Method for ozone and particulate matter prediction and early warning based on multi-mode air quality model

Abstract

The application discloses a method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model, and the method comprises the following steps: obtaining meteorological prediction data, satellite remote sensing data, pollution source emission inventory and air quality monitoring data of a target region; inputting the obtained data into a CMAQ air quality model, a CAMx air quality model, a particulate matter ozone neural network prediction model and a satellite remote sensing aerosol inversion model to obtain model prediction values; adopting a neural network model architecture, and on the basis of long-term historical related data, cyclically training the model prediction values and actual monitoring values to establish a multi-mode air quality model; inputting the latest related data of the target region into the multi-mode air quality model, and predicting and early warning the ozone and particulate matter of the target region according to the output results of the multi-mode air quality model. The method has good effects on particulate matter and ozone prediction and early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of weather prediction, and in particular to a method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model. BACKGROUND

[0002] At present, particulate matter and ozone have become the main air pollutants affecting the air quality of cities and regions in China, and their coordinated control has become the focus of air quality improvement and the key to winning the battle for blue skies. Particulate matter and ozone have complex correlations, and they not only have common precursors, but also interact with each other in the atmosphere through various pathways.

[0003] In terms of particulate matter and ozone coordinated governance, existing scientific research methods are facing challenges. The mechanism of the mutual influence of ozone and particulate matter has not been clearly concluded, and the technology for predicting and warning ozone and particulate matter is not mature.

[0004] In terms of particulate matter and ozone prediction and early warning, numerical prediction and statistical prediction are mainly used. Numerical prediction mainly uses air quality models to systematize complex atmospheric physics and chemistry models, establishes models related to pollutant emissions, weather, and chemical reactions, and simulates changes in air quality. Statistical prediction is based on historical data from online automatic air quality monitoring, and predicts future air quality by analyzing the change rules of various pollution factors. These two methods have their own advantages and disadvantages, and the data cannot support each other, resulting in unstable accuracy of the prediction results.

[0005] In the specific implementation process, a large amount of data for analysis, the instability of long-distance transoceanic transmission makes it impossible to obtain the required data in time, and the calculation amount required by high-resolution mode simulation is huge, the existing mainstream computing framework cannot meet the supercomputing demand, and the timeliness of prediction and early warning cannot be guaranteed. SUMMARY

[0006] The purpose of the present application is to overcome the deficiencies in the prior art and respond to the call of China's "adherence to the application of information technology innovation road", the present application provides a method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model, which has good effect on particulate matter and ozone prediction and early warning.

[0007] Technical scheme: In order to achieve the above purpose, the method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model comprises the following steps:

[0008] Step 1: Obtain long-term historical related data of the target area, including meteorological prediction data, satellite remote sensing data, pollution source emission inventory and air quality monitoring data;

[0009] Step two: meteorological prediction data and pollution source emission inventory into two third generation air quality model of CMAQ and CAMx, obtain the air quality data, meteorological data and rainfall prediction value of the target area; air quality monitoring data into the particulate matter ozone neural network prediction model, obtain the particulate matter concentration and ozone concentration prediction value of the target area; satellite remote sensing data into the satellite remote sensing aerosol inversion model, obtain the AOT prediction value of the target area;

[0010] Step three: using the model architecture of neural network, on the basis of long-term historical related data, the model prediction value and actual monitoring value are trained circularly, so as to establish a multi-mode air quality model; the steps of establishing a multi-mode model include: (d1) generating training data; (d2) performing validity processing and normalization processing on the data; (d3) establishing an LSMT neural network model, and then using an adaptive gradient descent algorithm AdaGrad as a back propagation iteration; (d4) after training, saving the recurrent neural network model; the model predicts the hourly concentration of particulate matter and ozone in the future period according to the monitoring data, and extracts some days of data output; (d5) entering the next period, the LSTM neural network model is trained again;

[0011] Step four: input the latest related data of the target area into the multi-mode air quality model, and predict and warn the ozone and particulate matter of the target area according to the output result of the multi-mode air quality model.

[0012] Further, in step two, the specific implementation steps of inputting data into two third generation air quality models of CMAQ and CAMx are as follows:

[0013] (a1) formulate the geographical area of model simulation, determine the grid parameters, generate meteorological prediction data through WRF meteorological research and prediction model, and obtain hourly emission inventory data through emission source list algorithm;

[0014] (a2) input meteorological prediction data and emission inventory data, simulate SO2, NO2, CO, ozone, PM2.5, temperature, humidity, wind speed, wind direction and rainfall of each grid and each period in the key area through two air quality models of CMAQ and CAMx;

[0015] (a3) according to the fitting function, the prediction results are optimized and fitted through the nonlinear least square fitting algorithm; the formula for fitting the nonlinear least square formula using high order polynomial is:

[0016] ;

[0017] In the formula, f is the fitting value, and p is the prediction value;

[0018] And the error is calculated by least square method, the error function is:

[0019] ;

[0020] Through machine learning, the final expression of f(x) is obtained; and the fitting formula f(x) is used to output the prediction fitting values of CMAQ and CAMX models respectively.

[0021] Further, in step two, the data is substituted into the particulate matter ozone neural network prediction model, and the specific implementation steps are as follows:

[0022] (b1) Export all hourly historical data of air quality automatic monitoring stations in a specific area from the platform database;

[0023] (b2) According to the time sequence, generate a training data list and a verification data list;

[0024] (b3) Train the training data using a recurrent neural network, use several days as a cycle, and output the hourly concentration of ozone and particulate matter in several days within a cycle:

[0025] (b4) After the training is completed, save this recurrent neural network model; after 0 o'clock in the morning every day, the model will predict the hourly concentration of particulate matter and ozone in several days within a cycle according to the 24-hour monitoring data of the previous day;

[0026] (b5) After entering a cycle, the neural network is trained again to obtain better model prediction accuracy.

