An artificial intelligence monitoring and early warning method for air pollution integrating multi-source meteorological data

Abstract

The present invention discloses an artificial intelligence monitoring and early warning method for atmospheric pollution that integrates multi-source meteorological data, which belongs to the field of atmospheric pollution monitoring technology, including preliminary selection of model factors, acquisition of model factors, screening of model factors, model establishment and training, model effect testing, and monitoring and early warning. The present invention screens model factors that are sensitive to atmospheric pollution, combines the forecast results of air quality numerical models, integrates minute-level three-dimensional multi-source observation data, and establishes an atmospheric pollution monitoring and early warning model to achieve automatic reminder services, providing an objective basis for atmospheric pollution prevention and control decisions; compared with numerical models, the calculation efficiency is high, the inversion speed is fast, and the accuracy rate is significantly improved; the minute-level observation data is rolled and assimilated into the model in real time, which can greatly improve the refinement and objectivity of the monitoring results.

Description

Technical Field

[0001] The present invention relates to the technical field of air pollution monitoring, and in particular to an artificial intelligence monitoring and early warning method for air pollution that integrates multi-source meteorological data. Background Art

[0002] The occurrence of air pollution weather is the result of the combined effects of natural and human factors. As one of the important factors affecting the atmospheric environment, meteorological conditions restrict the dilution, diffusion, transportation and transformation of atmospheric pollutants, and thus affect the distribution and concentration of atmospheric pollutants. The quality of urban air is closely related to meteorological conditions. To further improve the ability to accurately prevent and control air pollution, we must first have the ability to finely monitor and warn of meteorological conditions for air pollution.

[0003] Traditional meteorological condition analysis mainly relies on the subjective analysis and judgment of weather forecasters, and relies more on human experience. With the increase of various weather observation methods and the improvement of observation refinement, various minute-level three-dimensional observation data have formed a massive data set. Relying solely on subjective analysis, the data utilization rate is low, the conclusion accuracy is poor, and the service effect is not good. Therefore, an artificial intelligence monitoring and early warning method for atmospheric pollution that integrates multi-source meteorological data is proposed. Summary of the invention

[0004] The technical problem to be solved by the present invention is: how to solve the problems existing in the prior art of relying solely on subjective analysis, low data utilization, poor conclusion accuracy, poor service effect, etc., and provide an artificial intelligence monitoring and early warning method for atmospheric pollution that integrates multi-source meteorological data.

[0005] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0006] S1: Preliminary model factors

[0007] Based on the forecaster's experience and air pollution monitoring indicators, the boundary layer temperature, humidity and wind, pollutant concentration, environmental meteorological parameters and temperature, humidity and wind reconstruction variables were initially selected to predict PM2.5 and O 3 Model factors for pollution occurrence;

[0008] S2: Get model factors

[0009] Obtain model factors from actual meteorological and environmental data and meteorological and environmental forecast results as training samples;

[0010] S3: Screening model factors

[0011] Use the Gini index score of the random forest algorithm to evaluate the feature importance of the training samples and screen out the model factors;

[0012] S4: Model building and training

[0013] Based on the selected model factors, a training data set is established, and an air pollution monitoring and early warning model is established using the LSTM network. The air pollution monitoring and early warning model includes a PM2.5 monitoring and early warning model, an O 3 Monitoring and early warning models;

[0014] S5: Model effect test

[0015] The PM2.5 monitoring and early warning model, O 3 The forecasting effects of the monitoring and early warning models are tested respectively;

[0016] S6: Monitoring and early warning

[0017] For PM2.5 monitoring and early warning models that achieve the expected forecast effect, 3 Monitoring and early warning models can realize PM2.5, O 3 Pollution monitoring and early warning.

[0018] Furthermore, in step S2, the meteorological live data include temperature, humidity, wind speed, horizontal wind UV component, air pressure, 24-hour pressure change, and precipitation obtained from the ground meteorological observation station, horizontal wind UV component, horizontal wind direction, horizontal wind speed, and vertical speed at different heights obtained from the wind profiler radar, and temperature, humidity, potential temperature, variable temperature, and inversion at different heights obtained from the microwave radiometer; the environmental live data include near-ground PM2.5, O 3 、NO 2 Hourly concentration, daily average concentration; weather forecast results come from meteorological numerical models; environmental forecast results come from air quality numerical models.

