Atmospheric pollutant concentration prediction method and device based on deep machine learning

Abstract

The invention discloses an atmospheric pollutant concentration prediction method and device based on deep machine learning, and the method comprises the steps: collecting the atmospheric pollutant concentration data of a to-be-detected region in a given time period, and carrying out the preprocessing of the data; training the constructed concentration prediction model by using the preprocessed atmospheric pollutant concentration data as a training set; utilizing the trained concentration prediction model to predict the concentration change of at least one given target pollutant in a to-be-detected area at different time nodes, and screening high-value point locations in real time according to predicted values and performing traceability analysis. According to the method, the high-value point pollutant concentration management and control refinement level can be greatly improved, and the real-time prediction precision of pollutants and the reliability of a traceability result are improved.

Description

Technical Field

[0001] The present invention relates to the field of air pollution monitoring, and in particular to a method and device for predicting air pollutant concentrations based on deep machine learning. Background Art

[0002] The impact of pollution source changes on particulate matter pollution is a complex environmental issue involving multiple factors, including changes in the types, quantities, emissions, and environmental conditions of pollution sources. When pollution sources shift from activities that produce less particulate matter to activities that produce more particulate matter, the increase in the number of pollution sources or the increase in emissions will increase the level of particulate matter pollution, directly affecting the concentration of particulate matter. With the growth of population and the intensification of economic activities, especially in urban areas, the increase in pollution sources has a particularly significant impact on particulate matter pollution. Therefore, it is particularly important to combine seasonal pollution source changes with pollution source analysis in different regions.

[0003] In the prior art, prediction methods based on numerical models and prediction methods based on data-driven models are two commonly used methods.

[0004] The prediction method based on numerical models mainly relies on meteorological principles and statistical methods to simulate the emission, diffusion and disappearance of atmospheric pollutants, thereby predicting the concentration of pollutants. However, this method has some limitations, such as the uncertainty of pollution sources, meteorological conditions and transformation processes, as well as the high complexity and large amount of calculation of numerical models, which makes this method difficult to apply to the field of regional pollutant concentration prediction and cannot accurately predict the traceability changes of pollution sources.

[0005] Prediction methods based on data-driven models, especially machine learning algorithms, predict pollutant concentrations by learning and analyzing historical pollutant data. This method further improves model performance and greatly improves the accuracy of prediction models. It is widely used in the prediction of conventional air pollutants such as air quality index (AQI), particulate matter, and nitrogen oxides. However, existing studies on environmental pollutant concentrations using data-driven methods are mainly applied to regional ambient atmospheres with relatively small concentration changes. They are not suitable for seasonal pollution source changes and pollution source analysis in different regions, resulting in poor accuracy in identifying pollution source changes. Summary of the invention

[0006] In order to solve the problems existing in the background technology, the present invention proposes a method and device for predicting atmospheric pollutant concentrations based on deep machine learning, aiming to solve the problems of insufficient recognition accuracy and poor real-time performance faced by the prior art in monitoring changes in pollution sources and pollutants in different regions. By real-time screening of high-value points and conducting source tracing analysis, the level of refined control of pollutant concentrations at high-value points can be greatly improved, the real-time prediction accuracy of pollutants is improved, and the increase in hardware investment in monitoring equipment is avoided.

[0007] On the one hand, the present invention provides a method for predicting atmospheric pollutant concentrations based on deep machine learning, comprising:

[0008] Collect atmospheric pollutant concentration data at different time points in the test area and pre-process the data;

[0009] The pre-processed atmospheric pollutant concentration data is used as a training set to train the constructed concentration prediction model;

[0010] The trained concentration prediction model is used to predict the concentration change of at least one target pollutant in the test area at different time points.

[0011] In some embodiments, predicting the concentration change of at least one target pollutant in the area to be detected includes:

[0012] Divide the area to be tested into a plurality of different sub-areas of interest;

[0013] The target pollutant concentration data collected in each sub-area of ​​concern are interpolated using the inverse distance weighted method.

[0014] The point distribution of different levels in each sub-area of ​​interest after interpolation processing is identified to obtain the predicted result of the concentration change of the target pollutant. The higher the point level, the higher the concentration of the corresponding target pollutant.

[0015] In some embodiments, identifying the distribution of points of different levels in each sub-region of interest after interpolation processing includes:

[0016] Through the global spatial autocorrelation analysis method, the target pollutant concentrations in each sub-area of ​​interest in the area to be tested are analyzed for overall clustering, and the first clustering result is obtained;

[0017] Based on the first clustering result, a high / low clustering analysis method is used to determine whether the concentration of the target pollutant in each sub-area of ​​concern is spatially clustered at a high value or a low value;

[0018] Statistically identify the high-value clusters or low-value clusters of the target pollutant concentration in each sub-area of ​​interest in space, and identify the high-value points and low-value points in each sub-area of ​​interest. The point distribution of different levels includes the high-value points above the first preset threshold and the low-value points below the second preset threshold.

