Air quality model coupled recursive spatiotemporal deep learning PM 2.5 Hourly concentration prediction method

Abstract

The application discloses an air quality mode coupled recursive space-time deep learning PM 2.5 hour concentration prediction method, which can provide relatively accurate PM 2.5 hour concentration prediction value, comprising: constructing a data set; performing feature screening on meteorological factors; constructing an eRSDL model and training the eRSDL model by using the training set, wherein the eRSDL model comprises a sliding average model, a data fusion model and a recursive space-time deep learning model RSDL of 24-hour scale; using the trained eRSDL model to perform regional PM 2.5 concentration prediction, and obtaining PM 2.5 concentration prediction result. The eRSDL method for input data optimization by adding a sliding average model and a data fusion method on the basis of the traditional RSDL is proposed, the prediction quality of WRF-CMAQ is improved before RSDL deviation correction, the prediction deviation can be kept in a stable range, and the PM 2.5 hourly concentration prediction process has better adaptability to training data and long-time sequence rule mining capacity, so that the PM 2.5 hourly concentration prediction accuracy is effectively improved.

Description

Technical Field

[0001] This invention belongs to PM 2.5 In the field of concentration analysis technology, and more specifically, involving air quality models coupled with recursive spatiotemporal deep learning PM... 2.5 Hourly concentration forecasting method. Background Technology

[0002] There are many current air quality early warning and forecasting systems, such as a coupled NO2 forecasting method combining air quality simulation and observational machine learning (application number 2023103946762) and an ozone hourly concentration forecasting method integrating recursive spatiotemporal learning and simulation monitoring (application number 2023107680202), which provide accurate forecasting models for NO2 and O3, respectively. However, existing air quality early warning and forecasting systems are less accurate for PM2.5. 2.5 Current forecasting models rely solely on numerical models, statistical models, or machine learning methods, which can lead to significant deviations in air quality forecasts under uncertain meteorological conditions and extreme weather events. Most existing air pollutant concentration forecasts depend on subsequent manual corrections; numerical models and machine learning alone cannot provide accurate concentration forecasts.

[0003] Accurate and stable air quality forecasts provide crucial scientific basis for relevant departments to formulate governance plans and are a key technology for monitoring and early warning systems. Accurate PM2.5 data is essential. 2.5 Forecasts can help governments predict PM levels in advance. 2.5 During periods of high pollution, measures such as increasing supervision of industrial production and construction sites, improving traffic flow, increasing the frequency of road watering, and activating spray systems can be used to mitigate PM2.5 pollution. 2.5 To pollute or reduce the degree of pollution.

[0004] Current PM 2.5 Forecasting technologies are mainly divided into: ① Numerical model forecasting: based on known patterns, the corresponding numerical models are converted to predict future air quality; ② Statistical model forecasting: forecasting models based on statistical methods using data such as weather and pollutant concentration monitoring; ③ Numerical-statistical coupled forecasting: based on numerical model forecasting, statistical methods are used as preprocessing or post-processing to optimize the input or output.

[0005] Current numerical-statistical coupled forecasting methods mostly employ artificial neural networks for forecasting. However, when historical data is insufficient or the accuracy of input data is low, the sensitivity of neural networks to input errors and their lack of interpretability are problems that neural network methods themselves cannot solve. In terms of interpretability and stability, they are also inferior to traditional time series models and regression models. Summary of the Invention

[0006] To at least solve one of the problems existing in the prior art, the present application provides a PM 2.5 Concentration prediction method, the data of the current numerical-statistical coupling method is optimized and expanded, the eRSDL method of adding a moving average model and a data fusion method to optimize the input data of RSDL is proposed on the basis of the traditional RSDL, has higher training data adaptability and long time sequence rule mining ability; meanwhile, the present application can make up for the problem of insufficient periodic change ability of the traditional RSDL simulation, keep the prediction deviation in a stable range, and effectively improve the PM 2.5 Hourly concentration prediction accuracy.

[0007] To achieve the purpose of the present application, the present application provides a PM 2.5 Concentration prediction method, comprising the following steps:

[0008] A PM 2.5 Monitoring, simulation concentration data and meteorological data set, the data set is divided into training set and test set and the data set is pretreated;

[0009] The meteorological factors are selected, and the simulation meteorological factors directly related to the PM 2.5 Concentration prediction is excluded.

[0010] The eRSDL model is constructed, and the training set is used to train the eRSDL model, the eRSDL model includes a moving average model, a data fusion model and a 24-hour scale recursive spatio-temporal deep learning model RSDL, the moving average model is used to combine the PM 2.5 Monitoring data to obtain the preliminary PM 2.5 Concentration prediction value, the preliminary PM 2.5 Concentration prediction value and the PM 2.5 Concentration value simulated by WRF-CMAQ is linearly regressed to obtain the future concentration prediction value of each monitoring site PM 2.5 Based on the latest monitoring data fitting; the data fusion model is used to combine the future concentration prediction value of each monitoring site PM 2.5 Based on the latest monitoring data fitting and the PM 2.5 Concentration value simulated by WRF-CMAQ is eVNA inverse distance weighted data fusion to obtain the further corrected PM 2.5 Simulation concentration value; the 24-hour scale recursive spatio-temporal deep learning model RSDL is used for PM 2.5 Hourly concentration prediction;

[0011] The eRSDL model trained is used to carry out regional PM 2.5Concentration forecast, obtaining PM 2.5 Concentration forecast results.

[0012] Based on further revised PM 2.5 Simulated values, WRF-CMAQ meteorological simulation values, and PM2.5 2.5 Using monitored values ​​(pre-reporting data as model input and subsequent data as prediction target values), a 24-hour recursive spatiotemporal deep learning (RSDL) network is built. This network can obtain PM levels for the next 24 hours. 2.5 Forecast values. The established 24-hour eRSDL model is extended using a recursive model to predict PM2.5 at multiple sites over the next 72 hours. 2.5 concentration.

