Ozone forecasting method based on multi-dimensional meteorological factors and variables

Abstract

The invention belongs to the technical field of environmental science, and provides an ozone forecasting method based on multi-dimensional meteorological factors and variables, which comprises the following steps: S1, collecting and sorting the multi-dimensional meteorological factors, and calculating change variables of each element; s2, taking all processed multi-dimensional meteorological factors as forecast factors, and carrying out standardization processing; s3, screening all processed forecasting factors by adopting a stepwise regression method to obtain an optimal forecasting factor combination; s4, performing model training on historical data by adopting a multi-layer perceptron machine learning method according to the screened forecast factors; s5, checking and evaluating the trained perceptron learning model, and comparing and evaluating the trained perceptron learning model with a linear regression model and a radial basis function model; s6, predicting the ozone pollution process in different time periods and different weather backgrounds; according to the invention, the accuracy and timeliness of ozone forecasting can be obviously improved, and more reliable data support is provided for weather and environment management departments.

Description

Technical Field

[0001] The invention belongs to the field of environmental science and technology, and specifically is an ozone forecasting method based on multi-dimensional meteorological factors and variables. Background Art

[0002] The formation process of ozone is complex, involving the interweaving of numerous chemical reactions and multiple influencing factors, such as temperature, humidity, light intensity, terrain characteristics and changeable meteorological conditions, all of which play an important role.

[0003] In recent years, the scientific research field has attached unprecedented importance to the development of ozone forecast models, and researchers have continuously pioneered and innovated, striving to develop more accurate and efficient forecasting tools. However, existing ozone forecasting technologies are mostly linear regression models or statistical models based on a single meteorological factor. These methods often have difficulty in providing accurate ozone concentration forecasts under complex meteorological conditions. Specifically, the existing technologies have the following significant shortcomings: Traditional ozone forecasting methods mainly rely on a single meteorological factor, such as temperature, humidity, etc., and fail to comprehensively consider the interactions and influences between multiple meteorological factors, which leads to a significant decrease in the accuracy and reliability of forecast results under extreme weather events or different seasonal conditions; existing models are insufficient in time resolution and real-time response capabilities, and cannot respond to ozone levels under rapidly changing meteorological conditions in a timely manner, limiting their application value in environmental management and public health warnings. Given the complexity and uncertainty of ozone pollution, building a model that can accurately predict its changing trends remains an extremely challenging task;

[0004] To this end, technicians in this field have proposed an ozone forecasting method based on multidimensional meteorological factors and variables, aiming to significantly improve the accuracy and timeliness of ozone forecasts and provide more reliable data support for meteorological and environmental management departments. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an ozone forecasting method based on multi-dimensional meteorological factors and variables to solve the problems raised in the background technology.

[0006] An ozone forecasting method based on multidimensional meteorological factors and variables comprises the following steps:

[0007] S1. Collection and collation of multidimensional meteorological factors: Collect 3-hour interval meteorological data on the ground and ERA5 grid points and daily data on ozone concentration, and calculate the time variables of the meteorological data at 3h, 6h, and 24h intervals, as well as the vertical change variables of each factor;

[0008] S2. Standardized forecast factors: All processed multidimensional meteorological factors, bimonthly classification, sunshine hours, and daily ozone values ​​of the previous day are used as forecast factors and standardized;

[0009] S3, stepwise regression screening factors: all processed prediction factors are screened by stepwise regression method to obtain the optimal combination of prediction factors;

[0010] S4, perceptron machine learning model training: Based on the screened predictive factors, a multi-layer perceptron machine learning method is used to train the model on historical data;

[0011] S5. Model evaluation: Test and evaluate the trained perceptron learning model, and compare and evaluate it with the linear regression model and radial basis function model;

[0012] S6. Ozone pollution model prediction: Select ozone pollution processes in different time periods and under different weather backgrounds for prediction.

[0013] Preferably, in S1, by extracting the time series characteristics and vertical structure characteristics of the meteorological data, the meteorological data is calculated for 3h, 6h, and 24h time variables, and the time variable calculations are respectively expressed as:

[0014] ΔX 3h =X(t)-X(t-3h)

[0015] ΔX 6h =X(t)-X(t-6h)

[0016] ΔX 24h =X(t)-X(t-24h)

[0017] Among them, X(t) represents the meteorological element value at time t, ΔX 3h , ΔX 6h , ΔX 24h It means that for each meteorological element, the difference between two consecutive moments, i.e., the intervals of 3h, 6h and 24h, is calculated respectively;

[0018] For the vertical change, by calculating the difference of meteorological elements at different altitudes, the vertical change variable is expressed as:

[0019] ΔX Z =X(Z2)-X(Z1)

[0020] Among them, Z1 and Z2 represent data of two different altitude layers, and X(Z1) and X(Z2) represent the meteorological element values ​​at altitudes Z1 and Z2 respectively.