[0019] Furthermore, in step S2, the weather forecast results are updated twice a day, the environmental forecast results are updated once a day, the environmental real-time data are updated once an hour, and the weather real-time data are updated every minute, among which the update interval of the meteorological observation station is 5 minutes, the update interval of the wind profiler radar is 6 minutes, and the update interval of the microwave radiometer is 2 minutes.

[0020] Furthermore, in step S2, the model factors include near-ground temperature, humidity, wind speed, wind UV component, air pressure, 24h pressure change, precipitation, low cloud cover, and total cloud cover; horizontal wind UV component, horizontal wind direction, horizontal wind speed, vertical speed, temperature, humidity, potential temperature, and variable temperature on pressure surfaces of 1000hPa, 925hPa, 850hPa, and 700hPa; mixing layer height, convergence coefficient, inversion, potential temperature difference, ventilation volume U component, and ventilation volume V component calculated from physical parameters; near-ground PM2.5, O 3 、NO 2Hourly concentration and multi-year average concentration; date factor reflecting time characteristics.

[0021] Furthermore, in the step S3, based on the sklearn integrated learning library in the Python language, sklearn is used to perform preliminary learning and training. During the training process, the score of each feature is calculated, and the result is standardized so that the sum of the importance of all features is equal to 1, and the importance indexes of different feature quantities are obtained. The importance indexes are sorted from large to small, and the voting cycle is performed to finally screen out the model factors, wherein the features are the obtained model factors.

[0022] Furthermore, in step S3, the model factors of the PM2.5 monitoring and early warning model screened out are specifically as follows:

[0023] Total cloud cover, sea level pressure, 24h pressure change, air temperature, 24h temperature change, inversion, humidity, precipitation, height of mixed layer, 10m wind U component, 10m wind V component; 925hPa U, V wind, wind direction, vertical speed; PM2.5 hourly concentration at the last moment, date factor;

[0024] The screened O 3 The model factors of the monitoring and early warning model are as follows:

[0025] 2-meter temperature, 10-meter wind V component, 10-meter wind U component, mixing layer height, 2-meter relative humidity, total cloud cover, precipitation, multi-year average O 3 Hourly concentration, NO 2 Hourly concentration, multi-year average NO 2 Hourly concentration, date factor.

[0026] Furthermore, in step S4, the WOA algorithm is used to tune the model during training to confirm the parameters in the LSTM to optimize the model performance, wherein the parameters include the number of hidden units, training cycle, learning rate, and L2 regularization parameter.

[0027] Furthermore, in step S6, the forecast result is displayed in a graphical manner; and monitoring and early warning prompts are implemented by setting monitoring thresholds.

[0028] Compared with the prior art, the present invention has the following advantages: the artificial intelligence monitoring and early warning method for atmospheric pollution that integrates multi-source meteorological data, by screening model factors that are sensitive to atmospheric pollution, combining the forecast results of air quality numerical models, integrating minute-level three-dimensional multi-source observation data, and establishing an atmospheric pollution monitoring and early warning model, realizes automatic reminder services, and provides an objective basis for decision-making on atmospheric pollution prevention and control; compared with numerical models, the calculation efficiency is high, the inversion speed is fast, and the accuracy is significantly improved; the real-time rolling assimilation of minute-level observation data into the model can greatly improve the refinement and objectivity of the monitoring results. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of an artificial intelligence monitoring and early warning method for air pollution integrating multi-source meteorological data in an embodiment of the present invention;

[0030] Figure 2 It is a schematic diagram of a technical route for establishing an air pollution monitoring and early warning model using an LSTM machine learning algorithm in an embodiment of the present invention;

[0031] Figure 3a is an iteration curve of the number of hidden units when optimizing the PM2.5 model in an embodiment of the present invention;

[0032] Figure 3b It is a training cycle iteration curve when tuning the PM2.5 model in an embodiment of the present invention;

[0033] Figure 3c is a learning rate iteration curve when tuning the PM2.5 model in an embodiment of the present invention;

[0034] Figure 3d It is the L2 parameter iteration curve when the PM2.5 model is tuned in the embodiment of the present invention;

[0035] Figure 4 This is a comparison chart of the errors of the PM2.5 model on the training set before (blue) and after (red) parameter optimization in the embodiment of the present invention;

[0036] Figure 5a In the embodiment of the present invention, 3 Iterative curve of the number of hidden units during model tuning;