[0027] Further, in step (b1), for historical data with missing data, abnormal values, and unsuitable for model training, data cleaning and normalization processing are performed; the normalization equation is:

[0028] .

[0029] Further, in step two, the data is substituted into the satellite remote sensing aerosol inversion model prediction, and the specific implementation steps are as follows:

[0030] (c1) Download MODIS data from NASA website;

[0031] (c2) Invert the apparent reflectivity of the target through the polarization radiation transfer formula, and the polarization radiation transfer formula is:

[0032] ;

[0033] Wherein R is the apparent reflectance, F0 is the extraterrestrial solar irradiance flux, I is the top of the atmosphere radiance, μ is the observation zenith angle cosine of the satellite view angle, μ0 is the solar zenith angle cosine, φ is the relative direction angle of the scattering radiation propagation analysis and the incident solar direction;

[0034] (c3) According to the target area, the apparent reflectance of 0.412um wavelength is inverted, and the temperature brightness value inverted according to 11nm and 12nm wavelengths and the spatial variation of 1.38um MODIS reflectance are used to judge the weather conditions of the target area, such as cloud layer or ground ice and snow cover, stop inversion, and use the historical model prediction value as the result of this model operation;

[0035] (c4) The AOT estimated value is obtained by calculating the normalized vegetation index and the normalized ice and snow index parameters;

[0036] (c5) By identifying the land type and aerosol type in the target area grid, the corresponding parameters are adjusted to obtain the AOT simulation value after parameter adjustment;

[0037] (c6) The model inversion effect is fitted, and finally the final optimized AOT value of the target area is output;

[0038] (c7) Through the algorithm, the future AOT value is predicted.

[0039] Further, in step three, the specific implementation steps of the multi-mode model are as follows:

[0040] (d1) Training data generation, obtain various historical prediction data of the multi-mode model from the database to generate training data and validation data, take several days as a cycle, take the data in a cycle as the input training set, and take the data in the next cycle as the validation set;

[0041] (d2) Data validity processing, data normalization processing;

[0042] (d3) Establishing the LSMT model, the first layer of the LSMT model is the LSTM network, and the second layer is the full connection layer; Set the mean square error MSE as the loss function of this neural network, and the calculation method of the loss function is to calculate the square sum of the distance between the predicted value and the true value, and its formula is:

[0043] ;

[0044] Then the adaptive gradient descent algorithm AdaGrad is used as the back propagation iteration, and its updating process is as follows:

[0045] ;

[0046] Wherein s is the accumulation of gradient square, when updating the parameters, the learning rate is divided by the square root of this accumulation plus a very small value, which helps the parameters to move in a direction closer to the bottom of the slope, thereby accelerating the convergence;

[0047] (d4) After the training is completed, the recurrent neural network model is saved, and after 0 o'clock every morning, the model will predict the hourly concentration of particulate matter and ozone in the future period according to the 24-hour monitoring data of the previous day, and output the data of several days;

[0048] (d5) Enter the next period, and the LSTM neural network model is trained again.

[0049] Further, in step one, the GFS weather forecast data, MODIS satellite remote sensing data, air quality monitoring station data, national pollution source emission mesoscale list and target area pollution source emission small scale list are downloaded through the download server.

[0050] Further, the download server uses a distributed parallel operation architecture to process data; the download server includes a master server A, a basic server B and a computing power server resource pool; the master server A is responsible for the management of the entire distributed parallel operation, and the multi-mode air quality model is arranged on the master server A; the master server A and the basic server B provide basic computing power for the multi-mode air quality model, and the basic server B and the servers in the computing power server resource pool provide supercomputing computing power for the multi-mode air quality model; all servers communicate through a dedicated network under the coordination of the master server A.

[0051] Further, in step one, the air quality monitoring data of the target area is obtained by setting the air quality monitoring station, and the wind speed and wind direction data of the target area is obtained by setting the wind speed and wind direction monitoring device;

[0052] The outer side of the body of the wind speed and wind direction monitoring device is provided with a protective frame; the outer side of the protective frame is provided with a cleaning rod; the cleaning rod slides along the outer surface of the protective frame under the driving of a driving part; the protective frame is overall arc-shaped, the driving part is a wind-driven rotating piece, the rotating piece is rotationally matched with a supporting seat below the wind speed and wind direction monitoring device, two wings are arranged at an angle on the rotating piece, the cleaning rod is located between the two wings, and the wings drive the rotating piece to rotate when the wings are wind; the supporting seat is provided with two limiting blocks for limiting the rotation limit position of the rotating piece.

[0053] Beneficial effects: the method for ozone and particulate matter prediction and early warning based on the multi-mode air quality model has the following beneficial effects:

[0054] 1) The ozone and particulate matter are predicted and warned through a multi-mode air quality model, the prediction result is high in accuracy, and the prediction result is more stable;

[0055] 2) The download server with a distributed parallel operation architecture is designed, the supercomputing demand is more easily met, and the timeliness of the prediction and warning can be guaranteed. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification. Fig. 1 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification.

[0057] Fig. 2 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification.

[0058] The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification. Fig. 3 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification.

[0059] Fig. 4 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification.

[0060] The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification. Fig. 5 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification.

