Air quality prediction method based on deep learning model of latent source contribution analysis

Abstract

This invention discloses an air quality prediction method based on a deep learning model using potential source contribution analysis. Applied to the field of air quality prediction technology, it addresses the problem that existing technologies only consider the geographical distance of the research object, lacking a deeper exploration of its spatial diffusion impact. This invention utilizes potential source contribution analysis based on airflow trajectories to construct a spatial impact map to represent the spatial correlation between cities. Then, by using a graph convolution algorithm, it extracts effective spatial features from the spatial impact map in conjunction with the pollution status of each city. Furthermore, it employs a long short-term memory model to learn the temporal characteristics of past air pollutant concentration data, generating prediction results. First, the results of potential source contribution analysis are applied to a graph convolutional network, and then combined with the long short-term memory model to predict urban air pollution data. The method of this invention can accurately predict urban air pollutants on both spatiotemporal scales.

Description

Technical Field

[0001] This invention belongs to the field of air quality prediction technology, and specifically relates to an air quality prediction technology based on potential source contribution analysis. Background Technology

[0002] With continuous urbanization and the expansion of cities, urban environments are deteriorating, and problems such as excessive emissions and environmental pollution occur frequently. The increasing number of vehicles, coupled with air pollutant emissions from factories around cities and vehicle exhaust, further damages air quality. Urban air quality has become one of the environmental problems facing my country, seriously affecting the health of the Chinese people. The main pollutants causing air pollution include carbon monoxide, ozone, carbon dioxide, some nitrogen oxides, and particulate matter, among which particulate matter poses the greatest threat to human health. Fine particulate matter has a small particle size and can act as a carrier of toxic substances; people exposed to fine particulate matter pollution for a long time are more likely to face serious risks of cardiovascular and respiratory diseases. Air pollution forecasting is of great significance for environmental protection and air pollution control. It can effectively predict air quality, reflect future trends in the atmospheric environment, and prevent air pollution events in advance. This is of great importance for strengthening atmospheric environmental protection, improving the current state of air pollution prevention and control, and promoting ecological restoration.

[0003] Currently, most studies on spatiotemporal prediction of air quality only consider the geographical distance of the research object in terms of space, lacking a deeper exploration of its spatial diffusion impact. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes an air quality prediction method based on a deep learning model using potential contribution analysis. First, the results of potential source contribution analysis are applied to a graph convolutional network, and then combined with a long short-term memory model to predict urban air pollution data, thereby enabling accurate prediction of urban air pollutants on a spatiotemporal scale.

[0005] The technical solution adopted in this invention is: an air quality prediction method based on a deep learning model of potential contribution analysis, comprising:

[0006] S1. Based on the potential source contribution analysis method, calculate the air pollution contribution factors of each surrounding city to the city to be predicted.

[0007] S2. Based on the contribution factors, select cities that have an impact on the city to be predicted and include them in the analysis area. Collect pollution data in the analysis area and construct a spatial impact map. The analysis area includes the city to be predicted and the selected cities that have an impact on the city to be predicted.

[0008] S3. Input the spatial impact map and pollution data into the graph convolutional network to transform the raw pollution data into time series data with spatial characteristics;

[0009] S4. Input the time series data with spatial characteristics into the long short-term memory model for training, extract the time features of the time series data, and predict the future air pollution situation.

[0010] The beneficial effects of this invention are as follows: This invention applies graph convolutional networks and long short-term memory networks to predict urban air pollution data. It utilizes potential source contribution analysis based on airflow trajectories to deeply explore the impact of air pollution from surrounding cities on a specific city, constructing a spatial impact map to represent the spatial correlation between cities. Then, by using a graph convolutional algorithm to combine the spatial impact map with the pollution status of each city, effective spatial features are extracted. Finally, a long short-term memory model is used to learn the temporal characteristics of past air pollutant concentration data to generate prediction results. This invention takes a dual spatiotemporal approach and proposes a new method for building spatial graph models to explore deeper environmental correlations between cities, which is beneficial for improving the accuracy and efficiency of air pollution prediction. Attached Figure Description

[0011] Figure 1 This invention relates to an air quality prediction method based on a deep learning model using potential source contribution analysis.

[0012] Figure 2 This is a schematic diagram of the graph model establishment method based on potential source contribution analysis of the present invention;

[0013] Figure 3 It is the airflow trajectory of city d within 24 hours;

[0014] Figure 4 This is a schematic diagram of the grid division;

[0015] Figure 5 This is a PSCF distribution map;

[0016] Figure 6 This is a schematic diagram of the graph convolutional network model of the present invention;

[0017] Figure 7 This is a schematic diagram of the long short-term memory model of the present invention. Detailed Implementation

[0018] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0019] Potential source contribution analysis (LPCA) is widely used to identify potential source areas of high-concentration pollutants observed at receptor sites. Therefore, by using LCA, it is possible to analyze the pollution impact of surrounding areas on a specific region more deeply, obtaining a more accurate spatial distribution map from the perspective of pollution conditions. Using the results of LCA, spatial distribution map models can be constructed, taking both temporal and spatial perspectives to effectively predict air pollution and improve the accuracy of air quality forecasts.