[0013] The preprocessing includes any one or more of the following: outlier detection, outlier handling, missing value handling, and dimensionless processing of the dataset. The dimensionless processing includes any one or more of z-score standardization and min-max normalization.

[0014] The steps for determining outliers include:

[0015] Calculate PM 2.5 The mean, standard deviation, and variance of the concentration monitoring data were used. The rule determines the range of outliers, which is located within... Data outside of these categories is considered to be outliers. It is variance.

[0016] The steps for handling missing values ​​include:

[0017] Determine the data column containing the missing values, and record the position and value of the known data points in that column;

[0018] Sort the known data points;

[0019] Find all missing values ​​or the two data points before and after a missing data segment, and use linear interpolation to calculate the missing values.

[0020] The z-score standardization:

[0021] Calculate the sample mean with standard deviation , , , This is the original data;

[0022] Perform z-score standardization on the individual data. For each input variable in the sequence... Standardized data can be obtained using the following formula. , .

[0023] The min-max normalization is:

[0024] The min-max normalization formula is: Wherein X is the original data, Y is the normalized data, and inf and sup are the lower and upper bounds of the data set.

[0025] Further, the meteorological factors are screened by correlation analysis and principal component analysis, including the steps of:

[0026] The principal component analysis is performed on the 24-hour value of each meteorological factor in the input set by using the singular value decomposition algorithm, and the principal components of the input set with high correlation are screened out by combining correlation, and then the correlation between each meteorological component and PM 2.5 Concentration is analyzed by using Spearman correlation analysis, and then the principal component analysis of cross-meteorological elements is performed to extract the meteorological elements related to each other, and the required meteorological factors are screened.

[0027] The sliding average regression model is constructed.

[0028] Considering that the WRF-CMAQ simulation result has certain reference value on the trend change of PM 2.5 , the WRF-CMAQ simulated PM 2.5 concentration variable is additionally added to the original sliding average model, and the model formula is as follows:

[0029] ;

[0030] Wherein is the simulation prediction result of PM 2.5 concentration by the WRF-CMAQ system in the current prediction hour node.

[0031] Further, when the eVNA inverse distance weighted data fusion is performed, the interpolation formula is:

[0032]

[0033]

[0034] Wherein, is the inverse distance square weight between the monitoring point and the model prediction area E, is the PM 2.5 future concentration prediction value of the monitoring point on the monitoring point after preliminary optimization by the sliding average model, and are the model prediction area and the monitoring point WRF-CMAQ simulated PM of the small range area of the monitoring point i 2.5 concentration value, and are the horizontal and vertical distances between the monitoring point i and the center of the small range area E of the target, respectively, is the square inverse.

[0035] Further, the recursive spatiotemporal sensitive deep learning model RSDL includes an encoder and a decoder, the encoder uses a down-sampling method to convert the feature map of high resolution-coarse particulate matter into a feature map of low resolution-fine particulate matter, and the decoder uses an up-sampling method to convert the feature map of low resolution-fine particulate matter into a feature map of high resolution-coarse particulate matter.

[0036] Further, in the recursive spatiotemporal sensitive deep learning model RSDL, the encoder-decoder function is:

[0037] Encoder:

[0038] Decoder: ;

[0039] is the input value, which is a four-dimensional array containing time, independent variable number, grid row, and grid column: ; is the input PM 2.5 monitoring value, ; is the hidden layer, which is a four-dimensional array containing time, channel number, grid row, and grid column: ; is the output layer, which is a three-dimensional array containing time, grid row, and grid column: ; is the time number, representing the reporting time and all hour points within the next 24 hours; is the index number of meteorological and PM 2.5 simulation data, is the index summary; is the row and column number of the grid, and are the width and height of the grid domain; is the input value of the first index of row, column, ; is the PM 2.5 monitoring concentration value at ; is the predicted PM2.5 concentration; is the process output data of the encoder; is the channel number, is the number of channels; and respectively represent the encoder and decoder processes.

[0040] Further, the encoder comprises three layers of neural network modules, each of which comprises a convolutional gated recurrent unit network module and a convolutional module, leaky ReLU is used as an activation function between layers, and a Dropout function with a preset probability is added in the encoder to randomly design part of the hidden layer input;

[0041] The decoder comprises three layers of neural network modules, and uses transposed convolution to up-sample the grid data.

[0042] Further, the established 24-hour scale eRSDL model can be expanded to a PM 2.5 concentration prediction for more than 24 hours in the future by using a recursive mode.

[0043] Further, the established 24-hour scale eRSDL model can be expanded to a PM 2.5 concentration prediction for 72 hours in the future by using a recursive mode. When making a prediction, it is divided into three 24-hour prediction stages, the previous stage provides the input required for the moving average model and the initial concentration input of the eRSDL model for the next stage, and when predicting the PM 2.5 concentration for 0-23 hours, the PM 2.5 concentration monitored before the report is used as the initial value of the eRSDL model; when predicting the PM 2.5 concentration for 24-47 hours and 48-71 hours, the initial values are the 0-23 hour concentration value predicted by the eRSDL model and the 24-47 hour concentration value predicted by the eRSDL model, respectively.

[0044] Compared with the prior art, the present application can at least achieve the following beneficial effects:

[0045] (1) The present application provides a novel PM 2.5 concentration prediction method based on recursive spatiotemporal deep learning, which uses correlation analysis and principal component analysis methods to study the degree of influence of meteorological conditions on PM 2.5 concentration, which is crucial for the establishment and analysis of the prediction model.

[0046] (2) By inputting the time series of PM 2.5 concentration values monitored at the monitoring point, the MA model preliminarily corrects the PM 2.5The simulated value, the latest PM of each monitoring point based on the monitoring data fitting 2.5 The future concentration prediction value. The PM of each monitoring point is obtained by the MA model 2.5 The future concentration prediction value and the PM simulated by CMAQ 2.5 The concentration simulation value is inversely weighted and fused by using the eVNA data fusion method, and the PM of the whole simulation domain is obtained 2.5 The concentration simulation value is further modified to obtain the PM of the recursive spatiotemporal convolution network 2.5 The simulated data is used for the purpose of preliminary optimization.