[0019] In some embodiments, the global spatial autocorrelation analysis method is used to perform overall clustering of target pollutant concentrations in each sub-area of ​​interest of the area to be tested, including:

[0020] The global Moran index is calculated based on the target pollutant concentration value of each sub-area of ​​concern and the average target pollutant concentration of all sub-areas of concern. The global Moran index is expressed as:

[0021]

[0022] in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; The value range of the global Moran index I is (-1 to 1). When I>0, it indicates positive spatial correlation, which indicates that the space is clustered; when I<0, it indicates negative spatial correlation, which indicates that the space is dispersed; when I=0, it indicates that the space is randomly distributed. The closer the I value is to 1 or -1, the more intense the degree of spatial autocorrelation is, and the higher the degree of spatial clustering is.

[0023] The result of the global Moran's index is statistically tested to obtain the first clustering result, which is expressed as:

[0024]

[0025] Among them, Var(I) is the theoretical variance of I, E(I) is the theoretical expectation, and Z(I) is the first clustering result.

[0026] In some embodiments, a high / low clustering analysis method is used to determine whether the concentration of the target pollutant in each sub-area of ​​interest is spatially clustered at a high value or a low value, including:

[0027] The target pollutant concentration values ​​of each sub-area of ​​concern are counted to obtain the second clustering result, which is expressed as:

[0028]

[0029] Among them, G(d) represents the second clustering result, d is the distance, W ij(d) is the spatial weight defined by the distance rule, i and j represent the codes of different sub-areas of interest; n represents the number of all sub-areas of interest; x represents the target pollutant concentration value of each sub-area of ​​interest;

[0030] The expected value of the second clustering result is calculated under the assumption that there is no spatial aggregation:

[0031] E(G)=W ij (d) / [n(n-1)]

[0032] The statistical test value of the second clustering result under the condition of normal distribution is calculated as follows:

[0033]

[0034] Among them, when the second clustering result G(d) is higher than the expected value E(G) and the statistical test value Z is significant, it indicates that there is a high-value aggregation in space; when the second clustering result G(d) is lower than the expected value E(G) and the Z value is significant, it indicates that there is a low-value aggregation in space; when the second clustering result G(d) tends to the expected value E(G), it indicates that there is a random distribution feature in space.

[0035] In some embodiments, statistically identifying the high-value clusters or low-value clusters of the target pollutant concentration in each sub-area of ​​interest in space includes:

[0036] According to the target pollutant concentration value of each sub-area of ​​concern and the average value of the target pollutant concentration of all sub-areas of concern, the score value of each cluster position of the target pollutant is calculated. The score value is expressed as:

[0037]

[0038] in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; score value G i *The higher the value, the tighter the clustering of high values, and the score G i *The lower the value, the tighter the clustering of low values;

[0039] According to the size relationship between the score value and different preset thresholds, the point distribution of different levels is determined, wherein when the score value exceeds the first preset threshold, it is identified as a high-value point, and when the score value exceeds the second preset threshold, it is identified as a low-value point.

[0040] In some embodiments, it further comprises:

[0041] Identify the key pollution sources that cause the concentration change of at least one target pollutant at different time points, and output the prediction results of the key pollution sources, including:

[0042] Analyze the key pollution sources that cause changes in the concentration of each target pollutant in the overall tested area at different time points;

[0043] Analyze the key pollution sources that cause the concentration changes of each target pollutant in each sub-area of ​​interest of the area to be tested at different time points;

[0044] For the high-value points screened out in each sub-area of ​​concern, different high-value concentration periods are selected to analyze the key pollution sources that cause changes in the concentration of each target pollutant.

[0045] In some embodiments, preprocessing the data includes:

[0046] The collected atmospheric pollutant data are cleaned, screened and processed for outliers, including statistics of missing values, statistics of outliers, and filling in missing values ​​using the k-nearest neighbor mean interpolation method, and the data are normalized to the interval [0,1].

[0047] In some embodiments, an LSTM long short-term memory network is used to construct the concentration prediction model.

[0048] On the other hand, the present invention further provides an atmospheric pollutant concentration prediction device based on deep machine learning, which adopts the above-mentioned atmospheric pollutant concentration prediction method based on deep machine learning, and the atmospheric pollutant concentration prediction device based on deep machine learning comprises:

[0049] The data set construction module is used to collect the atmospheric pollutant concentration data at different time points in the test area and pre-process the data;

[0050] A model training module is used to train the constructed concentration prediction model using the pre-processed atmospheric pollutant concentration data as a training set;

[0051] The prediction module is used to use the trained concentration prediction model to predict the concentration change of at least one target pollutant in the test area at different time nodes.