[0020] like Figure 1 As shown, an air quality prediction method based on a deep learning model of potential source contribution analysis includes the following steps:

[0021] S1. Collect pollution and meteorological data of the city to be predicted, and calculate the air pollution contribution factors of each surrounding city to the city to be predicted using the potential source contribution analysis method. In this embodiment, the surrounding cities mainly consider the first-ring cities adjacent to the city to be analyzed, as well as the second-ring cities adjacent to the first ring. Of course, in practical applications, a larger urban area can be considered as needed.

[0022] Based on the results of potential source contribution analysis, the impact factors of pollution from surrounding cities are calculated. The steps of potential source contribution analysis are as follows: Figure 2 As shown, the specific steps include the following:

[0023] S11. Taking city d as an example, firstly, collect meteorological data and its map vector data for multiple time periods. Then, use the meteorological data to perform backward trajectory analysis on city d to be predicted, and obtain the following results: Figure 3 The airflow trajectory of city d over 24 hours is shown; the study area is divided into several small horizontal grids, such as... Figure 4 The grid resolution is mostly set to 0.25°×0.25°. The number of trajectory points passing through grid (i,j) within the region is counted and denoted as n. ij ;

[0024] S12. Set the pollutant concentration threshold, which can be set using the average pollutant concentration or according to the secondary standard limit in the "Ambient Air Quality Standard" (GB3095-2012). Using the original pollutant concentration values ​​and the airflow trajectory obtained in S11, calculate the pollutant concentration values ​​of some points on each trajectory through reverse trajectory calculation. Count the number of pollution trajectory points in the study area that pass through grid (i,j) and whose pollutant concentration exceeds the threshold, and record this as m. ij The PSCF value of any grid (i,j) can be obtained using the following formula:

[0025]

[0026] PSCF represents the potential source contribution factor, which is finally calculated as follows: Figure 5 The results are shown.

[0027] S13. Calculate the total PSCF value of the surrounding urban areas of the city to be predicted, divide it by the corresponding city area to obtain the PSCF value per unit area of ​​each surrounding city. This gives the PSCF-based spatial influence factor σ of a certain city c for the area to be predicted. c The formula is as follows:

[0028]

[0029] Among them, S c This indicates the area of ​​the city to be predicted.

[0030] S2. Based on the contribution factors, select urban areas with spatial impact factors greater than 0 as surrounding cities that have an impact on the city to be predicted, collect their pollution data, and construct a spatial impact map.

[0031] S21. Establish a graph model adjacency matrix A with city d as the point to be predicted. factor :

[0032]

[0033] Thus, the spatial influence diagram G can be obtained. factor =(V,E,A) factor ), where V represents the number of nodes in the spatial influence graph. v i Let i represent city; E represent the edge set; A represent city i; E represent the edge set; factor It is the adjacency matrix of the graph.

[0034] S3. Input the spatial impact map and the PM2.5 monitoring data of all relevant cities into the graph convolutional network to transform the raw data into time series data with spatial characteristics;

[0035] Graph convolutional networks are used to extract the spatial characteristics of air pollution in various cities within the study area. Using a spatial impact map constructed with S1, graph convolution calculations are performed on the pollution concentration data of each city to extract the spatial characteristics of air pollution between cities. The graph convolutional network model flow is as follows: Figure 6 As shown, v0-v9 represent nodes in the graph, referring to different cities. v0 represents the city to be predicted. After graph convolution, the spatial features of the city to be predicted are extracted, which includes the following steps:

[0036] S31. Collect air pollution data for all cities within the study area, and standardize the data using deviation standardization, as shown in the following formula:

[0037]

[0038] Where, x max x represents the maximum value of PM 2.5 monitoring data. min The minimum value of PM 2.5 monitoring data is given, x* represents the target value after deviation standardization, and x represents the data that needs to be standardized.

[0039] S32. Perform symmetric normalization on the spatial influence map obtained in S1, that is...

[0040]

[0041] in, I is an identity matrix of size N×N. It is a diagonal matrix.

[0042] S33. Define the eigenvalue matrix This is used to store all standardized feature data of N cities at time t, where N is the number of cities and t is the monitoring time.

[0043] S34. Perform graph convolution on the symmetric normalized result. The calculation formula is as follows:

[0044]

[0045] in, W is the output of the (l+1)th layer graph convolution. (l) For the parameters of the l-th layer, The output of the l-th layer graph convolution is the input layer. σ(·) represents the activation function.

[0046] S4. Input the time-series data with spatial characteristics into the Long Short-Term Memory model for training, extract the temporal features of the time-series data, and predict future air pollution. Specifically, the spatially characteristic result of the feature matrix processed by the convolutional network in step S2 for any time t is set as G. t By combining the convolution results from the past T time steps, and since a spatial influence map was previously constructed based on the 24-hour airflow trajectory, T is set to 1, thus obtaining a set of time series sequences with spatial characteristics. The time series sequence is input into the Long Short-Term Memory model, and combined with the temporal features of the extracted time series data, the predicted air pollution values ​​for the next T′ time steps are obtained. Air pollution exhibits a certain periodicity, with a cycle of one week. T′ can be selected from an integer between 1 and 7, depending on the forecast requirements. Figure 7 As can be seen, each recurrent structure in the Long Short-Term Memory (LSTM) model consists of a unit state, three gate structures, and four neural network layers. The key part of the LTM model is the unit state, i.e. Figure 7 The horizontal arrow at the top of the middle outputs c t The cell state proceeds downwards through the entire chain, and information can easily flow to the next node along the direction of the arrows.