[0037] Figure 5b In the embodiment of the present invention, 3 Training cycle iteration curve during model tuning;

[0038] Figure 5c In the embodiment of the present invention, 3 Learning rate iteration curve during model tuning;

[0039] Figure 5d In the embodiment of the present invention, 3 L2 parameter iteration curve during model tuning;

[0040] Figure 6 : is the parameter before optimization (blue) and after optimization (red) in the embodiment of the present invention. 3 Error comparison chart of the model on the training set;

[0041] Figure 7 It is a comparison diagram of the correlation coefficient between the PM2.5 model and the actual situation (blue) and the correlation coefficient between the CMA model and the actual situation (red) in the embodiment of the present invention;

[0042] Figure 8 : is a comparison diagram between the root mean square error (blue) of the PM2.5 model and the actual situation and the root mean square error (red) of the CMA model and the actual situation in an embodiment of the present invention;

[0043] Fig. 9 In the embodiment of the present invention, 3 Comparison of the correlation coefficient between the model and the actual situation (blue) and the correlation coefficient between the CMA model and the actual situation (red);

[0044] Fig.10 In the embodiment of the present invention, 3 Comparison of the root mean square error between the model and the actual situation (blue) and the root mean square error between the CMA mode and the actual situation (red). DETAILED DESCRIPTION

[0045] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0046] The model establishment cycle is determined by statistically analyzing the frequency and time distribution of polluted weather in the monitoring area. The frequency and distribution of air pollution in a certain province in China from 2013 to 2022 are statistically analyzed. Polluted weather with PM2.5 as the primary pollutant is the most common, mostly occurring in autumn and winter, with the highest frequency from November to February of the following year. Since 2018, summer ozone pollution has gradually become prominent, mostly occurring in June and July. Here we focus on the PM2.5 in autumn and winter and the O 3 Pollution, seasonal air pollution monitoring and early warning model, in which the PM2.5 modeling period is from October to March of the following year, 3 The modeling period is from April to October.

[0047] like Figure 1 As shown, the method for artificial intelligence monitoring and early warning of atmospheric pollution by integrating multi-source meteorological data in this embodiment is further described below, which specifically includes the following steps:

[0048] 1. Preliminary model factors

[0049] Based on the forecaster's experience and the air pollution monitoring indicators of the province, the boundary layer temperature, humidity and wind, pollutant concentration, environmental meteorological parameters and temperature, humidity and wind reconstruction variables were initially selected to predict PM2.5 and O 3 The model factors closely related to pollution occurrence are shown in Table 1 below:

[0050] Table 1 Preliminary selected model factors

[0051]

[0052]

[0053] 2. Obtain model factors

[0054] This method mainly obtains model factors from actual meteorological and environmental data and meteorological and environmental forecast results, and gives PM2.5 and O 3 The concentration is refined and the forecast results are dynamically corrected. The actual meteorological data include the temperature (2T 0 )、Humidity(2R 0 ), wind speed (10mwspd 0 ), horizontal wind UV component (10U 0 , 10V 0 ), air pressure (MSL 0 ), 24-hour voltage change (MSL_24H 0 ), precipitation (TP 0 ), horizontal wind UV component, horizontal wind direction, horizontal wind speed, vertical speed at different heights (1000hPa, 925hPa, 850hPa, 700hPa) obtained from wind profiler radar, temperature, humidity, potential temperature, temperature change, inversion, etc. at different heights (1000hPa, 925hPa, 850hPa, 700hPa) obtained from microwave radiometer; environmental real-time data include near-ground PM2.5, O 3 、NO 2 hourly concentration, daily average concentration, etc.; meteorological forecast results come from meteorological numerical models; environmental forecast results come from air quality numerical models, here we use the PM2.5, O 3 Hourly concentration results.

[0055] In this embodiment, the training data in the modeling process of the method of the present invention comes from the historical data of EAR5 of the European Center, and the fine grid forecast results of the European Center EC are used when the model is actually running; the model factors obtained include near-ground temperature, humidity, wind speed, wind UV component, air pressure, 24-hour (h) pressure change, precipitation, low cloud cover, total cloud cover; horizontal wind UV component, horizontal wind direction, horizontal wind speed, vertical speed, temperature, humidity, potential temperature, and variable temperature on the isobaric surfaces of 1000hPa, 925hPa, 850hPa, and 700hPa; the height of the mixing layer, the convergence coefficient, the inversion temperature, the potential temperature difference, the ventilation volume U component, and the ventilation volume V component calculated by the physical parameters; near-ground PM2.5, O 3 、NO 2 Hourly concentration and multi-year average concentration; date factors reflecting time characteristics, etc.