[0047] (3) The PM of the simulation domain is modified based on the MA model and the eVNA data fusion method 2.5 The concentration simulation value, and the recursive spatiotemporal deep learning (RSDL) model is established to cooperate with the RSDL multi-site PM 2.5 The prediction result and the regional prediction result of WRF-CMAQ constitute an enhanced RSDL model (eRSDL), and the regional PM 2.5 Concentration prediction is realized, and the prediction accuracy of the prediction region is improved. 2.5 Concentration prediction accuracy.

[0048] (4) Since the prediction model result has complex regional transportation and historical accumulation characteristics, it is not discrete independent data, and the eRSDL model is more suitable for capturing and capturing the spatiotemporal relationship characteristics of the numerical prediction result than the traditional bias correction technology (such as long short-term memory recursive neural network LSTM and convolutional neural network GRU).

[0049] (5) The present application can use the recursive mode to expand the 24-hour scale eRSDL model prediction result to 72 hours in the future, and realize the PM 2.5 Concentration prediction, the prediction effect is improved in general, which is characterized by high prediction accuracy, small maximum relative error, and certain robustness.

[0050] (6) Compared with the traditional numerical model prediction bias correction method, the secondary prediction model can make full use of the numerical prediction data, reduce the prediction error, and correct the region with monitoring stations, realize the regional overall air quality prediction, and better serve the prediction and control of environmental PM 2.5 Concentration, so as to improve the environmental quality.

[0051] (7) Time series model and regression model are also a low complexity method, which can not only be used as a data preprocessing process to reduce sample noise, but also can determine a reasonable prediction result range under a certain confidence by parameter hypothesis test analysis. The present application optimizes and extends the data of the current numerical-statistical coupling method, reduces the data quality and quantity requirements of the model, and fully utilizes the accuracy advantage of the neural network model to achieve the purpose of fast and accurate prediction of PM 2.5 . BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0053] Figure 1 The encoder and decoder schematic diagram of the recursive spatiotemporal sensitive deep learning. 2.5 The step flow chart of the air quality model coupled recursive spatiotemporal deep learning PM

[0054] Figure 2 The encoder and decoder schematic diagram of the recursive spatiotemporal sensitive deep learning.

[0055] Figure 3 The multi-layer encoder and decoder structure schematic diagram. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Referring to Figure 1 , the present application provides an air quality model coupled recursive spatiotemporal deep learning PM 2.5 hour concentration prediction method, comprising the following steps:

[0058] Step 1: Prepare the numerical model data and PM 2.5 concentration monitoring data of the station to be predicted in the prediction period, and PM 2.5 simulation concentration data as a data set and pre-process the data in the data set, and divide the data set into a training set and a test set. The training set is used to train the model parameters, and the validation set is used to verify the accuracy of the model parameters.

[0059] The hourly forecast values of the meteorological indexes are obtained from the WRF model, the hourly concentration forecast values of the pollutants are obtained from the CMAQ model, and the PM 2.5 concentration monitoring data are obtained from the China Environmental Monitoring Center.

[0060] In this step, the numerical model data in the prediction period are the simulation values of the meteorological factors and the PM 2.5 concentration simulation values in a relatively long time period, and the time resolution scale of both data is hourly data.

[0061] In some embodiments of the present application, the numerical model data are provided by the meteorological simulation of the mesoscale meteorological model WRF in version 3.9.1, and are input into the CMAQ air quality model as one of the elements. The meteorological simulation data provided by the mesoscale meteorological model WRF and the air quality model CMAQ in version 5.3 are used to simulate the PM 2.5 concentration.

[0062] Table 1 WRF-CMAQ output fields and meanings

[0063]

[0064] The preprocessing includes but is not limited to outlier determination, outlier processing, missing value processing, and dimensionless processing, and the dimensionless processing includes z-score standardization and min-max normalization processing.

[0065] In some embodiments of the present application, the outliers of the data set are determined by the method, and the processing steps are as follows:

[0066] S101. Calculate the mean value and the standard deviation of the monitoring data of the PM 2.5 concentration , is a loop parameter, is the value of the monitoring data, is the number of the monitoring data.

[0067] S102. Determine the range of outliers, according to the method, the data outside the range are considered as possible outliers; is the variance.

[0068] S103. Further determine the outliers based on the range of outliers, and according to the data verification experience, the verification is performed.

[0069] In some embodiments of the present application, the missing values of the data set are interpolated and completed, and the processing steps are as follows: ​​

[0070] S104. Determine the data column where the missing value is located, record the position and value of the known data points in this column;

[0071] S105. Sort the known data points to ensure that they are arranged in order in the data column;

[0072] S106. Find the two nearest data points before and after all missing values or missing data segments, such as missing values are or ;

[0073] S107. Use linear interpolation formula for interpolation calculation, that is , .

[0074] In some embodiments of the present application, the data set is z-score standardized:

[0075] S108. Calculate the mean and standard deviation of the selected data set, , , , , , .

[0076] In some embodiments of the present application, the data set is min-max normalized:

[0077] min-max normalization formula: , where is the original data, is the normalized data, and inf and sup are the lower and upper bounds of the data set.

[0078] Table 2: final selected dimensionless method

[0079]

[0080] Step 2: Feature selection of meteorological factors by correlation analysis and principal component analysis, select the simulated meteorological factors that have a greater impact on PM 2.5 concentration prediction, and exclude indirect related effects.

[0081] In this step, the correlation analysis and principal component analysis method is used to perform principal component analysis on the 24-hour value of each meteorological data set, and the input set principal components with high correlation are screened out by combining correlation analysis. Among them, after extracting the hourly value principal component, the Spearman correlation analysis is used to analyze the correlation between each meteorological component and PM 2.5 concentration, as shown in Table 3. Then, principal component analysis across meteorological elements is performed to extract meteorological factors that are correlated with each other, and the dimension of the input set is reduced to the order of magnitude. Finally, the simulation meteorological factors input into the eRSDL are determined. According to the correlation analysis results, the directly related index items are selected, and the indirect correlation influence is excluded.

