparation, model initialization, training loop, and model saving. An early stopping mechanism is used to prevent overfitting, and model performance is monitored using a validation set.

[0081] In the training data preparation phase, the preprocessed multi-source feature data was constructed into spatiotemporal sequence samples. A data preprocessing module was used to perform batch loading and boundary processing on the data. The air pollutant and meteorological element data from the stations were divided into training and test sets in a ratio of approximately 8:2, with the training set consisting of 52 stations and the test set consisting of 6 stations.

[0082] The model training employs an improved trainer, whose core training loop includes four steps: forward propagation, loss calculation, backpropagation, and parameter update. During training, gradient explosion is controlled using gradient clipping techniques, and the Adam optimizer is used for parameter optimization, with the learning rate set to 1e-4.

[0083] An early stopping mechanism is employed during training to obtain the optimal training model. Training automatically stops and the best model is saved when the validation set loss no longer decreases for 10 consecutive epochs (training rounds). This mechanism effectively prevents overfitting and ensures the model's generalization performance. After training, the model weights are saved as a `best_improved_model.pth` file, containing information such as the model state dictionary, training rounds, and validation loss.

[0084] The high-precision missing value completion process employs a conditional diffusion generation strategy, applying the trained model to complete missing values ​​in the test data. The completion process consists of four stages: missing value detection, conditional diffusion generation, iterative optimization, and post-processing. First, by detecting NaN and zero values ​​in the test data, a missing value mask matrix is ​​constructed to accurately identify the locations of data that need to be completed.

[0085] During conditional diffusion completion, fused features serve as strong conditional information guiding the generation direction of the denoising network. The completion process directly invokes the model's forward propagation, generating completed values ​​that conform to the data distribution through a single-step diffusion process. During completion, the model directly predicts air quality values ​​at missing locations based on fused spatiotemporal features (topography, meteorology, station relationships, etc.), keeping known observations unchanged and only completing the missing locations, ensuring that the completion process does not compromise the integrity of the original data.

[0086] The iterative optimization strategy gradually improves accuracy through multiple rounds of completion. Each round uses the completion result from the previous round as input for a new round of diffusion completion. This iterative approach can utilize the already completed information to improve the subsequent completion results.

[0087] The post-processing stage applies physical constraints and spatiotemporal smoothing to ensure the rationality of the completed results. Physical constraints include non-negativity constraints on pollutant concentrations and concentration range constraints, such as limiting PM2.5 concentrations to 0-500 μg / m³. 3 Within the specified range. Spatiotemporal smoothing eliminates abnormal fluctuations in the completion results through moving average filtering, maintaining the continuity of data in both time and space dimensions.

[0088] The experimental accuracy verification employs a multi-index comprehensive evaluation system to validate the accuracy of the completion results, ensuring the effectiveness and reliability of the completion method. The verification process includes data preparation, accuracy index calculation, and visualization analysis to comprehensively evaluate the quality of the completion results.

[0089] The validation data represents the actual values ​​of the missing items. By comparing the differences between the completed data and the validation data at the missing locations, various accuracy indicators are calculated to objectively evaluate the completion effect.

[0090] Accuracy assessment uses five core indicators: mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Mean Absolute Percentage Error (MAPE).

[0091] The model was used to complete 35 missing values ​​for different air pollutants at different times for 6 stations in the test set. The results were compared with the corresponding validation set values. The results are shown in Table 1 below:

[0092] Table 1

[0093]

[0094] The overall accuracy assessment is shown in Table 2:

[0095] Table 2

[0096]

[0097] like Figures 3-4 As shown, the method of this invention maintains stable completion performance under different pollutant types and different missing modes, with an overall accuracy of 96.08%. The completion accuracy is particularly outstanding for major pollutants such as PM2.5, PM10, and SO2. However, the accuracy is slightly lower for NO2 and CO due to larger fluctuations in local areas, leading to some error between predicted and actual values. Of course, the excellent completion accuracy is also due to the fact that air pollutants (such as SO2) are at their minimum values ​​over time, and the attention mechanism optimized for spatiotemporal consistency makes their predicted values ​​more accurate, resulting in higher data completion accuracy.

[0098] Compared with the accuracy of traditional interpolation methods and time series prediction methods, the accuracy of the method of this invention, the traditional kriging interpolation method, and the random forest-based time series prediction method for completing missing values ​​is compared in Table 3:

[0099] Table 3

[0100]

[0101] Compared with traditional Kriging interpolation and random forest-based time series prediction methods, the overall accuracy of the method in this invention is better than the other two. This shows that the method is better than traditional methods in terms of completion effect for different missing patterns such as random missing and continuous missing. It also proves the wide applicability and robustness of the method in practical applications.

[0102] On the other hand, the present invention provides a geographic meteorological multimodal fusion air pollutant monitoring data completion device, which includes various modules capable of implementing the various steps of the aforementioned method, specifically including:

[0103] A multimodal data preprocessing module is used to acquire and preprocess multimodal data of the target area, including geospatial information, air pollutant monitoring data, and meteorological element data.