[0056] Among all model factors, meteorological forecast results (weather forecast model EC) are updated twice a day (reported at 08:00 and 20:00 Beijing time respectively), environmental forecast results (air quality model CMA) are updated once a day (reported at 14:00 Beijing time), environmental real-time data (air quality real-time data) are updated once an hour, and the update frequency of meteorological real-time data is at the minute level (5 minutes for meteorological observation stations, 6 minutes for wind profile radars, and 2 minutes for microwave radiometers). To facilitate business applications, the start time of the monitoring model can be determined in combination with service requirements. Here, the model is modeled at 15:00. All model factors use the most recent data, and the minute-level three-dimensional real-time data is integrated into the model as soon as possible to improve the timeliness of the model factors. You can also establish a rolling update model as needed (for example, establish an hourly update model with the fastest update frequency of minutes), so that the latest real-time data can be iterated into the forecast results to realize the rolling assimilation application of real-time data.

[0057] 3. Use the random forest algorithm to screen model factors

[0058] Random forest is an ensemble learning algorithm based on decision trees. In random forest, the importance of a feature can be measured by calculating the average change in its impurity in all decision trees. This average change in impurity is the Gini index. For the i-th feature (the feature here refers to the model factor obtained above), the Gini index score calculation process is as follows:

[0059] For each decision tree, calculate the Gini index when using the i-th feature for splitting. Assuming that the Gini index before the split is Gini_before and the Gini index after the split is Gini_after, then the change in the Gini index of this feature is Gini_change = Gini_before-Gini_after.

[0060] The average Gini index change in all decision trees is calculated to get the average Gini index score of the i-th feature.

[0061] Specifically, for a given dataset D, assuming that the two sub-datasets obtained by splitting using the i-th feature F are D1 and D2, the impurity Gini_after after splitting can be calculated as follows:

[0062] First, calculate the number of samples in each category in D. Assuming there are k categories, the number of samples in the jth category in D is N_j.

[0063] Secondly, calculate the number of samples of each category in D1 and D2. Assuming that there are a_j samples of the jth category in D1 and b_j samples of the jth category in D2, then the number of samples of the jth category in D1 and D2 is a_j and b_j respectively.

[0064] Then, calculate the impurity of D1 and D2. Assuming that the proportion of samples of the jth category in D1 and D2 is p1_j and p2_j, the impurity of D1 and D2 is Gini1 = 1-sum(a_j / N*p1_j) and Gini2 = 1-sum(b_j / N*p2_j) respectively.

[0065] Finally, the impurity change Gini_change = (Gini1 + Gini2) / 2 - Gini(D) when splitting using F is calculated, where Gini(D) is the impurity of D.

[0066] For each feature F, repeat the above steps to get its Gini index score in all decision trees. Then sort the features according to the size of these scores. The higher the score, the greater the importance of the feature.

[0067] In this embodiment, the method uses the Gini index score of the random forest algorithm to evaluate the feature importance of the training samples: based on the sklearn integrated learning library in the Python language, sklearn is used for preliminary learning and training. During the training process, the score of each feature (the "feature" here refers to the model factor obtained above) is calculated, and the results are standardized so that the sum of the importance of all features is equal to 1, and the importance indexes of different feature quantities are obtained. They are sorted from large to small according to the importance index, and the voting cycle is determined to finally screen out the model factor.

[0068] As a specific example, the PM2.5 model factors screened out are as follows:

[0069] Total cloud cover (TCC), sea level pressure (MSL), 24h pressure change (MSL_24h), air temperature (2T), 24h temperature change (2T_24h), inversion (InvT), humidity (2R), precipitation (TP), mixing layer height (MLH), 10m wind U component (10U), 10m wind V component (10V); 925hPa U, V wind, wind direction, vertical speed; PM2.5 hourly concentration at the previous moment, date factor (month, day).