[0082] Table 3 Correlation between selected meteorological indicators and PM 2.5 concentration

[0083]

[0084] Step 3: Constructing an eRSDL model and training the eRSDL model using the training set, the eRSDL model including a moving average model, a data fusion model, and a 24-hour scale recursive spatiotemporal deep learning model RSDL, the moving average model being used to combine the PM 2.5 monitoring data to obtain a preliminary PM 2.5 concentration prediction value, the preliminary PM 2.5 concentration prediction value being linearly regressed with the PM 2.5 concentration value simulated by WRF-CMAQ to obtain a future PM 2.5 concentration prediction value of each monitoring site based on the latest monitoring data fitting; the data fusion model being used to perform data fusion on the future PM 2.5 concentration prediction value of each monitoring site based on the latest monitoring data fitting and the PM 2.5 concentration value simulated by WRF-CMAQ to obtain a further corrected PM 2.5 simulation value; and the 24-hour scale recursive spatiotemporal deep learning model RSDL being used to perform PM 2.5 hourly concentration prediction.

[0085] This step includes the following sub-steps:

[0086] S301. Constructing a moving average model (MA model) based on the enumeration method and PM 2.5 monitoring data, and then obtaining a preliminary PM 2.5 concentration prediction value based on the MA model and PM 2.5 monitoring data, and then obtaining a preliminary PM 2.5 concentration prediction value based on the MA model and PM2.5 The concentration values are linearly regressed to obtain the future concentration prediction of each monitoring site PM 2.5 based on the latest monitoring data fitted.

[0087] S3.1.1: Construct a moving average model based on historical data, expressed as

[0088]

[0089] is a polynomial function of , , is the coefficient of the polynomial function, is the random error, is the size of the moving average regression window, is the value of the current requirement, is the value obtained by the moving average formula;

[0090] S302. On the basis of step S3.1.1, add the WRF-CMAQ simulated PM 2.5 concentration value variable, and the improved moving average model is

[0091]

[0092] wherein is the assignment weight of the MA model to the WRF-CMAQ simulated PM 2.5 concentration value variable, is the random error, is the simulated value of WRF-CMAQ to PM 2.5 concentration in the current forecast hour node.

[0093] In some embodiments of the present application, the parameters of the moving average model are set as shown in Table 4.

[0094] Table 4 Moving average parameters

[0095]

[0096] Step 3.2: Perform eVNA inverse distance weighted data fusion on the future concentration prediction of each monitoring site PM 2.5 based on the latest monitoring data fitted in step 3.1 and the WRF-CMAQ simulated PM 2.5 concentration value to obtain a further corrected PM 2.5 simulation value.

[0097] ​​​In this step, the eVNA inverse distance weighting data fusion method is used in PM 2.5 The monitoring data is fused by eVNA after being processed by the sliding average model and before being input into the recursive spatiotemporal convolutional neural network, and the specific steps are as follows:

[0098] S3.2.1: Prepare data: inverse distance square weight between monitoring point i and model prediction area E PM 2.5 predicted value in the future , PM 2.5 concentration value simulated by WRF-CMAQ (numerical model) in the small area where the model prediction area E and the monitoring point i are located (which can be divided according to the actual situation of the site distribution) and , , is the square inverse of

[0099] S3.2.2: Set the distance as the Euclidean distance between the center point of the small area and the monitoring point in the inverse distance weight calculation, and define as follows:

[0100]

[0101]

[0102] is the square inverse of ; and are the horizontal and vertical distances between the monitoring point i and the center of the target small area E, when there is a monitoring point i on the model prediction area E, the corrected value of the monitoring point is directly used, n is the number of monitoring points.

[0103] Step 3.3: Based on the PM 2.5 monitoring data and the corrected PM 2.5 simulation value obtained in step 3.2, a 24-hour scale recursive spatiotemporal sensitive deep learning model (RSDL) is established.

[0104] In this step, the recursive spatiotemporal sensitive deep learning model RSDL is built, and its structure is as follows: Figure 2The recursive spatiotemporally sensitive deep learning model RSDL consists of two parts: an encoder and a decoder. The encoder and decoder are used to identify the local and regional impacts of meteorology and pollutants. Specifically, the encoder uses a downsampling method to convert a high-resolution coarse-grained feature map into a low-resolution fine-grained feature map, while the decoder uses an upsampling method to convert a low-resolution fine-grained feature map into a high-resolution coarse-grained feature map.

[0105] Both the encoder and decoder include several Conv_GRU networks. In the encoder and decoder structures, different Conv_GRU networks are used to extract the spatiotemporal features of the monitoring data and simulation data for grid data or feature maps of different resolutions. The extracted spatiotemporal features are then used as hidden layer data and input into the corresponding next-step Conv_GRU network.

[0106] The encoder and decoder functions are as follows:

[0107] Encoder: ;

[0108] Decoder: ;

[0109] The input value is a four-dimensional array containing time, independent variable index, grid row, and grid column: ; For input PM 2.5 Monitoring values ; The hidden layer is a four-dimensional array containing time, channel number, grid row, and grid column: ; The output layer is a three-dimensional array containing time, grid rows, and grid columns: .

[0110] in, The time sequence number represents the starting time of the report. and all hourly points within the next 24 hours; For weather and PM 2.5 The index number of the simulated data. This is a summary of the indicators; These are the row and column numbers of the grid. and These are the width and height of the grid domain; for hour, OK, The first of the columns The input values ​​for each indicator, yes PM 2.5 Monitor concentration values; is the predicted PM of the encoder 2.5 concentration; is the process output data of the encoder; is the channel number, is the number of channels; and respectively represent the encoder and decoder processes.

[0111] In some embodiments of the present application, the encoder comprises three layers of neural network modules, each of which comprises a Convolutional Gate Recurrent (Conv GRU) module and a convolution module, and leaky ReLU is used as the activation function between layers; the encoder part also adds a Dropout function with a probability of a preset value (the parameter can be valued according to the actual situation, and in some embodiments of the present application, the preset value is 0.2) to randomly design part of the hidden layer input, so as to ensure not to rely too much on a certain index and reduce the possibility of overfitting. The decoder also comprises three layers of neural network modules, and uses transposed convolution to up-sample the grid data.