[0052] In addition, another aspect of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above-mentioned atmospheric pollutant concentration prediction method based on deep machine learning are implemented and the same technical effect can be achieved.

[0053] The present invention has outstanding substantive features and significant progress compared to the prior art, specifically:

[0054] The present invention proposes a method for predicting atmospheric pollutant concentrations based on deep machine learning. The method collects atmospheric pollutant concentration data at different time nodes in the area to be tested and preprocesses the data; and uses the preprocessed atmospheric pollutant concentration data as a training set to train the constructed concentration prediction model. Using the trained concentration prediction model, the concentration change of at least one given target pollutant in the area to be tested at different time nodes is predicted. The method is based on deep machine learning, and through the deep learning method and the introduction of computing mechanisms such as spatial autocorrelation analysis, high / low clustering analysis, and hot spot analysis, the model performance is further improved, so that the accuracy of the prediction model is greatly improved. At the same time, the method takes into account the changes in pollution sources in different seasons and the analysis of pollution sources in different regions, and by real-time screening of high-value points and conducting source tracing analysis, it can greatly improve the refinement level of pollutant concentration control at high-value points, improve the accuracy of pollutant concentration prediction and pollution source change identification, and avoid increasing hardware investment such as monitoring equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The overall flow chart of the atmospheric pollutant concentration prediction method based on deep machine learning according to an embodiment of the present invention is shown;

[0056] Figure 2 A schematic diagram showing the flow of the target pollutant concentration change prediction process;

[0057] Figure 3 The predicted results of the monthly average concentration change of the target pollutant PM2.5 in the overall tested area are shown.

[0058] Figure 4 A schematic diagram of the process of tracing the source of target pollutants is shown;

[0059] Figure 5(a) and Figure 5(b) show the changes in key pollution sources of the same target pollutant in different sub-areas of concern at different time points;

[0060] Figure 6(a) shows the concentration evolution sequence of the target pollutant during the high concentration period;

[0061] Figure 6(b) shows the contribution ratio of each pollution source corresponding to the target pollutant during the high concentration period;

[0062] Figure 7 A schematic block diagram showing the overall structure of an atmospheric pollutant concentration prediction device based on deep machine learning according to an embodiment of the present invention is shown;

[0063] Wherein, the accompanying drawings are marked as follows:

[0064] 300-Atmospheric pollutant concentration prediction device based on deep machine learning;

[0065] 301-dataset construction module;

[0066] 302-model training module;

[0067] 303-Prediction module. DETAILED DESCRIPTION

[0068] In order to make the above features and effects of the present invention more clearly understood, embodiments are given below and described in detail with reference to the accompanying drawings.

[0069] See also Figure 1 As shown, Figure 1 A schematic diagram of the overall process of an atmospheric pollutant concentration prediction method based on deep machine learning disclosed in one embodiment of the present invention is shown.

[0070] A method for predicting atmospheric pollutant concentration based on deep machine learning includes at least the following steps:

[0071] Step S1: Collect the atmospheric pollutant concentration data at different time points in the test area and pre-process the data.

[0072] In some embodiments, hourly atmospheric pollutant concentration data for consecutive years are captured from a network monitoring platform as a sample data set, including atmospheric pollutants such as AQI, PM2.5, PM10, O3, NO2, SO2, and CO. In this embodiment, the real-time nature of the sample data is ensured by capturing data hourly.

[0073] In addition, in one embodiment, due to the presence of missing data and outliers in the sample data set, in one embodiment, the data is further pre-processed such as cleaning, screening and outliers. Missing values ​​and outliers are counted and filled with knn nearest neighbor mean interpolation. Usually, in the process of evaluating and comparing indicators, we often choose to discard the units of the original analysis data and transform them into dimensionless values ​​in order to facilitate weighting or comparison of indicators of different units and magnitudes. The process of scaling the original analysis data to a smaller range is called data normalization. In this embodiment, the data is normalized to the interval [0,1].

[0074] Step S2: Use the pre-processed atmospheric pollutant concentration data as a training set to train the constructed concentration prediction model.

[0075] In machine learning, in order to effectively train a model and evaluate its performance, it is usually necessary to divide the training data set into a training set and a validation set. First, the model is trained using the training set, and the model's hyperparameters and optimization algorithms are continuously optimized to improve its prediction accuracy and stability. This process aims to make the model better fit the data and have stronger generalization capabilities. Subsequently, the trained model is evaluated using the validation set. This includes calculating error metrics between the model's predictions and the actual values, such as root mean square error, mean absolute error, etc. Through this step, we can objectively evaluate the performance of the model on unseen data and further adjust the model to ensure its reliable performance in practical applications. In general, dividing the data set, training the model, adjusting parameters, and evaluating performance is an iterative process aimed at making the model more accurate and stable to adapt to a variety of practical scenarios. This systematic approach helps to build machine learning models with good generalization capabilities for various data.