[0070] As a specific example, the selected O 3 The model factors are as follows:

[0071] 2-meter temperature (2T), 10-meter wind V component (10V), 10-meter wind U component (10U), mixing layer height (MLH), 2-meter relative humidity (2R), total cloud cover (TCC), precipitation (TP), multi-year average O 3 Hourly concentration, NO 2 Hourly concentration, multi-year average NO 2 Hourly concentration, date factor (month, day).

[0072] 4. Model training

[0073] According to the selected model factors, a training data set is established with a data set period of 2 years (due to the adjustment of atmospheric pollution emissions, the upgrade of numerical models, etc., the training period should not be too long, and the model can be upgraded by dynamically adjusting the training data set). The model is established using LSTM (long short-term memory neural network algorithm). The model includes PM2.5 monitoring and early warning model (i.e. PM2.5 model), O 3 Monitoring and early warning model (O 3 Model).

[0074] LSTM refers to the long short-term memory network, which is a time recurrent neural network that is mainly used to solve the long-term dependency problem in RNN (recurrent neural network). Compared with ordinary RNN units, LSTM adds a memory unit C that stores past information, allowing it to process and update past state information. The reason why the model uses a long short-term memory network is that the evolution characteristics of pollutants have strong spatiotemporal continuity. Compared with traditional statistical methods, deep learning methods have added evolution characteristics that consider changes before and after. It has great advantages in studying time series changes. The technical route diagram of using LSTM machine learning algorithm to establish an air pollution monitoring and early warning model is as follows Figure 2 shown.

[0075] In this embodiment, the whale swarm intelligent optimization algorithm (WOA) is used to tune the model during training. The algorithm draws inspiration from the group hunting behavior of humpback whales and solves the optimization problem by simulating the process of whale groups searching, surrounding, chasing and attacking prey. In the whale swarm intelligent optimization algorithm, each whale represents a potential solution. The algorithm simulates the process of whales searching for prey in the ocean, allowing whales to continuously explore and hunt prey, and try to improve their solutions. Specifically, each whale will update its position and speed according to the current optimal solution and its own experience during the search process.

[0076] In this embodiment, the method of the present invention uses the WOA algorithm to tune the model to confirm the parameters in the LSTM to optimize the model performance. The PM2.5 model parameter tuning results are as follows: Figure 3a-3d As shown, O 3 The model parameter adjustment results are as follows Figure 5a-5d shown.

[0077] To be more specific, the optimal parameters of the PM2.5 model are: number of hidden units: 61; maximum training cycle: 256; InitialLearnRate (learning rate): 0.0115; L2 regularization parameter: 0.0027.

[0078] As more specific, O 3 The optimal model parameters are: number of hidden units: 163; maximum training cycle: 65; InitialLearnRate: 0.015; L2 regularization parameter: 0.0013.

[0079] 5. Model effect test

[0080] (1) PM2.5 model

[0081] The prediction effect of the PM2.5 model mentioned above was tested using a city in the province from October 2023 to February 2024 as a test data set. The daily average value was calculated using the LSTM model (PM2.5 model) and the CMA forecasted PM2.5 hourly average value, and compared with the actual PM2.5 daily average value. The results (see Figure 7 ) shows that the correlation coefficient between the LSTM model and the actual situation on days 1-3 is above 0.5, which is improved to varying degrees compared with CMA. The correlation coefficient on days 4-7 is between 0.44-0.49, which is slightly improved compared with CMA except for the 144h time period. Figure 8 ) Except for the 96-hour forecast time, the other time periods are lower than that of CMA. The model significantly improves the forecast of PM2.5 concentration within 3 days.

[0082] As a more specific comparison table of the correlation coefficient between PM2.5 model and actual situation and the correlation coefficient between CMA model and actual situation is shown in Table 2 below:

[0083] Table 2 Comparison of correlation coefficients between PM2.5 model and actual situation and correlation coefficients between CMA model and actual situation

[0084] Correlation coefficient 24h 48h 72h 96h 120h 144h 168h CMA 0.45 0.49 0.48 0.43 0.41 0.48 0.46 LSTM 0.67 0.6 0.53 0.44 0.45 0.47 0.49

[0085] As a more specific comparison table of the root mean square error between the PM2.5 model and the actual situation and the root mean square error between the CMA model and the actual situation is shown in Table 3 below:

[0086] Table 3 Comparison table of PM2.5 model and actual root mean square error and CMA model and actual root mean square error

