An air pollution prediction method fusing meteorological driving mechanism constraints

Abstract

The application discloses a kind of air pollution prediction methods of fusion meteorological driving mechanism constraint, belong to air quality prediction field.The application obtains the time series input data of the multi-source data of target area, extracts key meteorological variable and constructs meteorological driving characteristic vector, constructs graph structure based on monitoring site and utilizes graph attention network to extract pollutant space propagation characteristic, utilizes time convolution network to extract pollutant time evolution characteristic;Joint loss function containing prediction error term and physical constraint term is constructed, the physical mapping relationship constructed based on meteorological driving characteristic is embedded as constraint term in model training process;Based on the model that training is completed, the prediction result of different time scale is output.The application embeds meteorological driving mechanism in the form of explicit physical constraint in deep learning model training, realizes the collaborative modeling of data driving and physical mechanism, improves the physical consistency, explainability and prediction stability under complex meteorological conditions of prediction result.

Description

Technical Field

[0001] This invention relates to the fields of air quality forecasting and machine learning, and specifically to an air pollution forecasting method that integrates meteorological driving mechanism constraints. Background Technology

[0002] Air pollution forecasting is a crucial research area in atmospheric environmental science. The formation and evolution of pollutants such as PM2.5 and O3 are influenced not only by emission sources but also by meteorological conditions. Factors such as temperature, humidity, wind speed, and boundary layer height affect the diffusion, transport, and chemical reaction processes of pollutants, thus determining air quality trends. Therefore, accurately characterizing the driving effect of meteorological conditions on pollutants is key to achieving high-precision air quality forecasting.

[0003] Currently, air pollution prediction methods mainly include numerical simulation methods, statistical methods, and machine learning methods:

[0004] While numerical simulation methods (such as the WRF-CMAQ coupled model) can describe atmospheric physical and chemical processes well, they are computationally complex, have long running times, and are highly dependent on high-precision input data and emission inventories, resulting in high computational resource and operation and maintenance costs in practical operational applications.

[0005] Statistical methods, such as multiple linear regression, time series analysis, and autoregressive moving average models, are simple in structure and computationally efficient, and have certain applicability under conditions of small samples or stationary series. However, these methods typically assume that the data follows a specific statistical distribution, making it difficult to characterize the complex nonlinear spatiotemporal dependencies in air quality data. Furthermore, they have limited ability to model nonlinearity, nonstationarity, and interactions between multiple variables, resulting in insufficient prediction accuracy when dealing with drastic changes in meteorological conditions and the nonlinear evolution of pollutants.

[0006] In recent years, deep learning methods have been widely applied in the field of air quality prediction due to their powerful nonlinear modeling capabilities. In particular, the introduction of Graph Neural Networks (GNNs) has made it possible to model the complex spatial relationships between monitoring stations. Existing technologies have made significant progress in pollutant concentration prediction by constructing static or dynamic graph structures and combining them with temporal processing methods such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Temporal Convolutional Networks (TCNs), or Transformers.

[0007] However, existing deep learning methods generally suffer from the problem of "strong fitting, weak mechanism," meaning that model training mainly relies on data-driven fitting and lacks explicit constraints on the inherent physical laws governing the transport and diffusion of atmospheric pollutants. This leads to systematic biases and physical inconsistencies in prediction results under complex meteorological conditions (such as a sudden shift from stable weather to strong winds). Specifically, existing dynamic graph neural networks often establish a simple linear relationship between increased wind speed and pollutant transport rates when meteorological conditions change drastically, ignoring the inherent "dilution before transport" temporal characteristics of atmospheric boundary layer physics. This results in causal confusion at meteorological turning points. For example, when strong winds arrive, the model may immediately increase the connection weights of downwind stations and predict an increase in their concentrations. However, in reality, strong winds first dilute local pollutants, and the concentration at downwind stations only increases after a sustained pollution source upwind and a certain time lag. This neglect of the time constant in the physical mechanism causes existing methods to produce predictions that contradict the actual laws when weather systems undergo abrupt changes.