[0021] Preferably, in S2, the variable standardization process is to use the sorted bimonthly classification factors as independent variables and all other factors as covariates and rescale them. The standardized expression of the rescaled covariates is:

[0022]

[0023] Among them, x represents the original data point, μ represents the mean of the data, and σ represents the standard deviation of the data.

[0024] Preferably, in S3, after all predictors are standardized, forward stepwise regression is used. Initially, there are no variables in the model. A variable is added to the model in each iteration, and the variable has the largest correlation with the response variable among the remaining variables. Each variable in the model is checked. If an existing variable becomes insignificant after adding the new variable, it is removed from the model. The above steps are repeated until no more variables can be added to the model, and the iteration is stopped.

[0025] Preferably, in S4, the multilayer perceptron neural network structure includes an input layer, a hidden layer and an output layer, wherein the input layer is composed of independent variables and covariates, with a total of 14 input nodes, each of which corresponds to a screened predictor; the hidden layer contains 7 neurons, each of which is a function of the sum of an input weight, and a hyperbolic tangent activation function is used for nonlinear transformation; the output layer is used to generate the final prediction result, i.e., ozone concentration;

[0026] The model training process is as follows: S401, the initial weights and biases of the network are randomly initialized by the system;

[0027] S402, using mean square error as a loss function and adopting an adjusted conjugate gradient method as an optimization algorithm;

[0028] S403, calculating the gradient of weights and biases according to the loss function, and updating the parameters using a gradient descent algorithm;

[0029] S404. Use a hyperbolic tangent activation function in the hidden layer to introduce nonlinear characteristics.

[0030] Preferably, in S5, the perceptron learning model trained in S4 is evaluated, and a linear regression method and a radial basis function machine learning method are used to build a model. By using the same predictor as an input feature, the linear regression model and the radial basis function model are trained respectively, and the performance indicators of the three models are compared and evaluated.

[0031] Preferably, in S6, by selecting historical ozone concentration data under different time periods and weather conditions, and screening the results according to the S3 factor, the input features used for prediction are determined, and then the trained perceptron learning model is loaded into the prediction system, and the corresponding input features are extracted according to the real-time monitoring data or the short-term weather forecast data, and the extracted input features are input into the perceptron learning model to obtain the predicted value of the ozone concentration.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention adopts the perceptron machine learning method to construct an ozone forecast model based on multidimensional meteorological factors. By introducing multidimensional meteorological factors and changes in independent variables of time and space, the accuracy and timeliness of ozone forecasts are significantly improved. The precise calculation of time variables enables the model to respond more promptly to ozone levels under rapidly changing meteorological conditions. Considering the changes in spatial variables, the model can more intuitively reflect the relationship between ozone concentration and meteorological elements, providing more reliable data support for meteorological and environmental management departments, helping the government make scientific decisions, protect public health, and promote sustainable social and economic development.

[0034] 2. The model constructed by the present invention shows good forecasting ability in practical applications, can accurately reflect the changes in ozone concentration in different years, has high stability and credibility, and its model has high prediction accuracy and uniform residual distribution. It can better simulate the changes in ozone concentration before and after the ozone pollution process, and the upward and downward trends are reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The present invention is a flow chart of an ozone forecasting method based on multi-dimensional meteorological factors and variables;

[0036] Figure 2 A training graph of a perceptron machine learning model of the present invention;

[0037] Figure 3 This is the model forecast and actual time series diagram of the ozone pollution process in a certain city in the first half of June 2023 in the embodiment. DETAILED DESCRIPTION

[0038] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0039] As attached Figure 1 To Attachment Figure 3 As shown:

[0040] Embodiment: The present invention provides an ozone forecasting method based on multidimensional meteorological factors and variables, comprising the following steps:

[0041] S1. Collection and collation of multidimensional meteorological factors: Collect meteorological data at 3-hour intervals on the ground and at ERA5 grid points and daily data on ozone concentration, and calculate the time variables of the meteorological data at 3h, 6h, and 24h intervals, as well as the vertical change variables of each element; extract the temporal characteristics and vertical structural characteristics of the meteorological data; calculate the time variables of the meteorological data at 3h, 6h, and 24h intervals, and the time variable calculations are expressed as follows:

[0042] ΔX 3h =X(t)-X(t-3h)

[0043] ΔX 6h =X(t)-X(t-6h)

[0044] ΔX 24h =X(t)-X(t-24h)