[0087] Root mean square error 24h 48h 72h 96h 120h 144h 168h CMA 39.7 38.3 39.3 40.8 41.9 40.4 40.5 LSTM 35.3 35.5 38 41.1 40.4 39.3 38.5

[0088] (2)O 3 Model

[0089] A city in the province from April to October 2023 is used as a test data set to test O 3 The model prediction effect was tested. The results show (see Fig. 9 )1-4 days LSTM model (O 3 The correlation coefficient between the model and the actual situation is above 0.75, and is about 0.6 on the 5th to 7th day, which is significantly higher than that of CMA at each time point. Fig.10 ) decreased significantly at each time point, and the prediction effect of LSTM model was significantly improved compared with CMA.

[0090] As more specific, O 3 The comparison table of the correlation coefficient between the model and the actual situation and the correlation coefficient between the CMA model and the actual situation is shown in Table 4 below:

[0091] Table 4O 3 Comparison table of correlation coefficients between model and actual situation and correlation coefficients between CMA model and actual situation

[0092]

[0093]

[0094] As more specific, O 3 The comparison table of the root mean square error between the model and the actual situation and the root mean square error between the CMA mode and the actual situation is shown in Table 5 below:

[0095] Table 5O 3 Comparison table of root mean square error between model and actual situation and root mean square error between CMA mode and actual situation

[0096] Root mean square error 24h 48h 72h 96h 120h 144h 168h CMA 59.4 60.6 62 61.8 64.2 64.1 67.2 LSTM 35.9 36.5 38.6 42.4 44.9 48 45.2

[0097] 6. Automatic monitoring and early warning

[0098] For models that achieve the expected forecasting effect, the model business operation is realized by accessing various data of business applications, and the forecast results are displayed in a graphical manner. Automatic warning prompts are achieved by setting monitoring thresholds, such as setting the threshold to PM2.5 hourly concentration greater than 75ug / m 3 , O 3 Hourly concentration greater than 160ug / m 3 ,The threshold can be set arbitrarily according to service requirements.

[0099] 7. Dynamic model upgrade

[0100] In order to solve the problem of reduced model forecasting effect caused by the adjustment of pollution emission sources in various places and changes in numerical models, this method has established a process-based and automated design for model training. As long as the training data set is updated in accordance with the specified data format, the model can be retrained to achieve dynamic model adjustment. In addition, the actual data and numerical model data used in the method can also be adjusted to other modes. For example, ground meteorological station data can be replaced by regional automatic stations, wind profile radar data can be replaced by laser wind radar and other equipment, microwave radiometers can be replaced by Raman temperature and humidity radar data, EC and CMA numerical models can also be replaced by other model results. Similarly, after standardizing the data format, model training and application can be carried out conveniently and quickly. At the same time, by replacing the data sets with different sites, this method can quickly establish artificial intelligence models in different locations, which is convenient for the promotion and application of the method.

[0101] To sum up, the artificial intelligence monitoring and early warning method for atmospheric pollution that integrates multi-source meteorological data in the above-mentioned embodiment, by screening model factors that are sensitive to atmospheric pollution, combining the forecast results of the air quality numerical model, integrating minute-level three-dimensional multi-source observation data, and establishing an atmospheric pollution monitoring and early warning model, realizes automatic reminder services, and provides an objective basis for atmospheric pollution prevention and control decision-making; compared with the numerical model, it has high calculation efficiency, fast inversion speed, and significantly improved accuracy; the real-time rolling assimilation of minute-level observation data into the model can greatly improve the refinement and objectivity of the monitoring results; through data format conversion and model factor adjustment, it can quickly and easily carry out model reconstruction, which is convenient for transplantation and application in different regions, different meteorological characteristics, and different pollution scenarios.