[0076] In this embodiment, after constructing the sample data set, it is further divided into a training set and a validation set, and the constructed concentration prediction model is trained using the training set.

[0077] In one embodiment, the concentration prediction model is constructed using an LSTM long short-term memory network. Unlike a general feedforward neural network, the LSTM long short-term memory network can use time series to analyze inputs and use LSTM to achieve hourly prediction and source tracing of atmospheric pollutant concentrations in the test area.

[0078] Step S3: using the trained concentration prediction model, predict the concentration change of at least one target pollutant in the test area at different time nodes. The concentration prediction model outputs the predicted results of the concentration change of each target pollutant and the predicted results of the key pollution sources that cause the target pollutant to exceed the standard at different time nodes.

[0079] Specifically, the prediction process includes two parts. The first is the prediction process of the concentration change of at least one target pollutant in the test area to obtain the predicted results of the concentration change of each target pollutant; the second is the prediction process for the concentration change, further tracing the key pollution sources that cause the target pollutants to exceed the standard at different time nodes, and obtaining the prediction results of the key pollution sources.

[0080] In one embodiment, the prediction process of the concentration change of at least one target pollutant in the area to be tested is as follows: Figure 2 As shown in Figure 2 A flow chart showing the concentration change prediction process of the target pollutant is shown.

[0081] First, the area to be tested is divided into a plurality of different sub-areas of interest (step S31). For example, the area to be tested is the entire Beijing city, and each district is a divided sub-area of ​​interest.

[0082] Then, the target pollutant concentration data collected in each sub-area of ​​interest are interpolated by the inverse distance weight method (step S32). In the specific implementation, the inverse distance weight method is a commonly used spatial interpolation method, particularly suitable for the interpolation processing of pollutant concentration data. The method is based on an assumption that points close in space have greater similarity than points far apart. Therefore, points closer in distance have a larger weight in the interpolation result, while points farther away have a smaller weight. When the target pollutant concentration data collected in each sub-area of ​​interest are interpolated using the inverse distance weight method, the main steps are as follows: collect the target pollutant concentration data collected by each monitoring point in each sub-area of ​​interest, such as PM2.5, PM10, etc., and the geographic coordinate information of these monitoring points. In IDW interpolation, it is usually necessary to set one or more distance power parameters (usually denoted as p or n), which will affect the degree to which the weight changes with distance. The higher the power, the smaller the impact of points farther away on the interpolation result. Commonly used power values ​​include 2 (which means that the weight is inversely proportional to the square of the distance). Then, for the interpolation point (i.e., the point without direct monitoring data), it is necessary to calculate the weight of each monitoring point in the interpolation result based on its distance from each monitoring point and the selected power parameter. The weight calculation formula is usually: i =1 / d i p , where w i is the weight of the i-th monitoring point, d i is the distance from the interpolation point to the i-th monitoring point, and p is the distance power parameter. According to the weights and concentration data of each monitoring point, the concentration value of the interpolation point is calculated by weighted average: Among them, c i is the concentration value of the ith monitoring point, and n is the number of monitoring points involved in the interpolation. Through the above interpolation process, the gaps in pollutant concentrations between different monitoring points can be filled, and the amount of pollutant concentration data can be expanded to subsequently improve the accuracy of prediction.

[0083] Finally, the point distribution of different levels in each sub-area of ​​interest after interpolation processing is identified to obtain the predicted result of the concentration change of the target pollutant. The higher the point level, the higher the concentration of the corresponding target pollutant (step S33). In some embodiments, the point distribution of different levels in each sub-area of ​​interest after interpolation processing is identified, for example, different point levels are divided according to different preset thresholds, and each point level corresponds to a different pollutant concentration range. The different point levels at least include the high-value point higher than the first preset threshold, the low-value point lower than the second preset threshold, and the intermediate value point between the first preset threshold and the second preset threshold. In addition, the high-value point, the low-value point, and the intermediate value point can be further divided as needed, and the present invention is not limited to this. This process combines the global spatial autocorrelation analysis method, the high / low cluster analysis (Getis-OrdGeneral G), the hot spot analysis and other methods to identify the point distribution of different levels in each sub-area of ​​interest after interpolation processing, especially the high-value site.