[0047] Long Short-Term Memory (LSTM) models can control information through gate structures, allowing for the complete deletion or addition of information from a cell state. Gates can also selectively allow some information to pass through, achieved through a Sigmoid neural network layer and dot product operations. The Sigmoid layer outputs a number between 0 and 1, describing how much information can pass through each component; 0 indicates no information can pass through at all, and 1 indicates all information can pass through. The LTM model contains three such gate structures: an input gate, an output gate, and a forget gate. The input gate determines how much input information can be stored in the cell state c. t In the middle, the output gate control unit status How many of the current output values ​​h can be output to the Long Short-Term Memory model? t In this context, the forget gate determines how much of the cell state at time t-1 is retained in the cell state at time t. Through the interaction between the three gates, the cell state c at time t is updated. t It contains the information transmitted at time t-1 that needs to be discarded and the new information acquired from the input signal at time t that needs to be added. c t-1 It will continue to be passed on to the long short-term memory model at time t+1, and will be passed on as a new cell state.

[0048] Future data and historical data are highly correlated. By extracting time features using a long short-term memory model, it is possible to predict the concentration of air pollutants in the next T′ time steps.

[0049] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. An air quality prediction method based on a deep learning model using potential contribution analysis, characterized in that, include: S1. Based on the potential source contribution analysis method, calculate the air pollution contribution factors of each surrounding city to the city to be predicted. S2. Based on the contribution factors, select cities that have an impact on the city to be predicted and include them in the analysis area. Collect pollution data in the analysis area and construct a spatial impact map. The analysis area includes the city to be predicted and the selected cities that have an impact on the city to be predicted. Constructing a spatial influence map specifically involves establishing a graph model adjacency matrix with the cities to be predicted as prediction points. To obtain the spatial influence map ,in This indicates the number of nodes in the spatial influence graph. , Represents city N is the number of cities; Represents the set of edges; Represented as: ； in, Indicates the city to be predicted; S3. Input the spatial impact map and pollution data into a graph convolutional network to transform the raw pollution data into time-series data with spatial characteristics; step S3 specifically includes the following sub-steps: S31. Collect air pollution data for all cities within the study area, and standardize the data using deviation standardization, as shown in the following formula: ； in, This represents the maximum value of the air pollution data. This represents the minimum value of air pollution data. This represents the target value after standardization of the deviation. This represents data that needs to be standardized. S32. Perform symmetric normalization on the spatial influence map obtained in S1, that is... ； in, , It is the size of The identity matrix, It is a diagonal matrix. ; S33. Define the eigenvalue matrix This is used to store all standardized feature data of N cities at time t, where t is the monitoring time. S34. Perform graph convolution on the symmetric normalized result. The calculation formula is as follows: ； in, This is the output of the (l+1)th layer graph convolution. For the parameters of the l-th layer, The output of the l-th layer graph convolution is the input layer. , Indicates the activation function; S4. Input the time series data with spatial characteristics into the long short-term memory model for training, extract the time features of the time series data, and predict the future air pollution situation.

2. The air quality prediction method based on a deep learning model with potential contribution analysis according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S11. Take the city to be predicted and its surrounding cities as a study area, collect meteorological data and map vector data of the study area for multiple time periods, and use the meteorological data to perform backward trajectory analysis on the city to be predicted to obtain the airflow trajectory of the city to be predicted. The study area was divided into several small horizontal grids, with the grid resolution mostly set to 0.25° × 0.25°. Statistics were collected on the data passing through the grids within the study area. The number of trajectory points is denoted as ; S12. Set pollutant concentration thresholds; using the original pollutant concentration values ​​and combining them with the airflow trajectory obtained in S11, calculate the pollutant concentration values ​​at each trajectory point through reverse trajectory calculation, and statistically analyze the data within the study area after passing through the grid. The number of pollution trajectory points whose pollutant concentration exceeds the threshold is denoted as [missing information]. The mesh is calculated using the following formula. PSCF value: ； PSCF stands for Potential Source Contribution Factor; S13. Calculate the PSCF-based spatial influence factors of surrounding cities on the city to be predicted. The calculation formula is: ； in, This represents the area of ​​the surrounding city c.

3. The air quality prediction method based on a deep learning model with potential contribution analysis according to claim 2, characterized in that, The pollutant concentration thresholds are set based on the average pollutant concentration or the secondary standard limits in the "Ambient Air Quality Standard".

4. The air quality prediction method based on a deep learning model with potential contribution analysis according to claim 3, characterized in that, Step S2 includes surrounding cities with spatial influence factors greater than 0 into the area to be analyzed.

Images