[0061] Fig. 6 The accompanying drawings are intended to illustrate the present ozone and particulate matter prediction and warning method, and constitute a part of the present specification. DETAILED DESCRIPTION

[0062] The present ozone and particulate matter prediction and warning method will be further described below in combination with the accompanying drawings.

[0063] As shown in the accompanying drawings, the present ozone and particulate matter prediction and warning method based on a multi-mode air quality model comprises the following steps: Figs. 1 to 6 Step one: long-term historical related data of a target area are acquired, and the related data comprises meteorological prediction data, satellite remote sensing data, pollution source emission inventory and air quality monitoring data;

[0064] Step two: the meteorological prediction data and the pollution source emission inventory are substituted into two third-generation air quality models of CMAQ and CAMx to acquire prediction values of air quality data, meteorological data and rainfall of the target area; the air quality monitoring data are substituted into a particulate matter ozone neural network prediction model to acquire prediction values of particulate matter concentration and ozone concentration of the target area; and the satellite remote sensing data are substituted into a satellite remote sensing aerosol inversion model to acquire an AOT prediction value of the target area;

[0065]

[0066] ​​​​Step three: adopt the model architecture of neural network, on the basis of long-term historical related data, the model prediction value and the actual monitoring value are trained in a cycle, so as to establish a multi-mode air quality model; the steps of establishing a multi-mode model include: (d1) generating training data; (d2) performing validity processing and normalization processing on the data; (d3) establishing an LSMT neural network model, and then adopting an adaptive gradient descent algorithm AdaGrad as a backward transmission iteration; (d4) after the training is completed, the cycle neural network model is saved; the model predicts the hourly concentration of particulate matter and ozone in the future period according to the monitoring data, and extracts some data output; (d5) enter the next cycle, and train the LSTM neural network model again;

[0067] Step four: input the latest related data of the target area into the multi-mode air quality model, and predict and warn the ozone and particulate matter in the target area according to the output result of the multi-mode air quality model.

[0068] In step one, GFS weather forecast data and MODIS satellite remote sensing data are downloaded from the target data server by the download server A and the download server B. Specifically, the latest data is obtained from a specific data category according to the business scene requirement.

[0069] In this step, firstly, the two servers are optimized in the physical line, so as to ensure high speed and stability of connection with overseas network.

[0070] Secondly, considering that some target data servers limit the access number, download speed and download time, the access with large instantaneous download amount is limited, and a set of download mechanism is formulated to avoid being limited by the target data server.

[0071] Thirdly, the two target servers will be synchronized to the content distribution server in real time after completing the download of each data file, and the content distribution server will ensure that the project servers in various places can quickly and stably obtain data through load balancing, content distribution, scheduling and other ways.

[0072] After the download server A and the download server B complete the data upload, the content distribution server will check the data, generate a download completion marker file after judging that all files are complete and available.

[0073] The project server downloads the required files according to the geographical location of the customer, and builds in different areas. After confirming the download completion marker file on the content distribution server, the content distribution server automatically optimizes the transmission path through intelligent virtual network according to network congestion and corresponding speed, so as to ensure fast and stable file transmission.

[0074] All servers are built using Linux operating system, through CSHELL script according to the above data download mechanism to achieve their respective tasks, through the wget command to obtain data from the target data server, through the scp command to transfer files between servers.

[0075] In step one, meteorological prediction data needs to be prepared. After obtaining GFS meteorological prediction data, through the WRF meteorological research and prediction model, input the GFS global meteorological prediction data, calculate the future meteorological prediction results containing China and surrounding areas, provincial surrounding areas, target city surrounding areas, and other multi-level small-scale resolution, get the future meteorological prediction. In specific implementation, usually use 27KM, 9KM, 3KM, 1KM four-layer grid nesting method to gradually obtain high-resolution meteorological simulation values of key areas, and set radiation process parameters, physical process parameters, cumulus convection parameters, and road surface process parameters according to specific implementation requirements, improve the accuracy of prediction by refining underlying surface, LAC (location area code) and other basic data. The prediction length is from 72 hours to 168 hours, which is determined according to the actual project requirements.

[0076] In step one, emission inventory data needs to be prepared. Obtain the national pollution source emission mesoscale inventory and the target area pollution source emission small-scale inventory, calculate the emission inventory data in the prediction and early warning period, which contains emission species, emission time, and emission source identification. In the specific implementation process:

[0077] (e1) Different scales of emission inventory are different in statistical method, time classification, emission type, and data format. According to the project needs, output relatively consistent inventory;

[0078] (e2) When interpolating the large-scale grid emission source inventory to the small grid source inventory, there are multiple interpolation methods, different interpolation methods will affect the accuracy of the model simulation results, need to be tested in the early stage of the project;

[0079] (e3) Emission inventory data is used for CMAQ and CAMx to simulate the environment air, in order to obtain higher accuracy, need to use BVOCs (biogenic volatile organic compound emission model), SMOKE (atmospheric emission source inventory processing model) and other special industry emission calculation models.

[0080] In step one, the air quality historical monitoring data needs to be prepared. The air quality monitoring data can reflect the real-time situation of the air quality in the real environment, generally includes six air quality data of SO2, NOx, CO, O3, PM2.5 and PM10, and four meteorological data of wind speed, wind direction, temperature and humidity, according to different equipment hardware, the data is uploaded in different time periods, and the present application uses air quality monitoring hourly data.

[0081] The server will acquire the data of the air quality monitoring national control station of 336 cities nationwide in real time, the maintenance of the national control station equipment is timely, the data quality control requirement is strict, and it is the most authoritative data reflecting the environmental air quality.