[0084] Specifically, first, the global spatial autocorrelation analysis method is used to perform an overall clustering analysis on the target pollutant concentrations of each sub-area of ​​interest in the area to be tested, and the first clustering result is obtained. Spatial autocorrelation can measure whether things have autocorrelation in space. High autocorrelation represents the existence of clustering of things in space. In this analysis process, specifically, the global Moran index is calculated based on the target pollutant concentration values ​​of each sub-area of ​​interest and the average value of the target pollutant concentrations of all sub-areas of interest. The global Moran index is expressed as:

[0085]

[0086] in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; The value range of the global Moran's index I is (-1~1). When I>0, it indicates positive spatial correlation, which indicates spatial agglomeration; when I<0, it indicates negative spatial correlation, which indicates spatial dispersion; when I=0, it indicates random distribution in space. The closer the I value is to 1 or -1, the more intense the degree of spatial autocorrelation is and the higher the degree of spatial agglomeration is.

[0087] In addition, the result of the global Moran index is further statistically tested to obtain the first clustering result, which is expressed as:

[0088]

[0089] Among them, Var(I) is the theoretical variance of I, E(I) is the theoretical expectation, and Z(I) is the first clustering result.

[0090] Then, since the global Moran index is used to measure the clustering of things in space, it cannot determine whether things are clustered at high values ​​or low values ​​in space. Therefore, in one embodiment, based on the first clustering result, a high / low clustering analysis method is used to determine whether the concentration of target pollutants in each sub-area of ​​interest is clustered at high values ​​or low values ​​in space. In this analysis process, the concentration values ​​of target pollutants in each sub-area of ​​interest are counted to obtain the second clustering result, which is expressed as:

[0091]

[0092] Among them, G(d) represents the second clustering result, d is the distance, W ij (d) is the spatial weight defined by the distance rule, i and j represent the codes of different sub-areas of interest; n represents the number of all sub-areas of interest; and x represents the target pollutant concentration value of each sub-area of ​​interest.

[0093] The expected value of the second clustering result is calculated under the assumption that there is no spatial aggregation:

[0094] E(G)=W ij (d) / [n(n-1)]

[0095] The statistical test value of the second clustering result under the condition of normal distribution is calculated as follows:

[0096]

[0097] Among them, when the second clustering result G(d) is higher than the expected value E(G) and the statistical test value Z is significant, it indicates that there is a high-value aggregation in space; when the second clustering result G(d) is lower than the expected value E(G) and the Z value is significant, it indicates that there is a low-value aggregation in space; when the second clustering result G(d) tends to the expected value E(G), it indicates that there is a random distribution feature in space.

[0098] Finally, by using hotspot analysis to obtain the locations where high-value or low-value elements are clustered in space, we can intuitively understand where high and low values ​​are clustered and how densely they are clustered. Specifically, in this process, by statistically identifying the locations of high-value clusters or low-value clusters of target pollutant concentrations in each sub-area of ​​interest, we can identify the high-value points and low-value points of each sub-area of ​​interest. In this process, based on the target pollutant concentration value of each sub-area of ​​interest and the average value of the target pollutant concentration of all sub-areas of interest, we calculate the score value of each cluster position of the target pollutant. The score value is expressed as:

[0099]

[0100] in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; score value G i *The higher the value, the tighter the clustering of high values, and the score G i *The lower the value, the tighter the clustering of low values. Based on the relationship between the score value and different preset thresholds, the distribution of points of different levels is determined, where when the score value exceeds the first preset threshold, it is identified as a high-value point, and when the score value exceeds the second preset threshold, it is identified as a low-value point.

[0101] In this embodiment, in the process of predicting the concentration change of at least one target pollutant in the test area at different time nodes using the trained concentration prediction model, the prediction time nodes can be set, such as monthly prediction, daily prediction, annual prediction, etc. In addition, different sub-areas of interest can be set with different time nodes as needed, and the present invention is not limited to this. For example, Figure 3 It shows that the area to be tested is Beijing, the sub-areas of concern are the districts and counties of Beijing, and the target pollutant is PM 2.5 The prediction results of monthly average concentration change.

[0102] In one embodiment, the prediction process of the concentration prediction model further includes a prediction process for concentration changes, further tracing the key pollution sources that cause the target pollutants to exceed the standard at different time nodes, identifying the key pollution sources that cause the concentration changes of at least one target pollutant at different time nodes, and obtaining the prediction results of the key pollution sources, especially the tracing analysis of high-value points. Specifically, it includes the tracing analysis of the overall area to be tested, the tracing analysis of each sub-area of ​​concern, and the further tracing analysis of the high-value points selected in each sub-area of ​​concern. Figure 4 As shown, Figure 4 A flow chart of the traceability process is shown.