[0008] To address the aforementioned issues, CN114444796A discloses a cross-regional air pollution prediction method and system based on graph neural networks. This method uses pollutant concentrations simulated by a Gaussian diffusion model as one of the input variables, combining it with graph convolutional networks, GRU, and Transformer for prediction. However, this approach only uses the output of the physical model as an additional input feature, making the constraints of the physical mechanism on the model indirect. The model may still "forget" or deviate from physical laws during the learning process, making it difficult to guarantee the physical consistency of the prediction results. CN121524920A discloses an air quality prediction optimization method and system based on deep learning. This method calculates the gradient and curvature of the predicted concentration field using automatic differentiation techniques, constructs a composite optimization objective including physical consistency residuals, and uses an alternating optimization strategy for parameter updates. However, the embedding of physical constraints in this approach relies on the spatial gradient calculated by automatic differentiation, making the physical mapping relationship relatively indirect. Furthermore, the alternating optimization strategy causes the physical constraints and data fitting to separate during training, failing to achieve true collaborative optimization.

[0009] Therefore, how to effectively integrate meteorological driving mechanisms into deep learning models, achieve collaborative modeling of data-driven and physical mechanisms, and ensure that the prediction results not only conform to the observation data but also follow the physical laws of atmospheric pollutant diffusion, especially in complex terrain areas and meteorological transition periods, is a key technical problem that urgently needs to be solved in the field of air pollution prediction. Summary of the Invention

[0010] To address the issues of existing air pollution prediction methods lacking physical constraints and exhibiting poor prediction stability in complex terrain areas, this invention provides an air pollution prediction method that integrates meteorological driving mechanism constraints. By constructing an interpretable physical mapping function based on meteorological driving features and embedding it as an explicit constraint term into the joint loss function, the deep learning model can simultaneously optimize prediction accuracy and physical consistency during training. This achieves collaborative modeling of data-driven and physical mechanisms, fundamentally solving the problems of physical inconsistency and poor stability in prediction results under complex meteorological conditions in existing methods.

[0011] The technical solution adopted by this invention to solve its technical problem is:

[0012] An air pollution prediction method incorporating meteorological driving mechanisms includes the following steps:

[0013] Step 1), Multi-source data construction and preprocessing

[0014] Acquire multi-source data of the target area, construct and preprocess time-series input data, wherein the multi-source data includes air quality monitoring data, meteorological observation data and reanalysis data;

[0015] Multi-source data, including air quality monitoring data, meteorological observation data, and reanalysis data, are used to construct a time-series input feature vector. ,in For pollutant concentration data, For meteorological variable data; standardize the data to eliminate the influence of dimensions;

[0016] Step 2), Construction of meteorological driving features

[0017] Extract key meteorological variables, including temperature. relative humidity Wind speed Boundary layer height air pressure Construct meteorological driving feature vectors ,in It is a feature mapping function used to characterize the driving effect of meteorological elements on the diffusion and transport of pollutants;

[0018] Step 3), Spatial Dependency Modeling

[0019] Based on the graph structure constructed by monitoring stations, with stations as nodes and spatial adjacency as edges, a graph attention network is used to model the spatial dependencies between different stations and extract the spatial propagation characteristics of pollutants.

[0020] Step 4), Time Series Evolution Modeling

[0021] Temporal convolutional networks are used to model the temporal evolution of pollutants, and long-term temporal dependencies are captured by expanding the convolutional structure.

[0022] Step 5), embedding of meteorological driving mechanism constraints

[0023] Constructing a joint loss function ,in, For the prediction error term, For physical constraint terms, As the weighting coefficient, the physical constraint term constructs a physical mapping relationship based on meteorological driving characteristics to ensure that the prediction results conform to the laws of pollutant diffusion and meteorological driving.

[0024] Step 6), Multi-scale prediction output

[0025] Based on the trained model, air pollution prediction results at different time scales are output, including short-term prediction (2 hours), medium-term prediction (3 days), and extended-term prediction (7 days).

[0026] Advantages or beneficial effects of the present invention:

[0027] (1) Explicit embedding of physical constraints to achieve deep synergy between data-driven approaches and physical mechanisms:

[0028] This invention differs from existing technologies that treat the physical model output as merely input features or employ separate alternating optimization strategies. It is the first to propose embedding the meteorological driving mechanism as an explicit physical constraint term into the joint loss function of a deep learning model. This is achieved by constructing... The joint optimization objective ensures that the model is driven by both data fitting error and physical law deviation during backpropagation, guaranteeing that model parameters are updated in a direction that simultaneously satisfies data consistency and physical consistency at each training step. This "true fusion" mechanism fundamentally solves the problem of separation between physical constraints and data-driven learning in existing methods, significantly improving the physical rationality and predictive reliability of the model under sudden changes in meteorological conditions.