[0082] In the actual implementation process, according to the actual situation of the project, the reliable station data of provincial control, municipal control and the like is selected, and is also included in the multi-mode air quality model calculation. Through the multi-mode air quality model, the two key pollution factors of PM2.5 and ozone in the future air quality are predicted for 72 hours, 48 hours and 24 hours per hour. And the pollution generation mechanism is quantitatively traced back.

[0083] The multi-mode air quality model includes: CMAQ, CAMx, two third-generation air quality models, particulate matter ozone neural network prediction model and aerosol inversion model based on MODIS satellite remote sensing data.

[0084] Among them, CMAQ and CAMx are third-generation air quality models, based on the concept of "one atmosphere", combined with meteorological data, geographical data, pollution source emission data, through modeling of pollution generation and diffusion physical process and chemical process, the air quality of future multiple scale regions is predicted, the prediction accuracy of CMAQ and CAMx models under different meteorological conditions and chemical mechanisms is different, the present application fully combines their respective advantages, for the two kinds of air quality models, two kinds of chemical mechanisms are equipped, and the prediction results are fitted with the true observation results, and finally the ozone and PM2.5 prediction results of each hour in the future 72 hours, 48 hours and 24 hours are output respectively. In order to achieve pollution source tracing, CMAQ-ISAM and CAMx-PSAT two modules are started.

[0085] The third-generation air quality model uses each curved surface grid to simulate the physical and chemical processes of the atmospheric environment, and the continuous change rate of the pollution factor concentration of each grid can be expressed by the gas continuity equation formula:

[0086] ;

[0087] Wherein Ci is the pollution factor concentration, u is the transmission concentration rate of change, K is the diffusion concentration rate of change, E is the emission concentration rate of change, S is the sedimentation concentration rate of change, Rgas is the gaseous reaction concentration rate of change, Rpart is the particulate reaction concentration rate of change, and Rphase is the state item conversion concentration rate of change.

[0088] The air quality model needs to distribute pollutant emission data, meteorological prediction data, geographical environment data and boundary condition data to each grid according to different resolutions in multiple nested layers, and then carries out numerical simulation of future tens to hundreds of hours for each grid through atmospheric physics and chemistry modes and other algorithms, so that the calculation amount and data generation amount are very large.

[0089] The latest versions of international CMAQ and CAMx follow the traditional Fortran implementation mode, without considering computer cluster distribution and the latest efficient parallel computing, the present application carries out secondary development on the parameter adjustment mode, computing power distribution and data storage mechanism of the two air quality models, realizes cluster distributed parallel operation of the two models, and greatly shortens the simulation time.

[0090] In order to improve the accuracy of the air quality model, the output results of the CMAQ and CAMx models are fitted with the actual monitoring data by means of nonlinear least squares fitting, and a fitting function is obtained for further fitting of the prediction results.

[0091] In step two, the data is substituted into the specific implementation steps of the particulate matter ozone neural network prediction model as follows:

[0092] (a1) According to the requirements of each project, the geographical area of model simulation is formulated, and the grid parameters are determined, generally adopting a mode of four-layer grid nesting of 27KM, 9KM, 3KM and 1KM, so as to achieve the purpose of obtaining high-resolution prediction results in the target area.

[0093] After determining the model grid parameters, meteorological prediction data is generated by WRF, and hourly emission inventory data is obtained by emission source list algorithm.

[0094] The emission source data of this step is generally optimized by SMOKE model at 1*1KM accuracy, and high-rise emission pollution sources and key pollution sources are marked.

[0095] In this step, in order to improve the accuracy of meteorological model prediction, land underlying surface replacement, land utilization rate update, target area SMOKE emission inventory making, city canopy simulation and other technologies are adopted to improve the accuracy of various data.

[0096] (a2) input meteorological prediction data and emission inventory data, simulate SO2, NO2, CO, ozone, PM2.5, temperature, humidity, wind speed, wind direction and rainfall of each grid in the key area at each time period through CMAQ and CAMx air quality models;

[0097] At this step, the existing code is mainly decades ago, and with the increase of simulation area and resolution, the calculation amount is multiplied compared with before, and more time is needed to complete several days of simulation.

[0098] The calculation efficiency of CMAQ and CAMx has been tested, and it is found that in parallel operation, the number of CPU cores and the calculation speed are not in linear relationship, and the more the number of CPU cores, the faster the calculation speed, because the third generation air quality model uses relatively old technology in code implementation, so the bottom layer implementation code of CMAQ and CAMx model is optimized: replace the more efficient compilation library, install the distributed parallel operation dependent library; optimize the parallel operation scheduling mechanism; optimize the file output mechanism, so that the model can be used for distributed calculation.

[0099] After optimization, considering the implementation cost, the model calculation efficiency is improved by 4 times on the basis of basic computing power, and the distributed server cluster in the application is connected, which can additionally increase computing power according to the sufficiency of computing power resources to speed up model calculation.

[0100] (a3) through a nonlinear least squares fitting algorithm, the prediction results are optimized and fitted according to the fitting function.

[0101] Nonlinear least squares formula:

[0102] Use high-order polynomial fitting

[0103] ;

[0104] f is the fitting value, p is the prediction value, and in actual operation, n=3 is often set.

[0105] The error is calculated by the least squares method, and the error function is:

[0106] ;

[0107] After machine learning, the final expression of the appropriate f(x) is finally obtained.

[0108] The fitting formula f(x) is used to output the prediction fitting values of the two models CMAQ and CAMX respectively.