[0103] In this process, the key pollution sources that cause changes in the concentration of each target pollutant at different time points in the overall area to be tested are first analyzed (step S41). Through this process, the key pollution sources that cause changes in the concentration of target pollutants within the overall scope of the area to be tested can be identified. By determining the contribution value of each pollution source to the change in the concentration of the target pollutant, the key pollution sources can be identified. For example, in practice, the pollution sources that affect the target pollutant PM2.5 within the scope of the Beijing area to be tested in 2022 are obtained from high to low: road dust source (29.8%)> industrial source (23.7%)> construction site dust source (18.5%)> residential source (10.9%)> motor vehicle source (6.8%)> agricultural source (4.7%)> other sources (2.9%)> power source (2.7%). At this time, the top few pollution sources can be determined as key pollution sources as needed.

[0104] Due to geographical differences, there are also certain differences in the key pollution sources of the same target pollutant in different sub-areas of concern. Therefore, in one embodiment, the key pollution sources that cause changes in the concentration of each target pollutant at different time nodes in each sub-area of ​​concern in the area to be tested are further analyzed (step S42). Through this process, the key pollution sources that cause changes in the concentration of target pollutants at different time nodes in each sub-area of ​​concern in the area to be tested can be identified, and the key pollution sources can be identified by determining the contribution value of each pollution source in each sub-area of ​​concern to the change in the concentration of the target pollutant. For example, in practice, the Beijing area to be tested was obtained. In January 2022, the key pollution sources of the target pollutant PM2.5 in the three sub-areas of concern, Xicheng District, Pinggu District, and Fangshan District, were industrial sources (22.2%~22.5%) and residential sources (24.8%~30.8); the key pollution sources of the target pollutant PM2.5 in the sub-areas of concern, such as Changping District, Shunyi District, Huairou District, Yanqing District, Miyun District, Mentougou District, Tongzhou District, and Daxing District, are residential sources (21.8%~29. 4%) and road dust sources (23.3%-31.4%); the key pollution sources of target pollutants PM2.5 in sub-focus areas such as Dongcheng District, Chaoyang District, and Shijingshan District are road dust sources (28.6%-34.1%) and construction site dust sources (20.2%-29.7%); the key pollution sources of target pollutants PM2.5 in the two sub-focus areas of Haidian District and Fengtai District are road dust sources, with contributions accounting for 37.4% and 32.2% respectively, as shown in Figure 5(a). In April 2022, the key pollution sources of target pollutants PM2.5 in sub-areas of concern such as Xicheng District, Dongcheng District, Chaoyang District, and Shijingshan District are road dust sources (24.6%-33.5%) and construction site dust sources (28.4%-41.5%); the key pollution sources of target pollutants PM2.5 in sub-areas of concern such as Shunyi District, Huairou District, Yanqing District, Pinggu District, Miyun District, Mentougou District, Tongzhou District, and Daxing District are industrial sources (23.9%-34.9%) and road dust sources (23.2%-37.1%). The key pollution sources of PM2.5, the target pollutant in Haidian District and Fengtai District, are industrial sources (20.6%, 22.5%), road dust sources (32.8%, 32.6%), and construction site dust sources (25.0%, 26.2%). The key pollution source of PM2.5, the target pollutant in Changping District, is road dust sources, which contribute 41.4%; the key pollution source of PM2.5, the target pollutant in Fangshan District, is industrial sources, which contribute 34.2%, as shown in Figure 5(b).From this, it can be seen that the key pollution sources of the same target pollutant in the same sub-area of ​​concern at different time points will have certain differences, and the key pollution sources of the same target pollutant in different sub-areas of concern at the same time point will have certain differences. This fully reflects that the method of the present invention takes into account the changes in pollution sources in different seasons and the analysis of pollution sources in different regions, and improves the accuracy of pollutant concentration prediction and identification of pollution source changes.

[0105] In addition, in some embodiments, the high-value points selected from each sub-area of ​​interest identified in the above-mentioned pollutant concentration prediction process are further predicted to predict the key pollution sources corresponding to the high-value points. The key pollution sources that cause the change in concentration of each target pollutant are analyzed by selecting different high-value concentration periods (step S43). In this process, the key pollution sources that cause the change in the concentration of the target pollutant in the high-value points corresponding to different time nodes in each sub-area of ​​interest can be identified. By determining the contribution value of each pollution source to the change in the concentration of the target pollutant, the key pollution sources in the high-value points corresponding to different time nodes can be identified. For example, in practice, as shown in Figures 6 (a) and 6 (b), Figure 6 (a) shows the concentration evolution sequence of the target pollutant PM10 during the high-value concentration period from January 28 to January 31, and Figure 6 (b) shows the contribution ratio of each pollution source corresponding to the target pollutant PM10 during the high-value concentration period from January 28 to January 31. As can be seen from Figure 6 (a) and Figure 6 (b), from January 28 to January 31, the daily average concentration of the target pollutant PM10 ranged from 19.77μg / m3 to 105.19μg / m3. From January 28 to January 30, during the period of rising PM10 mass concentration, the contribution of industrial sources, residential sources and motor vehicle sources increased significantly, and the contribution of road dust sources decreased, with the proportions ranging from 8.0% to 18.54%, 12.2% to 17.1%, 2.4% to 5.1%, and 38.18% to 54.1%, respectively. On January 31, the PM10 concentration dropped to 19.77μg / m3, the contribution of industrial sources, residential sources and motor vehicle sources decreased, and the contribution of road dust sources rebounded. It can be seen that the method of the present invention can predict the concentration changes of target pollutants day by day, and realize the tracing of the corresponding key pollution sources, especially it can greatly improve the level of refined control of high-value point pollutant concentrations.