[0029] (2) Constructing interpretable meteorological driving characteristics to enhance the physical interpretability of the model:

[0030] This invention breaks through the existing technology's approach of simply stacking or implicitly encoding meteorological factors. It extracts key meteorological variables such as temperature, relative humidity, wind speed, boundary layer height, and air pressure, constructing a nonlinear mapping feature vector with clear physical meaning. Furthermore, it introduces second-order interaction terms (such as the product of wind speed and boundary layer height) to characterize the synergistic influence of meteorological elements on pollutant diffusion capabilities. Based on an empirical meteorological diffusion model, this invention constructs a physical mapping function, where wind speed and boundary layer height reflect diffusion and dilution capabilities, relative humidity reflects secondary generation potential, and temperature reflects chemical reaction rates. This physical mapping function possesses clear interpretability, ensuring that the prediction results not only conform to statistical data patterns but also follow the fundamental physical principles of atmospheric pollutant transport and diffusion, effectively improving the model's interpretability and reliability.

[0031] (3) Coupled spatial-temporal-mechanism modeling to improve the prediction stability of complex terrain regions:

[0032] This invention integrates Graph Attention Networks (GAT) and Temporal Convolutional Networks (TCN), along with meteorological driving mechanism constraints, to achieve unified and collaborative modeling of spatial dependence, temporal evolution, and physical mechanisms. The Graph Attention Network, through self-attention mechanism, adaptively learns the spatial dependencies between monitoring stations, effectively capturing the asymmetric propagation characteristics of pollutants in complex terrains (such as karst mountains). The Temporal Convolutional Network, through its expanded convolutional structure, captures the long-term temporal dependence of pollutant concentrations, avoiding the gradient vanishing problem in long-sequence modeling of recurrent neural networks. Meteorological driving constraints ensure that the spatiotemporal modeling process always follows physical laws. The organic combination of these three elements significantly improves the model's predictive stability in complex terrain areas and during meteorological transitions (such as from calm to strong winds), effectively avoiding causal confusion issues that arise in existing methods during sudden meteorological changes.

[0033] (4) The integrated architecture of multi-scale prediction enhances the value of engineering applications:

[0034] This invention achieves multi-timescale output of short-term (2-hour), medium-term (3-day), and extended-term (7-day) forecasts within a unified model framework, eliminating the need to train multiple models separately for different forecast timeframes. This integrated architecture simplifies the deployment and maintenance process of the forecasting system and reduces engineering implementation costs. Simultaneously, the multi-scale forecasting capability meets the needs of different business scenarios: short-term forecasts serve pollution emergency early warning, medium-term forecasts support air quality control decisions, and extended-term forecasts assist in seasonal pollution prevention and control planning, significantly enhancing the comprehensive application value of the forecasting system.

[0035] (5) Adapting to the physical laws governing meteorological transition periods to resolve the problem of causal confusion:

[0036] This invention pays particular attention to the predictive physical consistency in scenarios of rapid changes in meteorological conditions (such as a sudden shift from stable weather to strong winds). By explicitly embedding a diffusion mechanism based on the coupling of boundary layer height and wind speed, a secondary generation mechanism based on humidity, and a reaction rate mechanism based on temperature into physical constraints, the model ensures that it still follows the true physical time-series characteristic of "dilution before transmission" during meteorological transitions, rather than simply linearly linking wind speed increase with downwind concentration increase. This feature effectively solves the causal confusion problem commonly found in existing technologies, ensuring that prediction results remain consistent with the actual physical processes when weather systems undergo abrupt changes. Attached Figure Description

[0037] Figure 1 This is a flowchart of the air pollution prediction method described in this invention;

[0038] Figure 2 This is a schematic diagram of the air pollution prediction model structure described in this invention, which includes a graph attention network module, a temporal convolutional network module, and a physical constraint embedding module. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. This embodiment uses a karst mountain city as an example to specifically illustrate the method of the present invention, but the application scope of the present invention is not limited thereto.