[0112] In some embodiments of the present application, the detailed configurations of the encoder and the decoder can be referred to Table 5.

[0113] The Convolutional Gate Recurrent Unit Network used in the present application is a reset gate and an update gate, and the formulas are as follows:

[0114] Update gate: ;

[0115] Reset gate: ;

[0116] Candidate state: ;

[0117] Final layer output: ;

[0118] time is the time step, represents the update gate at the time step time, is the input data at the current time step, is the hidden layer output at the previous time point, and are the learnable parameters of the update gate corresponding to and respectively, and are the learnable parameters of the reset gate corresponding to and respectively, represents point-wise multiplication, Represents two-dimensional convolution. It is a deviation parameter. and They are also respectively and Learnable parameters.

[0119] When training the eRSDL model, the PM simulated by WRF-CMAQ is first used. 2.5 Preliminary PM values ​​obtained from the concentration value and moving average model 2.5 Linear regression was performed on the concentration prediction values ​​to obtain the PM2.5 concentration values ​​for each monitoring point based on the latest monitoring data. 2.5 The future concentration prediction values; then the PM2.5 at each monitoring point after regression. 2.5 Future concentration predictions were obtained using eVNA and WRF-CMAQ simulations of PM2.5. 2.5 Concentration values ​​are fused using inverse distance weighting to obtain further corrected PM2.5 values. 2.5 Simulated concentration values, and then corrected PM2.5 values. 2.5 Simulated concentration value and measured PM 2.5 Concentration and simulated meteorological data are input into RSDL for training, resulting in the final RSDL forecast model (eRSDL model) with additional preceding monitoring-simulation data preprocessing.

[0120] Step 4: Use the trained eRSDL model for regional PM 2.5 Concentration forecast, obtaining PM 2.5 Concentration forecast results.

[0121] Using a recursive approach, the established 24-hour eRSDL model trained in step 3 can be used to analyze PM. 2.5 Concentration can be predicted. Furthermore, the existing 24-hour scale recurrent spatiotemporal convolutional recurrent neural network can be extended to predict PM2.5 concentrations for other time periods using a recursive model. 2.5 Concentration, such as time-scale recursive spatiotemporal convolutional recurrent neural networks, are extended to predict PM2.5 concentration at multiple sites over the next 72 hours. 2.5 Concentration, specific steps are as follows:

[0122] S4.1: Predict PM for the next 72 hours 2.5 The forecast is divided into three 24-hour forecast phases. The previous phase provides the input required for the moving average model and the initial concentration input for the eRSDL model for the next phase.

[0123] S4.2: PM forecast for 0-23 hours 2.5 At the concentration, the PM2.5 concentration in the previous 24 hours was [data missing]. 2.5 The monitoring data was used as the initial values ​​for the eRSDL model to complete the first day's PM. 2.5 Hourly forecast;

[0124] S4.3: When the PM 2.5 concentration is forecasted for 24-47 hours, the hourly forecast values for 0-23 hours are input into the eRSDL model as an approximate substitute for the first day of monitoring data, and the PM 2.5 concentration for the second day is completed.

[0125] S604. When the PM 2.5 concentration is forecasted for 48-71 hours, the hourly forecast values for 24-47 hours are input into the eRSDL model as an approximate substitute for the second day of monitoring data, and the PM 2.5 concentration for the third day is completed.

[0126] In some embodiments of the present application, the environmental air PM 2.5 forecasting method is described.

[0127] A certain province and city in a certain region are selected as the research region, and the prediction time period is selected from August 1, 2018 to July 31, 2022 as the overall sample, and the period from August 1, 2018 to July 31, 2021 is divided into a training set range, and the remaining period from August 1, 2021 to July 31, 2022 is used as a test set. Using the atmospheric pollution source emission inventory and global forecast system (GFS) meteorological data in this region, a mesoscale meteorological model WRF and an air quality model CMAQ are built to forecast the hourly PM 2.5 concentration in Huizhou region for three days in 2022.

[0128] a) Use version 3.9.1 of the WRF model for meteorological simulation, and use version 5.3.2 of the CMAQ numerical model for air quality simulation under the driving of the meteorological simulation data provided by WRF.

[0129] b) The specific numerical prediction model knowledge is common knowledge, which will not be described in detail here.

[0130] Extract the monitoring data of each monitoring station in the region around the station in the city from 2018 to 2022, and the data comes from the China Environmental Monitoring Station. The extracted monitoring data is processed for abnormal value and missing value, abnormal value processing and dimensionless.

[0131] Data correlation analysis and principal component analysis were performed. First, the input set (i.e., all input data) was subjected to principal component analysis using the singular value decomposition algorithm (SVD), and the principal components of the input set with high correlation were screened out by combining correlation analysis. Among them, after extracting the hourly principal components, since the 9 indicators (8 indicators have direction, and the 9th is the static wind indicator) after preprocessing have a certain discrete type, Spearman correlation analysis was used to analyze the correlation between each meteorological component and PM 2.5 concentration. Then, cross-meteorological factor principal component analysis was performed to extract meteorological factors that are correlated with each other, and finally the dimension of the input set was reduced to the single-digit level.

[0132] Table 5 9 indicators after preprocessing

[0133]

[0134] The training and application framework of the RSDL method was built using the divided training set and test set.

[0135] In the Python3.10 environment, the MA model was first fitted by multi-dimensional high-order least squares. For the MA(q) model formula containing the WRF-CMAQ reference value, the formula is denoted as

[0136] ;

[0137] Among them,

[0138] ,

[0139] ;

[0140] Among them, is the polynomial coefficient of , and the superscript represents the transpose of the matrix.