[0109] Considering the calculation amount and actual business needs, generally selected future 72 hours of hourly data for output in the specific implementation process.

[0110] Currently, there are two main ways to predict air quality: mechanism-based prediction and non-mechanism-based prediction. CMAQ and CAMx air quality models use mechanism-based prediction, which predicts future air quality by establishing relatively complete pollution source emission inventory, relatively accurate meteorological field conditions, and atmospheric pollutant diffusion processes.

[0111] Non-mechanism-based prediction does not require complex pollutant boundary fields and meteorological boundary fields. It only needs to capture the characteristics of historical pollutant data to obtain the variation law of pollutant concentration.

[0112] PM2.5 and ozone have common precursors, namely nitrogen oxides and volatile organic compounds, and their roles in the formation of particulate matter and ozone are different. Their mutual influence is complex. Using a neural network algorithm based on machine learning, historical air quality data and meteorological data can be trained to predict the concentration of ozone and particulate matter in the future.

[0113] Recurrent neural networks use sequence data as input, all nodes are connected in a chain, and forward recursion is used to learn the nonlinear characteristics of sequences, which is particularly suitable for predicting particulate matter and ozone concentration.

[0114] In step two, the data is substituted into the specific implementation steps of the particulate matter and ozone neural network prediction model as follows:

[0115] (b1) Export all hourly historical data (such as 2 years) of specific regional air quality automatic monitoring stations from the platform database. In this process, data missing or abnormal due to equipment failure, data quality control adjustment, etc. will not be processed for the time being; data preprocessing is performed on historical data, and data cleaning and normalization are performed on data missing, abnormal values, etc. that are not suitable for model training. The normalization equation is:

[0116] ;

[0117] Because the data used by the model comes from environmental air quality monitoring national control sites and provincial control sites, and the operation and maintenance team is equipped 24 hours a year, there will be no continuous multi-day data anomalies or missing data. For data missing more than 8 hours a day, all data for that day is excluded from the training sequence. Considering the most extreme case, even if the data preprocessing data may have missing data for several days, the model accuracy can still be guaranteed.

[0118] (b2) According to the time sequence, generate a training data list and a validation data list. The data structure of the training data.

[0119] (b3) Using recurrent neural network to train the training data, each training input step is 7, which is convenient for simulating the accumulation and dissipation of atmospheric pollutants. The particulate matter ozone neural network prediction model uses a four-layer hidden layer structure, and the output result is the hourly concentration of ozone and particulate matter in the next 7 days.

[0120] (b4) After the training is completed, the recurrent neural network model is saved, and after 0 o'clock in the morning every day, the model will predict the hourly concentration of particulate matter and ozone in the next 7 days according to the 24-hour monitoring data of the previous day.

[0121] (b5) Every 7 days is a cycle, the neural network is retrained to obtain better model prediction accuracy.

[0122] It is one of the research hotspots of environmental remote sensing to obtain regional surface and atmospheric information through satellite remote sensing image inversion technology. AOT (Aerosol Optical Thickness) optical aerosol thickness is the path integral of the aerosol extinction function along the propagation direction from the ground to the top of the atmosphere, which is a physical quantity representing the degree of weakening of solar radiation by aerosol. Domestic and foreign experts and scholars have been exploring the method of using AOT information to predict the concentration of ground atmospheric pollutants, but the prediction effect of satellite remote sensing data alone is not ideal. Considering that the aerosol thickness data has obvious gain for particulate matter and ozone prediction, the aerosol inversion model based on MODIS data is introduced to predict the AOT value of each day in the future.

[0123] The aerosol inversion model of the application mainly uses the simplified deep blue algorithm and the 6S radiation transfer model, the purpose is to provide a strong correlation parameter for the multi-mode model of O3 and particulate matter. Improve the accuracy of neural network fusion of multi-mode model.

[0124] In step two, the data is substituted into the satellite remote sensing aerosol inversion model, and the specific implementation steps are as follows:

[0125] (c1) Download MODIS data from NASA website, considering the factors such as aging of optical devices and sensors, directly download Level 1 data after algorithm calibration and coordinate. According to the time point when the satellite sweeps China, download the relevant data.

[0126] (c2) Invert the apparent reflectivity of the target through the polarization radiation transfer formula, and the polarization radiation transfer formula is:

[0127] ;

[0128] Where R is the apparent reflectance, F0 is the extraterrestrial solar irradiance, I is the top-of-atmosphere radiance, μ is the observation zenith angle cosine of the viewing angle of the satellite, μ0 is the solar zenith angle cosine, and φ is the relative azimuth angle between the scattering radiation propagation analysis and the incident solar direction.

[0129] (c3) According to the target area, the apparent reflectance at 0.412 um wavelength is inverted, and the temperature brightness value inverted according to 11 nm and 12 nm wavelengths and the spatial variation of 1.38 um MODIS reflectance are used to determine the meteorological conditions of the target area, such as cloud layer or ground ice and snow cover, and stop inversion, and use the predicted value of the historical model as the result of the current model operation.

[0130] (c4) The AOT estimated value is obtained by calculating the normalized vegetation index (NDVI) and the normalized ice and snow index (NDSI) and other parameters.

[0131] (c5) By identifying the land types (water, city, vegetation, desert) and aerosol types (dust type, city type, ocean type, and biological combustion type) in the target area grid, the corresponding parameters are adjusted to obtain the AOT simulation value after parameter adjustment.

[0132] (c6) According to the data of 60 stations in China of AERONET, the inversion effect of the model is fitted, and finally the optimized AOT value of the target area is output.