[0040] Example:

[0041] like Figure 1 As shown, the air pollution prediction method proposed in this invention, which integrates meteorological driving mechanism constraints, specifically includes the following steps:

[0042] Step (1), Multi-source data construction and preprocessing

[0043] Hourly data were obtained from 20 air quality monitoring stations within the study area from January 2019 to December 2025. Data sources include:

[0044] ① Air quality monitoring data: PM2.5, PM10, O3, NO2, SO2, CO concentrations;

[0045] ② Meteorological observation data: temperature, relative humidity, wind speed, wind direction, and air pressure;

[0046] ③ Reanalysis data: Boundary layer height, derived from ERA5 reanalysis data;

[0047] First, construct the input feature vector. ,in For pollutant concentration data, For meteorological variable data, Z-score normalization is then used to process each feature dimension independently.

[0048] (1)

[0049] in, The sample mean. To avoid data leakage, the standardization parameters are calculated only based on the training set and applied to the validation and test sets, representing the standard deviation.

[0050] Step (2), Construction of meteorological driving features

[0051] Extract key meteorological variables and construct meteorological driving feature vectors:

[0052] (2)

[0053] In this embodiment, the feature mapping function Nonlinear transformation is adopted, specifically by normalizing each meteorological variable and then splicing them together, and introducing second-order interaction terms, such as the product of wind speed and boundary layer height, to characterize the synergistic influence of meteorological elements on the ability of pollutants to diffuse.

[0054] Step (3), Spatial Dependency Modeling

[0055] Based on the latitude and longitude coordinates of the monitoring stations, a spatial map structure is constructed using the K-nearest neighbor method. , among which, nodes Corresponding monitoring stations, edge Connect the K closest stations; in this example, K=5. The edge weights are calculated using a Gaussian kernel function.

[0056] (3)

[0057] Spatial dependencies are modeled using a Graph Attention Network (GAT). The GAT has two layers, each with four attention heads, and a hidden layer dimension of 64. The node update expression is as follows:

[0058] (4)

[0059] in, For nodes In the Layer feature representation; The attention coefficients between nodes are calculated through a shared attention mechanism. The learnable parameter matrix; For nodes The set of adjacent nodes; Using activation functions such as ReLU, GAT is used to extract pollution propagation and spatial correlation features between different sites;

[0060] Step (4), Time Series Evolution Modeling

[0061] A Temporal Convolutional Network (TCN) is used to model the temporal evolution of pollutants. The TCN structure consists of four residual blocks, each containing two dilated convolutional layers with dilation factors of 1, 2, 4, and 8, and a kernel size of 3. The formula for calculating the dilated convolution is as follows:

[0062] (5)

[0063] in, For a moment The simulated output value, For convolution kernel, The kernel size is [size]. As the expansion factor, To prevent overfitting, a batch normalization layer and a Dropout layer (with a Dropout rate of 0.2) are added after each residual block as input features.

[0064] Step (5), embedding of meteorological driving mechanism constraints

[0065] like Figure 2 As shown, the air pollution prediction model constructed in this invention mainly consists of three parts: a graph attention network module, a temporal convolutional network module, and a physical constraint embedding module. Specifically, the graph attention network module described in step (3) is used to capture the spatial dependencies between monitoring stations, the temporal convolutional network module described in step (4) is used to model the temporal evolution of pollutant concentrations, and the physical constraint embedding module integrates the meteorological driving mechanism into the model training process in the form of explicit constraints through a joint loss function. The specific implementation is as follows:

[0066] To enhance the physical consistency of the model, a meteorological driving mechanism constraint is introduced during model training, and a joint loss function is constructed:

[0067] (6)

[0068] in, This is the mean square error loss term. For physical constraint terms, These are the weighting coefficients. It is used to measure the difference between model predictions and actual observations, and its expression is:

[0069] (7)

[0070] in, The total number of samples, These are actual observations. These are the model's predicted values. The expression for the difference between model predictions and physical mechanism inferences is:

[0071] (8)

[0072] in, In this embodiment, the physical mapping function is constructed based on meteorological driving characteristics. An empirical meteorological diffusion model is adopted, specifically in the following form:

[0073] (9)

[0074] in, The emission source intensity is estimated from historical data, either through emission inventory inversion or based on the moving average method from historical monitoring data. , , These are empirical coefficients, determined by fitting historical data using the least squares method, specifically... , , The independent variable is the observed pollutant concentration. Multiple linear regression was performed with wind speed and boundary layer height as the dependent variable. This function characterizes the effects of wind speed and boundary layer height on diffusion capacity, humidity on secondary generation, and temperature on reaction rate. This constraint ensures that the prediction results conform to the laws of pollution diffusion and meteorological driving forces.