[0141] Considering all n fitting samples, denote:

[0142]

[0143] ;

[0144] Then the MA(q) model is expressed as:

[0145] ;

[0146] According to the principle of least squares, the coefficient matrix can be calculated by the following formula:

[0147] ;

[0148] Thus, the coefficients required for the MA model are obtained.

[0149] In the embodiment, the prediction area is divided into a grid form for the convenience of data processing and calculation. In the embodiment, the grid division method is that the size of one grid is 3 kilometers by 3 kilometers (the size can be adjusted by oneself), the stations are distributed as evenly as possible in the center of the grid, and the number of grids is as a multiple of 4 as possible. Finally, the prediction area is divided into 57 rows by 79 columns of grids.

[0150] The Convolutional Gated Recurrent Unit network (Conv GRU) is used to extract the spatiotemporal features of the WRF-CMAQ model. It is one of the widely used recurrent spatiotemporal convolutional recurrent neural networks (ConvRNN), and the other is the Convolutional Long Short-Term Memory network (ConvLSTM). Table 1 compares the performance of ConvRNN and ConvLSTM in predicting PM 2.5 In terms of concentration performance, the Conv GRU with better performance is finally selected as the substructure of the RSDL.

[0151] Table 6 Comparison of performance of Conv LSTM and Conv GRU on the validation set

[0152]

[0153] After selecting the Conv GRU as the substructure of the RSDL, the RSDL is trained in the Python 3.10 environment using the Pytorch platform to determine the optimal parameters of the RSDL. The smooth L1 loss function is used to calculate the relative loss between the predicted concentration of PM 2.5 and the monitored concentration Y of PM 2.5 . .

[0154] ;

[0155] In the formula, represents the number of training samples; represents the prediction time; represents the number of monitoring stations; represents the row number and column number of the monitoring station in the grid; . represents the weight coefficient, which balances the global accuracy and peak accuracy by giving greater weight to the peak time, is the predicted concentration of the i-th row, j-th column, and n-th sample at the q-th time point, is the monitored concentration of the i-th row, j-th column, and n-th sample at the q-th time point.

[0156] ​​Adam, which has been proven effective for different types of deep networks, is chosen as the adaptive stochastic gradient descent (SGD) algorithm, and the weight regularization is set to 0.0001. Once the training process is completed, the model parameters (including weights and biases) will be saved for application forecasting. Due to the high complexity of hyperparameter tuning, the Bayesian optimization algorithm of the automatic machine learning tool Neural Network Intelligence (NNI) is used to determine the optimal hyperparameters. The Bayesian optimization algorithm is an optimization method based on Bayesian theory, which selects the value of the next hyperparameter by establishing a model containing prior distribution and posterior distribution to minimize the objective function. In NNI, the Bayesian optimization algorithm is used to search for the optimal combination of hyperparameters to minimize the validation error of the model, and the implementation of the Bayesian optimization algorithm is based on the following steps:

[0157] a) Define the search space: define the possible value range of each hyperparameter, for example, set the learning rate range to [0.001, 0.004].

[0158] b) Design the objective function: choose an objective function as the evaluation indicator of hyperparameter performance (L1, L2, and Smooth L1 function), and the Smooth L1 function is chosen in this invention.

[0159] c) Establish the prior distribution: choose a prior distribution as the initial distribution of the hyperparameters, including the shape distribution, average distribution, exponential distribution, etc., and the shape distribution and average distribution are mainly used in this invention.

[0160] d) Establish a surrogate model: use Gaussian process or tree structure model to fit the relationship between hyperparameters and objective function.

[0161] e) Select the next hyperparameter: select the value of the next hyperparameter according to the surrogate model and Bayesian formula.

[0162] f) Evaluate the hyperparameters: evaluate the performance of the selected hyperparameters using the objective function.

[0163] g) Update the surrogate model: update the surrogate model using the new evaluation results.

[0164] h) Repeat steps e)-g) until the maximum number of iterations is reached or the objective function converges.

[0165] Through these steps, the Bayesian optimization algorithm can automatically find the optimal combination of hyperparameters for RSDL within the search space, thereby improving the performance and generalization ability of the model. The main hyperparameters are shown in Table 7.

[0166] Table 7 Main hyperparameters of RSDL method

[0167]

[0168] Based on the above training method, the RSDL is trained, and the parameters with the best performance on the training set and the validation set are selected as the final parameters of the RSDL.

[0169] In the model application stage, the eRSDL prediction method of 24-hour prediction scale can be extended to 72-hour PM 2.5 forecast through recursive mode. The 72-hour PM2.5 forecast of eRSDL prediction can be divided into three 24-hour prediction stages, and the initial concentration input of RSDL and the input required for the sliding average model of the previous stage are provided for the next stage. The specific steps are as follows: the meteorological and PM 2.5 simulation values of WRF-CMAQ, and the monitoring data of the previous 24 hours of the starting time are input as the initial value of the eRSDL model to complete the first day (0-23 hours) PM 2.5 hourly forecast; then the 0-23 hourly forecast value is input into the eRSDL model as an approximate replacement of the first day monitoring data to complete the next day (24-47 hours) PM 2.5 hourly forecast; finally, the 24-47 hourly forecast value is input into the eRSDL model as an approximate replacement of the next day monitoring data to complete the third day (48-71 hours) hourly forecast. In this way, the PM 2.5 hourly concentration prediction results of the next 72 hours can be obtained. After predicting the test set, the future three-day PM2.5 hourly prediction data set of the national control site in Huizhou City from August 1, 2021 to July 31, 2022 is obtained.

[0170] In each iteration process, the eRSDL model inputs the initial value and the simulated PM2.5 concentration into the sliding average model, corrects part of the discrete grid in the simulated PM2.5 concentration based on the monitoring point, and extends it to the entire study area through eVNA data fusion. Finally, the corrected PM2.5 simulation concentration value, the PM2.5 measured value one hour in advance of the day, and the simulated meteorological field are input into the trained RSDL model for the final prediction correction processing to obtain the final hourly PM2.5 prediction value.