[0106] In summary, the atmospheric pollutant concentration prediction method based on deep machine learning disclosed in the present invention collects atmospheric pollutant concentration data at different time nodes in the area to be tested and preprocesses the data; and uses the preprocessed atmospheric pollutant concentration data as a training set to train the constructed concentration prediction model. Using the trained concentration prediction model, the concentration change of at least one given target pollutant in the area to be tested at different time nodes is predicted. This method is based on deep machine learning, and through the deep learning method and the introduction of computing mechanisms such as spatial autocorrelation analysis, high / low clustering analysis, and hot spot analysis, the model performance is further improved, which greatly improves the accuracy of the prediction model. At the same time, this method takes into account the changes in pollution sources in different seasons and the analysis of pollution sources in different regions, and by real-time screening of high-value points and conducting source tracing analysis, it can greatly improve the refinement level of pollutant concentration control at high-value points, improve the accuracy of pollutant concentration prediction and pollution source change identification, and avoid increasing hardware investment such as monitoring equipment.

[0107] In addition, corresponding to the above-mentioned atmospheric pollutant concentration prediction method based on deep machine learning, another embodiment of the present invention also discloses an atmospheric pollutant concentration prediction device based on deep machine learning, referring to Figure 7 , Figure 7 The architecture diagram of the atmospheric pollutant concentration prediction device 300 based on deep machine learning is shown, which can be realized by Figure 1 , Figure 2 , Figure 4 The various processes of the atmospheric pollutant concentration prediction method based on deep machine learning are shown.

[0108] An atmospheric pollutant concentration prediction device 300 based on deep machine learning, the device at least comprising:

[0109] The data set construction module 301 is used to collect the atmospheric pollutant concentration data at different time points in the test area and pre-process the data;

[0110] A model training module 302 is used to train the constructed concentration prediction model using the pre-processed atmospheric pollutant concentration data as a training set;

[0111] The prediction module 303 is used to predict the concentration change of at least one target pollutant in the test area at different time points by using the trained concentration prediction model.

[0112] In addition, corresponding to the above-mentioned atmospheric pollutant concentration prediction method based on deep machine learning, another embodiment of the present invention also discloses an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of the above-mentioned atmospheric pollutant concentration prediction method based on deep machine learning are implemented, and the same technical effect can be achieved.

[0113] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0114] Another embodiment of the present invention further provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the above-mentioned steps of wide-band oscillation monitoring are implemented and the same technical effect can be achieved.

[0115] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be applied, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0116] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0117] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for predicting atmospheric pollutant concentration based on deep machine learning, characterized in that: Include: Collect the atmospheric pollutant concentration data within a given time period in the test area and pre-process the data; The pre-processed atmospheric pollutant concentration data is used as a training set to train the constructed concentration prediction model; The trained concentration prediction model is used to predict the concentration change of at least one target pollutant in the test area at different time points.

2. The method according to claim 1, characterized in that Predicting the concentration change of at least one target pollutant in the area to be tested, comprising: Divide the area to be tested into a plurality of different sub-areas of interest; The target pollutant concentration data collected in each sub-area of ​​interest are interpolated using the inverse distance weighted method. The point distribution of different levels in each sub-area of ​​interest after interpolation processing is identified to obtain the predicted result of the concentration change of the target pollutant. The higher the point level, the higher the concentration of the corresponding target pollutant.

3. The method according to claim 2, characterized in that Identify the distribution of points of different levels in each sub-area of ​​interest after interpolation, including: Through the global spatial autocorrelation analysis method, the target pollutant concentrations in each sub-area of ​​interest in the area to be tested are analyzed for overall clustering, and the first clustering result is obtained; Based on the first clustering result, a high / low clustering analysis method is used to determine whether the concentration of the target pollutant in each sub-area of ​​concern is spatially clustered at a high value or a low value; Statistically identify the high-value clusters or low-value clusters of the target pollutant concentration in each sub-area of ​​interest in space, and identify the high-value points and low-value points in each sub-area of ​​interest. The point distribution of different levels includes the high-value points above the first preset threshold and the low-value points below the second preset threshold.