[0075] The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. An early stopping mechanism (patience=20) was used to prevent overfitting.

[0076] Step (6),

[0077] Based on the trained model, predictive outputs at different time scales are achieved:

[0078] (10)

[0079] in, These correspond to short-term forecasts, medium-term forecasts, and extended-term forecasts, respectively.

[0080] The time scale for short-term forecasts is 2 hours, for medium-term forecasts it is 3 days, and for extended-term forecasts it is 7 days.

[0081] Through the above steps, the "space-time-mechanism" coupled prediction of air pollution can be achieved.

[0082] This invention deeply integrates meteorological driving mechanisms into the deep learning model training process in the form of explicit physical constraints, realizing the collaborative modeling of data-driven and physical mechanisms. This significantly improves the physical rationality, interpretability, stability, and engineering application value of air pollution prediction results, and provides an effective technical solution for solving the problem of physical consistency in air quality prediction under complex meteorological conditions.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An air pollution prediction method incorporating meteorological driving mechanisms, characterized in that, Includes the following steps: Step 1: Obtain multi-source data of the target area, construct and preprocess time series input data, wherein the multi-source data includes air quality monitoring data, meteorological observation data and reanalysis data; Step 2: Extract key meteorological variables and construct meteorological driving feature vectors; Step 3: Construct a graph structure based on the spatial distribution of monitoring stations, and use a graph attention network to extract the spatial propagation characteristics of pollutants; Step 4: Extract the temporal evolution features of pollutants using a temporal convolutional network; Step 5, construct the joint loss function ,in, For the prediction error term, For physical constraint terms, The physical mapping relationship constructed based on meteorological driving features is used as the weighting coefficient and embedded into the model training process as the physical constraint term. Step 6: Output air pollution prediction results at different time scales based on the trained model.

2. The air pollution prediction method based on the constraints of meteorological driving mechanisms according to claim 1, characterized in that, In the joint loss function, the prediction error term The mean squared error loss function measures the difference between the model's predicted values ​​and the actual observed values; the physical constraint term Used to measure the difference between model predictions and physical mechanism inferences, its expression is: ,in These are the model's predicted values. The feature vector is driven by meteorological conditions. This is the physical mapping function.

3. The method according to claim 1, characterized in that, The key meteorological variables include air temperature, relative humidity, wind speed, boundary layer height, and air pressure; the meteorological driving feature vector is constructed by performing a nonlinear transformation on the key meteorological variables through a feature mapping function and introducing a second-order interaction term.

4. The method according to claim 2, characterized in that, The physical mapping function An empirical meteorological diffusion model is adopted, the specific form of which is as follows: in, The emission source intensity is estimated from historical data, either through emission inventory inversion or based on the moving average method from historical monitoring data. , , These are empirical coefficients, determined by fitting historical data using the least squares method, specifically... , , The independent variable is the observed pollutant concentration. Perform multiple linear regression on the dependent variable.

5. The method according to claim 4, characterized in that, The physical mapping function The empirical coefficients were determined by fitting historical data using the least squares method, specifically by performing multiple linear regression with wind speed, boundary layer height, relative humidity, and temperature as independent variables and pollutant concentration observations as dependent variables.

6. The method according to claim 1, characterized in that, In step 3, the graph structure uses monitoring stations as nodes and spatial adjacency relationships as edges. A graph attention network is used to model the spatial dependencies between nodes. The graph attention network contains at least two layers, with multiple attention heads set in each layer. The association weights between nodes are adaptively learned through the attention mechanism.

7. The method according to claim 1, characterized in that, In step 4, the temporal convolutional network contains multiple residual blocks, each containing an expanded convolutional layer. The expanded convolutional structure expands the receptive field to capture the long-term temporal dependence of pollutant concentration, and the expansion factor increases with the network depth.

8. The method according to claim 1, characterized in that, The air pollution prediction results at different time scales include short-term prediction, medium-term prediction, and extended-term prediction; among them, the time scale of short-term prediction is 2 hours, the time scale of medium-term prediction is 3 days, and the time scale of extended-term prediction is 7 days.

9. The method according to claim 1, characterized in that, In step 1, the preprocessing of the multi-source data includes standardization to eliminate the influence of different units on feature extraction; the standardization parameters are calculated only based on the training set and applied to the validation and test sets.

Images