[0104] The spatiotemporal modeling and coding module is used to encode the time series of air pollutant monitoring data and meteorological element data through time embedding coding, extract station-level spatial features from geospatial information data, and construct a spatiotemporal map structure.

[0105] The multi-scale feature extraction and fusion module is used to extract spatial features using graph convolutional networks, extract temporal features by combining long short-term memory networks and attention mechanisms, and fuse spatial features, temporal features, meteorological features and temporal embedding features to generate a unified fused feature representation.

[0106] The conditional diffusion completion module is used to construct a conditional diffusion model based on the denoising diffusion probability model framework, using the fused features as conditions. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, it generates completed air pollutant monitoring data.

[0107] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for supplementing geographic and meteorological multimodal fusion air pollutant monitoring data.

[0108] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for completing geographic and meteorological multimodal fusion air pollutant monitoring data.

[0109] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for supplementing air pollutant monitoring data through geographic and meteorological multimodal fusion, characterized in that, The method includes: Step 1: Acquire multimodal data of the target area and preprocess it. The multimodal data includes geospatial information, air pollutant monitoring data, and meteorological element data. Step 2: Encode the time series of air pollutant monitoring data and meteorological element data through time embedding coding, extract station-level spatial features from geospatial information data, and construct a spatiotemporal map structure; wherein, the time embedding coding adopts the sinusoidal position coding method; the spatiotemporal map structure uses pollutant monitoring stations as nodes, and the node attributes include time features and spatial features; the spatial proximity relationship between stations is used as edges, and connections are constructed based on a preset distance threshold; the weight of the edges is calculated based on the distance between stations using a distance decay function; Step 3: Extract spatial features using a graph convolutional network, extract temporal features by combining a long short-term memory network and an attention mechanism, and fuse spatial features, temporal features, meteorological features, and temporal embedding features to generate a unified fused feature representation; wherein, the graph convolutional network extracts spatial dependencies between stations through multi-layer spatial propagation; the long short-term memory network and the attention mechanism collaboratively extract long-term dependencies of the time series and dynamically allocate time step weights; and fuse spatial features, temporal features, meteorological features, and temporal embedding features through a multi-layer perceptron. Step 4: Using the fusion features as conditions, construct a conditional diffusion model based on the denoising diffusion probability model framework. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, generate complete air pollutant monitoring data. The training of the conditional diffusion model adopts a composite loss function, which includes: the core noise prediction loss of the diffusion model, the temporal consistency loss, the spatial consistency loss, and the temporal smoothing loss.

2. The method for supplementing air pollutant monitoring data through geographic meteorological multimodal fusion according to claim 1, characterized in that, In step 1, the preprocessing includes: The multimodal data is spatiotemporally aligned to establish a unified spatiotemporal reference framework; Perform missing value detection and labeling; And feature standardization processing is performed on the multimodal data.

3. The method for supplementing air pollutant monitoring data through geographic meteorological multimodal fusion according to claim 1, characterized in that, Step 4 is followed by a post-processing step: Physical constraints are imposed on the completion results, including non-negativity constraints on pollutant concentration and concentration range constraints; Spatiotemporal smoothing is performed, and abnormal fluctuations in the completion results are eliminated by moving average filtering.

4. The method for supplementing air pollutant monitoring data through geographic meteorological multimodal fusion according to claim 1, characterized in that, In the composite loss function, the weights of noise prediction loss, temporal consistency loss, spatial consistency loss, and temporal smoothing loss are set to 1.0, 0.5, 0.3, and 0.05, respectively.

5. A geographic meteorological multimodal fusion air pollutant monitoring data completion device, applied to the method described in any one of claims 1-4, characterized in that, include: A multimodal data preprocessing module is used to acquire and preprocess multimodal data of the target area, including geospatial information, air pollutant monitoring data, and meteorological element data. The spatiotemporal modeling and coding module is used to encode the time series of air pollutant monitoring data and meteorological element data through time embedding coding, extract station-level spatial features from geospatial information data, and construct a spatiotemporal map structure. The multi-scale feature extraction and fusion module is used to extract spatial features using graph convolutional networks, extract temporal features by combining long short-term memory networks and attention mechanisms, and fuse spatial features, temporal features, meteorological features and temporal embedding features to generate a unified fused feature representation. The conditional diffusion completion module is used to construct a conditional diffusion model based on the denoising diffusion probability model framework, using the fused features as conditions. Through the iterative process of adding forward noise and denoising backward, combined with physical constraints and spatiotemporal consistency loss, it generates completed air pollutant monitoring data.

6. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the geographic meteorological multimodal fusion air pollutant monitoring data completion method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the geographic meteorological multimodal fusion air pollutant monitoring data completion method as described in any one of claims 1-4.

Images