[0133] (c7) Through the algorithm, the daily value in the next 3 days is predicted.

[0134] The multi-mode air quality model includes four independent models, each of which has its own strengths in simulating particulate matter and ozone. Among them, the CMAQ and CAMx models can predict all parameters of future air quality and weather, the particulate matter and ozone neural network prediction model has high prediction accuracy for ozone and PM2.5, and the aerosol inversion model can provide the important index of AOT. The multi-mode air quality model of the present application adopts the model architecture of LSTM neural network, and on the basis of long-term historical data, the model prediction value and the actual monitoring value are cyclically trained. In addition to being able to remember the long-term rules of air pollution in the target area, it can also analyze the short-term characteristics by integrating the air quality of the target area in the recent 7 days.

[0135] In step three, the specific implementation steps of the multi-mode model are as follows:

[0136] (d1) Training data generation, obtain various historical prediction data of the multi-mode model through the database, pay attention to the historical monitoring data of the site, generate training data and verification data. Seven days of 24-hour data are used as input training set, and future 7 days of data are used as verification set.

[0137] (d2) data validity processing, data normalization processing.

[0138] (d3) establishing an LSMT model, the first layer being an LSTM network, the second layer being a full connection layer, setting mean square error (MSE) as the loss function of the neural network, and the formula of the loss function being:

[0139] ;

[0140] The calculation method is to calculate the sum of squares of distances between the predicted value and the true value;

[0141] An adaptive gradient descent algorithm (AdaGrad) is used as the back propagation iteration: the updating process is as follows:

[0142] ;

[0143] Where s is the accumulation of the square of the gradient, and when updating the parameters, the learning rate is divided by the square root of the accumulation plus a very small value, which is conducive to moving the parameters in the direction closer to the bottom of the slope, thereby accelerating the convergence.

[0144] (d4) after the training is completed, the recurrent neural network model is saved, and after 0 o'clock in the morning every day, the model will predict the hourly concentration of particulate matter and ozone in the next 7 days according to the 24-hour monitoring data of the previous day, and output the data of 3 days.

[0145] (d5) every 7 days is a cycle, and the LSTM neural network model is trained again.

[0146] The multi-mode air quality model integrates four prediction models, and requires a large amount of data processing and numerical simulation calculation. In order to ensure the timeliness of the model prediction and the guiding nature of the daily environmental management, a distributed parallel operation architecture specially used for scientific calculation is designed.

[0147] The multi-mode air quality model is arranged on the main server A, which is responsible for the management of the entire distributed parallel operation, including task allocation, sub-server communication coordination, memory management and other functions. The debugging operation of the entire model is carried out on the main server A.

[0148] The basic server B and the servers in the computing resource pool provide supercomputing power for the multi-mode air quality model, and all the servers communicate through a dedicated network under the coordination of the main server.

[0149] The main server A and the basic server B provide basic computing power for the multi-mode air quality model, and at the same time, by evaluating the load rate of the computing resource pool, additional computing resources are loaded under the condition of sufficient computing power.

[0150] In step one, the air quality monitoring data of the target area is obtained by setting air quality monitoring sites, and the wind speed and direction data of the target area is obtained by setting wind speed and direction monitoring devices.

[0151] The conventional wind speed and direction monitoring device is difficult to maintain in use, and is easily affected by external debris in the long-term use process, such as a large amount of fallen leaves accumulated and attached, so that the wind speed and direction monitoring device cannot rotate, thereby affecting the reliability of the monitoring results of the wind speed and direction monitoring device. If only relying on manual cleaning of debris by maintenance personnel, the maintenance cycle is long, and it is difficult to clean the debris in time. Therefore, an automatic cleaning mechanism is added to the wind speed and direction monitoring device.

[0152] The outer side of the body of the wind speed and direction monitoring device is provided with a protective frame 1; the protective frame 1 can block foreign matter; the outer side of the protective frame 1 is provided with a cleaning rod 2; the cleaning rod 2 slides along the outer surface of the protective frame 1 under the driving of a driving component; thereby cleaning the foreign matter attached to the surface of the protective frame 1.

[0153] The driving component can be electrically driven, and a positioning module is arranged on the motor to drive the cleaning rod 2 to rotate back and forth one or more times every interval time.

[0154] The driving component can also be wind-driven. A rotating member is arranged on the support seat 6, and a wind-receiving structure is arranged on the rotating member. The wind-receiving structure drives the rotating member to rotate under the action of wind. The rotating member drives the cleaning rod 2 to slide along the outer surface of the protective frame 1 when rotating. A limiting member is further arranged on the support seat 6 to limit the limit position of the rotating member. When the cleaning rod 2 slides to one side of the protective frame 1 along the outer surface of the protective frame 1, it is blocked by the limiting member to prevent the rotating member from continuing to rotate. When there is an opposite wind force, the cleaning rod 2 slides to the other side of the protective frame 1 along the outer surface of the protective frame 1, so that the cleaning rod 2 can slide back and forth along the outer surface of the protective frame 1.

[0155] The wind-receiving structure includes two angle-shaped wings 3, and the included angle of the two wings 3 is less than or equal to 90 degrees. The cleaning rod 2 is located in the middle plane of the two wings 3, which is beneficial to the back-and-forth sliding of the cleaning rod 2 along the outer surface of the protective frame 1.

[0156] Two ends of the cleaning rod 2 are respectively provided with connecting rods 4, the connecting rods 4 are vertical rods, the upper end of the connecting rod 4 is connected with one end of the cleaning rod 2, and the lower end of the connecting rod 4 is provided with a sleeve 5. The support seat 6 is provided with a plug shaft 7 which is rotationally connected with the sleeve 5. The two wings 3 are arranged on one of the sleeves 5, and the structure is more compact and reasonable.