4. The method according to claim 3, characterized in that Through the global spatial autocorrelation analysis method, the target pollutant concentrations in each sub-area of ​​interest in the tested area are clustered as a whole, including: The global Moran index is calculated based on the target pollutant concentration value of each sub-area of ​​concern and the average target pollutant concentration of all sub-areas of concern. The global Moran index is expressed as: in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; The value range of the global Moran index I is (-1 to 1). When I>0, it indicates positive spatial correlation, which indicates that the space is clustered; when I<0, it indicates negative spatial correlation, which indicates that the space is dispersed; when I=0, it indicates that the space is randomly distributed. The closer the I value is to 1 or -1, the more intense the degree of spatial autocorrelation is, and the higher the degree of spatial clustering is. The result of the global Moran's index is statistically tested to obtain the first clustering result, which is expressed as: Among them, Var(I) is the theoretical variance of I, E(I) is the theoretical expectation, and Z(I) is the first clustering result.

5. The method according to claim 4, characterized in that The high / low clustering analysis method is used to determine whether the concentration of target pollutants in each sub-area of ​​concern is spatially clustered at high or low values, including: The target pollutant concentration values ​​of each sub-area of ​​concern are counted to obtain the second clustering result, which is expressed as: Among them, G(d) represents the second clustering result, d is the distance, W ij (d) is the spatial weight defined by the distance rule, i and j represent the codes of different sub-areas of interest; n represents the number of all sub-areas of interest; x represents the target pollutant concentration value of each sub-area of ​​interest; The expected value of the second clustering result is calculated under the assumption that there is no spatial aggregation: E(G)=W ij (d) / [n(n-1)] The statistical test value of the second clustering result under the condition of normal distribution is calculated as follows: Among them, when the second clustering result G(d) is higher than the expected value E(G) and the statistical test value Z is significant, it indicates that there is a high-value aggregation in space; when the second clustering result G(d) is lower than the expected value E(G) and the Z value is significant, it indicates that there is a low-value aggregation in space; when the second clustering result G(d) tends to the expected value E(G), it indicates that there is a random distribution feature in space.

6. The method according to claim 4, characterized in that Statistically identify the high-value clusters or low-value clusters of target pollutant concentrations in each sub-area of ​​interest, including: According to the target pollutant concentration value of each sub-area of ​​concern and the average value of the target pollutant concentration of all sub-areas of concern, the score value of each cluster position of the target pollutant is calculated. The score value is expressed as: in, i and j represent the codes of different sub-areas of concern; n represents the number of all sub-areas of concern; x represents the target pollutant concentration value of each sub-area of ​​concern; is the average value of the target pollutant concentration in all sub-areas of concern; W ij Spatial weight matrix; According to the size relationship between the score value and different preset thresholds, the point distribution of different levels is determined, wherein when the score value exceeds the first preset threshold, it is identified as a high-value point, and when the score value exceeds the second preset threshold, it is identified as a low-value point.

7. The method according to claim 3, characterized in that Also includes: Identify the key pollution sources that cause the concentration change of at least one target pollutant at different time points, and output the prediction results of the key pollution sources, including: Analyze the key pollution sources that cause changes in the concentration of each target pollutant in the overall tested area at different time points; Analyze the key pollution sources that cause the concentration changes of each target pollutant in each sub-area of ​​interest of the area to be tested at different time points; For the high-value points screened out in each sub-area of ​​concern, different high-value concentration periods are selected to analyze the key pollution sources that cause changes in the concentration of each target pollutant.

8. The method according to claim 1, characterized in that Preprocess the data, including: The collected atmospheric pollutant data are cleaned, screened and processed for outliers, including statistics of missing values, statistics of outliers, and filling in missing values ​​using the k-nearest neighbor mean interpolation method, and the data are normalized to the interval [0,1].

9. The method according to claim 1, characterized in that: The concentration prediction model is constructed using the LSTM long short-term memory network.

10. A device for predicting atmospheric pollutant concentration based on deep machine learning, characterized in that: The method for predicting the concentration of atmospheric pollutants based on deep machine learning according to any one of claims 1 to 9 is adopted, and the device for predicting the concentration of atmospheric pollutants based on deep machine learning comprises: The data set construction module is used to collect the atmospheric pollutant concentration data at different time points in the test area and pre-process the data; A model training module is used to train the constructed concentration prediction model using the pre-processed atmospheric pollutant concentration data as a training set; The prediction module is used to use the trained concentration prediction model to predict the concentration change of at least one target pollutant in the test area at different time nodes.

Images