[0171] In order to evaluate the performance of the model, the mean absolute error (MAE), the mean square error (RMSE), the Pearson correlation coefficient (R), and the consistency evaluation (IOA) are used as evaluation indexes.

[0172] According to the distribution of MAE and RMSE, the daily average error of RMSE is about 6-8 g / and the absolute error of the fusion of the air quality model and the recursive spatiotemporal deep learning method is controlled within 4-7 g / , wherein the first-day prediction error is about 4-5 g / ; according to the change of the correlation coefficient R, the Pearson correlation of the fusion of the air quality model and the recursive spatiotemporal deep learning method on the first day is 0.89135, close to 0.9, and the hypothesis is accepted at 99% confidence. The average correlation on the second and third days is 0.78576 and 0.68964, both of which are above 0.6, meeting the correlation requirement; in terms of the consistency evaluation index, the average IOA of the fusion of the air quality model and the recursive spatiotemporal deep learning on the first day is about 0.93456, which is the highest among all methods and all prediction days.

[0173] Further detailed analysis of the prediction results of each hour shows that the eRSDL model is generally superior in the hourly prediction effect. Through MAE and RMSE, the three-day prediction error can be well controlled within 5-7 g / , and the average absolute error and the average square error are reduced by about 3 g / and 5 g / , respectively. At the same time, the R value of the fusion of the air quality model and the recursive spatiotemporal deep learning on the first day is significantly higher than that of the other two methods, and the daily R value fluctuates within about 0.8-0.9, and the sliding average model greatly reduces the error caused by the accidental deviation of the one-hour concentration as the initial concentration. In terms of IOA, the sliding average regression can quickly capture the simple linear and low-order nonlinear time series correlation rules, thereby improving the trend accuracy of the pollution prediction by improving the accuracy of the initial concentration.

[0174] The foregoing embodiments of the present application integrate the spatiotemporal convolution recurrent neural network and the recursive cycle structure, develop a novel RSDL method capable of synergistically extracting spatiotemporal relationships, closely associate parameters and outputs, effectively capture the spatiotemporal distribution characteristics in the model prediction results, and improve the prediction accuracy; the present application has high application value, and through the MA model and the data fusion method of enhanced Voronoi region interpolation, the input data of the RSDL can be optimized, so that the RSDL has higher data adaptability and long-time rule mining capability. The present application fully utilizes the fitting capability of machine learning for nonlinear problems, realizes more accurate prediction of PM 2.5 , and the method has outstanding advantages, clear physical meaning of parameters, and strong applicability.

[0175] The method provided by the foregoing embodiments of the present application is based on a Weather Research and Forecasting-Community Multiscale Air Quality Model (WRF-CMAQ) simulation sample and PM 2.5 The monitoring data, combined with a recursive spatiotemporal deep learning (RSDL) network, are processed through a sliding average regression model (MA model) and a Voronoi area difference value method (eVNA) to form an enhanced RSDL method (eRSDL) by inputting historical monitoring values to the RSDL learning process at a pre-reporting time to fuse and correct the mode prediction values. The specific steps of the present application are as follows: selecting PM 2.5 The monitoring, simulation and meteorological data are used as a data set (for example, the time span is not less than three years); after the data set is preprocessed, the correlation analysis method and the principal component analysis method are used to screen out indexes with high correlation with the measured PM 2.5 concentrations; in the training of the eRSDL, firstly, the PM 2.5 concentrations simulated by the WRF-CMAQ are linearly regressed with the PM 2.5 concentrations of each monitoring point based on the latest monitoring data fitting; then, the PM 2.5 concentration prediction values of each monitoring point after regression are inversely distance weighted and fused with the PM 2.5 concentrations simulated by the WRF-CMAQ to obtain further corrected PM 2.5 concentrations; the simulation simulation values are further fused with the measured PM 2.5 concentrations and simulated meteorological data to obtain an RSDL prediction model with additional pre-processing of monitoring-simulation data, that is, an eRSDL model. In the business prediction, the meteorological and PM 2.5 concentrations simulated by the WRF-CMAQ, and the monitoring data of the previous 24 hours are input into the eRSDL model to complete the first-day (0-23 hours) hourly prediction; then, the 0-23 hourly prediction values are input into the eRSDL model as an approximate replacement of the first-day monitoring data to complete the second-day (24-47 hours) hourly prediction; the 24-47 hourly prediction values are input into the eRSDL model as an approximate replacement of the second-day monitoring data to complete the third-day (48-71 hours) hourly prediction, so that the PM 2.5The hourly concentration prediction result. The eRSDL method of the present application adds a moving average model and a data fusion method to optimize the input data on the basis of the traditional RSDL method, improves the prediction quality of WRF-CMAQ before the RSDL deviation correction, compensates for the insufficient ability of the long short-term memory neural network (LSTM) or the convolutional neural network (GRU) to simulate periodic changes, and keeps the prediction deviation within a stable range; in the PM 2.5 The hourly concentration prediction process has better adaptability to training data and long-term sequence rule mining ability, thereby effectively improving the PM 2.5 The hourly concentration prediction accuracy.

[0176] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application will not be limited to the embodiments shown herein but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. Air quality model coupled with recursive spatiotemporal deep learning PM 2.5 The hourly concentration forecasting method is characterized by, Includes the following steps: Build a PM including the site to be tested 2.5 The datasets of monitoring and simulating concentration data and meteorological data are divided into training and testing sets and preprocessed. Feature screening of meteorological factors, selecting those related to PM 2.5 Simulated meteorological factors directly related to concentration forecasts are used, while indirect influences are excluded. An eRSDL model is constructed and trained using the training set. The eRSDL model includes a moving average model, a data fusion model, and a 24-hour scale recursive spatiotemporal deep learning model RSDL. The moving average model is used to convert PM... 2.5 Preliminary PM levels were obtained by combining monitoring data. 2.5 Predicted concentration values ​​will be preliminary PM2.5 levels. 2.5 Predicted concentration values ​​and PM2.5 simulated by WRF-CMAQ 2.5 Linear regression was performed on the concentration values ​​to obtain the PM2.5 concentration values ​​for each monitoring station based on the latest monitoring data. 2.5 The future concentration forecast; the data fusion model is used to fit the PM2.5 concentration of each monitoring station based on the latest monitoring data. 2.5 Future concentration forecasts and PM2.5 levels compared with WRF-CMAQ simulations 2.5 Concentration values ​​were fused using eVNA inverse distance weighted data to obtain further corrected PM2.5 concentrations. 2.5 Simulated concentration values; a 24-hour recursive spatiotemporal deep learning model RSDL is used for PM2.5 analysis. 2.5 Hourly concentration forecast; Regional PM is performed using the trained eRSDL model. 2.5 Concentration forecast, obtaining PM 2.5 Concentration forecast results.

2. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 1 2.5 The hourly concentration forecasting method is characterized by, The preprocessing includes any one or more of the following: outlier detection, outlier handling, missing value handling, and dimensionless processing of the dataset. The dimensionless processing includes any one or more of z-score standardization and min-max normalization.

3. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 1 2.5 The hourly concentration forecasting method is characterized by, Feature screening of meteorological factors is performed through correlation analysis and principal component analysis, including the following steps: Principal component analysis (PCA) was performed on the 24-hour internal meteorological values ​​of each item in the input set using singular value decomposition. Correlation analysis was then used to select principal components with high correlation within the input set. Finally, Spearman correlation analysis was applied to analyze the relationship between each meteorological component and PM2.

5. 2.5 The correlation of concentrations was analyzed, and then principal component analysis was performed across meteorological elements to extract the interrelated meteorological elements and screen out the required meteorological factors.

4. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 1 2.5 The hourly concentration forecasting method is characterized by, The moving average model is Among them, X t The value is obtained through the moving average formula, where β0 is the PM simulated by the moving average model in WRF-CMAQ. 2.5 The weighting of the concentration value variable, where δ is the random error. For the current hourly forecast node, WRF-CMAQ is used to evaluate PM. 2.5 Simulated values ​​of concentration, It's about X i The polynomial function is given by q, where q is the size of the moving average regression window.

5. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 1 2.5 The hourly concentration forecasting method is characterized by, When performing eVNA inverse distance weighted data fusion, the interpolation formula is: Weight i =Distance i -2 =(Δrow i 2 +Δcol i 2 ) -1 Among them, Weight i It is the inverse squared distance weight between monitoring point i and the model prediction region E, Revised i PM2.5 at monitoring points after preliminary optimization using a moving average model 2.5 Future concentration prediction, Model E and Model i The PM2.5 concentrations simulated by WRF-CMAQ for the model prediction region E and the region where monitoring point i is located are respectively. 2.5 Concentration value, Δrow i and Δcol i These are the horizontal and vertical distances between monitoring point i and the center of a small area E of the target, respectively. i -2 It is Weight i The reciprocal of the square.

6. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 1 2.5 The hourly concentration forecasting method is characterized by, The recursive spatiotemporally sensitive deep learning model RSDL includes an encoder and a decoder. The encoder uses a downsampling method to convert a high-resolution, coarse-grained feature map into a low-resolution, fine-grained feature map, while the decoder uses an upsampling method to convert the low-resolution, fine-grained feature map into a high-resolution, coarse-grained feature map.

7. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 6 2.5 The hourly concentration forecasting method is characterized by, In the recursive spatiotemporally sensitive deep learning model RSDL, the encoder-decoder function is: Encoder: Decoder: The input value is a four-dimensional array containing time, independent variable index, grid row, and grid column: For input PM 2.5 Monitoring values The hidden layer is a four-dimensional array containing time, channel number, grid row, and grid column: The output layer is a three-dimensional array containing time, grid rows, and grid columns: t = t0, t0+1, t0+2, ..., t0+23 are time sequence numbers, representing the starting time t0 and all hour points within the next 24 hours; l = 1, 2, ..., N are the meteorological and PM data. 2.5 The index number of the simulated data, N is the index summary; (row,col)∈[0,Width-1; 0,Height-1] are the row and column numbers of the grid, and Width and Height are the width and height of the grid domain; x t,l,row,col This is the input value for the l-th indicator in row 't', column 'col'. PM at t0-1 2.5 Monitor concentration values; It is the encoder's forecast PM 2.5 Concentration; γ is the encoder's process output data; c = 1, 2, ..., M are the channel numbers, and M is the number of channels; E(·) and D(·) represent the encoder and decoder processes, respectively.

8. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 6 2.5 The hourly concentration forecasting method is characterized by, The encoder consists of three neural network modules. Each neural network module contains a convolutional gated recurrent unit network module and a convolutional module. Leaky ReLU is used as the activation function between layers. The encoder also incorporates a Dropout function with a preset probability to randomly design some of the hidden layer inputs. The decoder consists of a three-layer neural network module that uses transposed convolutions to upsample the grid data.

9. The air quality model coupled with recursive spatiotemporal deep learning PM according to any one of claims 1-8 2.5 The hourly concentration forecasting method is characterized by, The existing 24-hour eRSDL model can be extended to PMs with a future timeframe greater than 24 hours using a recursive approach. 2.5 Concentration forecast.

10. The air quality model coupled with recursive spatiotemporal deep learning PM according to claim 9 2.5 The hourly concentration forecasting method is characterized by, The established 24-hour eRSDL model is extended to PM for the next 72 hours using a recursive approach. 2.5 Concentration forecasting; the forecasting process is divided into three 24-hour forecast phases. The previous phase provides the input required for the moving average model and the initial concentration input for the eRSDL model for the next phase. This is done during the 0-23 hour PM forecast. 2.5 When the concentration is high, the PM2.5 concentration of the preceding 24 hours will be reported. 2.5 The monitored concentrations were used as initial values ​​for the eRSDL model; in the 24-47 hour and 48-71 hour PM forecasts 2.5 At the concentration level, the initial values ​​are the 0-23 hour concentration values ​​predicted by the eRSDL model and the 24-47 hour concentration values ​​predicted by the eRSDL model, respectively.

Images