[0157] The limiting member comprises two limiting blocks symmetrically arranged, which are respectively located at two sides of the connecting rod 4 and are used for limiting the movement position of the wing piece 3. When the lower plate surface of one wing piece 3 is in contact with the limiting block, the other wing piece 3 is in a vertical state, so that the wing piece 3 is more easily driven to rotate the whole rotating member when subjected to the wind force.

[0158] The protective frame 1 is in a whole arc shape and is composed of a plurality of arc-shaped rods arranged at intervals. The lower part of the cleaning rod 2 is provided with a plurality of tooth blocks, which are located in the gaps between adjacent two arc-shaped rods. The width of the tooth block is smaller than the interval of the adjacent arc-shaped rods. The arrangement of the tooth block can enhance the cleaning ability of the cleaning rod 2.

[0159] The two sides of the support seat 6 are provided with tooth-shaped grooves for accommodating the tooth blocks. After the tooth blocks move to the tooth-shaped grooves along the gaps between the arc-shaped rods, it is easier to scrape off the foreign matters from the outer surface of the protective frame 1, and the cleaning effect of the cleaning rod 2 is better.

[0160] The middle lower part of the arc-shaped rod is provided with a reinforcing frame composed of two vertical rods and a horizontal rod at the upper end of the two vertical rods. The middle lower part of the arc-shaped rod is fixedly connected with the horizontal rod through the vertical reinforcing rod. The length of the reinforcing rod is greater than the length of the tooth block, so as to avoid the movement interference between the tooth block and the horizontal rod. The arrangement of the reinforcing frame can improve the structural strength of the protective frame 1.

[0161] The upper surface of the support seat 6 is arc-shaped, and the foreign matters falling on the upper surface of the support seat 6 are more easily slipped off.

[0162] When the driving component is driven by the wind, due to the structure of the rotating member, the rotating member always inclines to one side in the balanced state. At this time, one wing piece 3 is vertical, and the other wing piece 3 is horizontal. The vertical wing piece 3 is driven to overturn the rotating member after being subjected to the wind force, so as to drive the cleaning rod 2 to slide along the surface of the protective frame 1 in the forward direction to clean the foreign matters, until the original vertical wing piece 3 becomes horizontal and the original horizontal wing piece 3 becomes vertical, and the balance is re-established. When the wind direction changes and the opposite wind force is generated, the wing piece 3 drives the cleaning rod 2 to slide along the surface of the protective frame 1 in the reverse direction. Thus, the cleaning rod 2 can reciprocate along the surface of the protective frame 1 under the action of the wind force, so as to clean the foreign matters. The cleaning rod can clean most of the foreign matters in time, but in actual use, there may be a small part of the foreign matters that cannot be cleaned. The maintenance personnel can completely remove the foreign matters during maintenance by using the wind speed and direction monitoring device.

[0163] The above only describes the preferred embodiments of the present application. It should be noted that those skilled in the art can make some improvements and refinements without departing from the principles of the present application. These improvements and refinements should also be considered within the protection scope of the present application.

Claims

1. A method for ozone and particulate matter prediction and early warning based on a multi-model air quality model, characterized in that: The method includes the following steps: Step 1: Obtain long-term historical relevant data for the target area, including meteorological forecast data, satellite remote sensing data, pollution source emission inventories, and air quality monitoring data; Step 2: Substitute meteorological forecast data and pollution source emission inventories into the CMAQ and CAMx third-generation air quality models to obtain air quality data, meteorological data, and predicted rainfall values ​​for the target area; substitute air quality monitoring data into the particulate matter and ozone neural network prediction model to obtain predicted particulate matter and ozone concentration values ​​for the target area; substitute satellite remote sensing data into the satellite remote sensing aerosol inversion model to obtain predicted AOT values ​​for the target area. Step 3: Using a neural network model architecture, based on long-term historical data, the model's predicted values ​​and actual monitoring values ​​are iteratively trained to establish a multi-mode air quality model. The steps for establishing a multi-mode model include: (d1) generating training data; (d2) performing validity processing and normalization on the data; (d3) establishing an LSTM neural network model, and then using the adaptive gradient descent algorithm AdaGrad as the backpropagation iteration; (d4) after training, saving this recurrent neural network model; based on the monitoring data, the model predicts the hourly concentrations of particulate matter and ozone for the next cycle, and extracts data from several days for output; (d5) entering the next cycle, and retraining the LSTM neural network model. Step 4: Input the latest relevant data of the target area into the multi-model air quality model, and predict and warn of ozone and particulate matter in the target area based on the output of the multi-model air quality model.

2. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 1, characterized in that: In step two, the specific implementation steps for substituting the data into the two third-generation air quality models, CMAQ and CAMx, are as follows: (a1) Define the geographical area to be simulated by the model, determine the grid parameters, generate meteorological forecast data through the WRF meteorological research and forecast model, and obtain hourly emission inventory data through the emission source inventory algorithm; (a2) Input meteorological forecast data and emission inventory data, and simulate SO2, NO2, CO, ozone, PM2.5, temperature, humidity, wind speed, wind direction and rainfall for each grid in the key area at each time period using the CMAQ and CAMx air quality models respectively; (a3) The prediction results are optimized and fitted using a nonlinear least squares fitting algorithm based on the fitting function; the formula for fitting the nonlinear least squares formula using a high-order polynomial is as follows: f(x)≈p n (x)=a0+a1(x-x0)+a2(x-x0) 2 +…+a n (x-x0) n ; In the formula, f is the fitted value and p is the predicted value; The error is calculated using the least squares method, and the error function is: Machine learning is used to obtain a suitable final expression for f(x); then, the fitting formula f(x) is used to output the predicted fitting values ​​of the CMAQ and CAMX models respectively.

3. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 2, characterized in that: In step two, the specific implementation steps for substituting the data into the particulate matter ozone neural network prediction model are as follows: (b1) Export hourly historical data of all automatic air quality monitoring stations in a specific region from the platform database; (b2) Generate a training data list and a validation data list according to the time sequence; (b3) Use a recurrent neural network to train the training data, taking several days as a period, and output the hourly concentrations of ozone and particulate matter for several days within a period: (b4) After training, save this recurrent neural network model; after 0:00 every day, the model will predict the hourly concentrations of particulate matter and ozone for several days in the future cycle based on the 24-hour monitoring data of the previous day. (b5) After entering a cycle, the neural network is retrained to obtain better model prediction accuracy.

4. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 3, characterized in that: In step (b1), historical data that is missing, numerically abnormal, or unsuitable for model training is cleaned and normalized; the normalization equation is:

5. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 1, characterized in that: In step two, the specific implementation steps for substituting the data into the satellite remote sensing aerosol inversion model for prediction are as follows: (c1) Download MODIS data from the NASA website; (c2) The apparent reflectivity of the target is inverted using the polarization radiative transfer formula, which is: Where R is apparent reflectance, F0 is the extraterrestrial solar radiation flux, I is the radiance at the top of the atmosphere, μ is the cosine of the observed zenith angle from the satellite's perspective, μ0 is the cosine of the solar zenith angle, and φ is the relative directional angle between the scattered radiation propagation analysis and the incident solar direction. (c3) Based on the target area, the reflectance at a wavelength of 0.412um is retrieved. At the same time, based on the temperature brightness values ​​retrieved at wavelengths of 11nm and 12nm, and the spatial variation of the MODIS reflectance at 1.38um, the meteorological conditions of the target area are determined. If there are clouds or snow and ice covering the ground, the retrieval is stopped, and the historical model prediction values ​​are used as the results of this model run. (c4) By calculating the normalized vegetation index and normalized snow and ice index parameters, the preliminary AOT estimate was obtained; (c5) By identifying the land type and aerosol type in the target area grid, the corresponding parameters are adjusted to obtain the AOT simulation value after parameter adjustment; (c6) Fit the model inversion effect and finally output the final optimized AOT value of the target region; (c7) Predict future AOT values ​​using algorithms.

6. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 1, characterized in that: In step three, the specific implementation steps for establishing the multi-modal model are as follows: (d1) Training data generation: Obtain various historical prediction data of the multi-mode model from the database to generate training data and validation data. Take several days as a cycle, use the data in one cycle as the input training set, and use the data in the next cycle as the validation set. (d2) Data validation and data normalization; (d3) Establish an LSMT model. The first layer of the LSMT model is an LSTM network, and the second layer is a fully connected layer. Set the mean squared error (MSE) as the loss function for this neural network. The loss function is calculated by summing the squares of the distances between the predicted and true values. The formula is as follows: Then, the adaptive gradient descent algorithm AdaGrad is used for the backpropagation iteration, and its update process is as follows: Where s is the accumulated amount of the squared gradient. When updating parameters, the learning rate is divided by the square root of this accumulated amount and then a small value is added. This helps the parameters move closer to the bottom of the slope, thereby accelerating convergence. (d4) After training, save this recurrent neural network model. Every day after 0:00, the model will predict the hourly concentration of particulate matter and ozone for the next cycle based on the 24-hour monitoring data of the previous day, and extract data from several days for output. (d5) Enter the next cycle and retrain the LSTM neural network model.

7. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 1, characterized in that: In step one, GFS meteorological forecast data, MODIS satellite remote sensing data, air quality monitoring station data, national mesoscale pollution source emission inventory, and target area pollution source emission inventory are downloaded via the download server.

8. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 7, characterized in that: The download server uses a distributed parallel computing architecture to process data; the download server includes a main server A, a basic server B, and a computing power server resource pool; the main server A is responsible for the management of the entire distributed parallel computing, and the multi-mode air quality model is deployed on the main server A; the main server A and the basic server B provide basic computing power for the multi-mode air quality model, and the basic server B and the servers in the computing power server resource pool provide supercomputing power for the multi-mode air quality model; All servers communicate via a dedicated network under the coordination of master server A.

9. The method for ozone and particulate matter prediction and early warning based on a multi-mode air quality model according to claim 1, characterized in that: In step one, air quality monitoring data of the target area is obtained by setting up air quality monitoring stations, and wind speed and direction data of the target area are obtained by setting up wind speed and direction monitoring devices. A protective frame (1) is provided on the outside of the main body of the wind speed and direction monitoring device; a cleaning rod (2) is provided on the outside of the protective frame (1); the cleaning rod (2) slides along the outer surface of the protective frame (1) under the drive of the driving component; the protective frame (1) is generally arc-shaped, the driving component is a wind-driven rotating component, the rotating component is rotatably engaged with the support base (6) below the wind speed and direction monitoring device, two blades (3) are provided at an angle on the rotating component, the cleaning rod (2) is located between the two blades (3), and the blades (3) drive the rotating component to rotate when exposed to wind; two limiting blocks (8) are provided on the support base (6) to limit the rotation limit position of the rotating component.

Images