[0102] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An artificial intelligence monitoring and early warning method for air pollution integrating multi-source meteorological data, characterized in that: The following steps are involved: S1: Preliminary model factors According to the forecaster's experience and air pollution monitoring indicators, model factors related to PM2.5 and O3 pollution occurrence were preliminarily selected based on boundary layer temperature, humidity and wind, pollutant concentrations, ambient meteorological parameters, and temperature, humidity and wind reconstruction variables; S2: Get model factors Obtain model factors from actual meteorological and environmental data and meteorological and environmental forecast results as training samples; In step S2, the model factors include near-ground temperature, humidity, wind speed, wind UV component, air pressure, 24h pressure change, precipitation, low cloud cover, and total cloud cover; horizontal wind UV component, horizontal wind direction, horizontal wind speed, vertical speed, temperature, humidity, potential temperature, and variable temperature on pressure surfaces such as 1000hPa, 925hPa, 850hPa, and 700hPa; mixing layer height, convergence coefficient, inversion, potential temperature difference, ventilation volume U component, and ventilation volume V component calculated from physical parameters; near-ground PM2.5, O3, and NO2 hourly concentrations and multi-year average concentrations; and date factors reflecting time characteristics; S3: Screening model factors Use the Gini index score of the random forest algorithm to evaluate the feature importance of the training samples and screen out the model factors; In step S3, the model factors of the screened PM2.5 monitoring and early warning model are specifically as follows: Total cloud cover, sea level pressure, 24h pressure change, air temperature, 24h temperature change, inversion, humidity, precipitation, height of mixed layer, 10m wind U component, 10m wind V component; 925hPa U, V wind, wind direction, vertical speed; PM2.5 hourly concentration at the last moment, date factor; The model factors of the selected O3 monitoring and early warning model are as follows: 2-meter temperature, 10-meter wind V component, 10-meter wind U component, mixing layer height, 2-meter relative humidity, total cloud cover, precipitation, multi-year average O3 hourly concentration, NO2 hourly concentration, multi-year average NO2 hourly concentration, date factor; S4: Model building and training According to the selected model factors, a training data set is established, and an air pollution monitoring and early warning model is established using the LSTM network. The air pollution monitoring and early warning model includes a PM2.5 monitoring and early warning model and an O3 monitoring and early warning model. S5: Model effect test The test data set is used to test the forecasting effects of PM2.5 monitoring and early warning model and O3 monitoring and early warning model respectively; S6: Monitoring and early warning For the PM2.5 monitoring and early warning model and O3 monitoring and early warning model that have achieved the expected forecasting effect, PM2.5 and O3 pollution monitoring and early warning are realized by accessing various data from business applications.

2. The method for air pollution artificial intelligence monitoring and early warning by integrating multi-source meteorological data according to claim 1 is characterized in that: In step S2, the real-time meteorological data include temperature, humidity, wind speed, horizontal wind UV component, air pressure, 24-hour pressure change, precipitation obtained from ground meteorological observation stations, horizontal wind UV component, horizontal wind direction, horizontal wind speed, vertical speed at different heights obtained from wind profiler radars, and temperature, humidity, potential temperature, temperature change, and inversion at different heights obtained from microwave radiometers; the real-time environmental data include hourly concentrations and daily average concentrations of near-ground PM2.5, O3, and NO2; the meteorological forecast results come from meteorological numerical models; and the environmental forecast results come from air quality numerical models.

3. The method for air pollution artificial intelligence monitoring and early warning by integrating multi-source meteorological data according to claim 2 is characterized in that: In step S2, the weather forecast results are updated twice a day, the environmental forecast results are updated once a day, the environmental real-time data are updated once an hour, and the weather real-time data are updated every minute, among which the update interval of the meteorological observation station is 5 minutes, the update interval of the wind profiler radar is 6 minutes, and the update interval of the microwave radiometer is 2 minutes.

4. The method for air pollution artificial intelligence monitoring and early warning by integrating multi-source meteorological data according to claim 1 is characterized in that: In the step S3, based on the sklearn integrated learning library in the Python language, sklearn is used to perform preliminary learning and training. During the training process, the score of each feature is calculated, and the result is standardized so that the sum of the importance of all features is equal to 1, and the importance indexes of different feature quantities are obtained. The importance indexes are sorted from large to small, and the voting cycle is performed to finally screen out the model factors, wherein the features are the obtained model factors.

5. The method for air pollution artificial intelligence monitoring and early warning by integrating multi-source meteorological data according to claim 1 is characterized in that: In step S4, the WOA algorithm is used to tune the model during training to confirm the parameters in the LSTM to optimize the model performance, wherein the parameters include the number of hidden units, training cycle, learning rate, and L2 regularization parameter.

6. The method for monitoring and early warning of atmospheric pollution by artificial intelligence integrating multi-source meteorological data according to claim 1 is characterized in that: In step S6, the forecast result is displayed in a graphical manner; monitoring and early warning prompts are implemented by setting monitoring thresholds.

Images