[0045] Among them, X(t) represents the meteorological element value at time t, ΔX 3h , ΔX 6h , ΔX 24h It means that for each meteorological element, the difference between two consecutive moments, i.e., the intervals of 3h, 6h and 24h, is calculated respectively;

[0046] For the vertical change, by calculating the difference of meteorological elements at different altitudes, the vertical change variable is expressed as:

[0047] ΔX Z =X(Z2)-X(Z1)

[0048] Among them, Z1 and Z2 represent data of two different altitude layers, and X(Z1) and X(Z2) represent the meteorological element values ​​at altitudes Z1 and Z2, respectively. By capturing the changing trends of meteorological elements over time and space, useful input features are provided for subsequent forecast models. By collecting and processing meteorological data at 3-hour intervals on the ground and at ERA5 grid points, meteorological factors in multiple dimensions including temperature, humidity, wind speed, air pressure, and their temporal characteristics and vertical structural characteristics are comprehensively considered. This multi-dimensional data input enables the model to more comprehensively capture the influencing factors of ozone generation and change. The introduction of multi-dimensional meteorological factors has significantly improved the accuracy of ozone forecasts. The model can more accurately simulate the changing trend of ozone concentration, allowing the model to adapt to the ozone prediction needs under different weather conditions and different time periods, and provide more reliable data support for meteorological and environmental management departments.

[0049] At the same time, the time variables of the meteorological data are calculated every 3h, 6h, and 24h, as well as the changing variables of each element in the vertical direction. This processing method can capture the changing trends of meteorological elements over time and space, and provide rich input features for the model. By calculating the differences in meteorological elements at different altitudes and considering the differences in meteorological data between the ground and high altitudes, the model can more comprehensively reflect the spatial variation characteristics of the atmospheric environment.

[0050] S2. Standardized forecast factors: All processed multidimensional meteorological factors, bimonthly classifications, sunshine hours, and daily ozone values ​​of the previous day are used as forecast factors and standardized. Variable standardization is to use the sorted bimonthly classification factors as independent variables, and all other factors as covariates and rescale them. The standardized expression of the rescaled covariates is:

[0051]

[0052] Among them, x represents the original data point, μ represents the mean of the data, and σ represents the standard deviation of the data. By calculating the mean and standard deviation of all covariates, and then applying the above standardization formula to each data point of each covariate, the data of all covariates are transformed, and the standardized data replaces the original data, ready for subsequent model training and evaluation, ensuring that all covariates have equal importance in the model, thereby improving the predictive ability and generalization ability of the model.

[0053] S3. Stepwise regression screening factors: All processed prediction factors are screened by stepwise regression method to obtain the optimal combination of prediction factors; after all prediction factors are standardized, forward stepwise regression is used. At the beginning, there are no variables in the model. Each iteration adds a variable to the model. This variable has the greatest correlation with the response variable among the remaining variables. Check each variable in the model. If an existing variable becomes insignificant after adding a new variable, remove it from the model. Repeat the above steps until no more variables can be added to the model and stop iterating. The optimal combination of prediction factors is screened by stepwise regression method to ensure that the model can still accurately predict ozone concentration under complex meteorological conditions. This method not only improves the prediction accuracy of the model, but also enhances the stability and reliability of the model. The factors that finally passed the screening are shown in Table 1:

[0054] Table 1 Multidimensional factor screening and its importance

[0055]

[0056]

[0057] S4. Perceptron machine learning model training: Based on the screened prediction factors, a multi-layer perceptron machine learning method is used to train the model on historical data; the multi-layer perceptron neural network structure includes an input layer, a hidden layer, and an output layer. The input layer is 14 layers consisting of independent variables and covariates, and the hidden layer is 1 layer, which contains 7 unobservable network units. Each hidden unit is a function of the sum of the input weights. This function is the system default hyperbolic tangent activation function. By continuously testing and tuning the model, an optimal ozone forecast model is obtained, and its weight value is automatically determined by the estimation algorithm, as shown in Table 1; the output layer is used to generate the final prediction result, that is, the ozone concentration;

[0058] The model training process is as follows: S401, the initial weights and biases of the network are randomly initialized by the system;

[0059] S402, using mean square error as a loss function and adopting an adjusted conjugate gradient method as an optimization algorithm;

[0060] S403, calculating the gradient of weights and biases according to the loss function, and updating the parameters using a gradient descent algorithm;

[0061] S404. Use a hyperbolic tangent activation function in the hidden layer to introduce nonlinear characteristics.

[0062] The relative errors of the training set and the test set of the model are 0.138 and 0.151 respectively. The error of the test set is close to that of the training set, and there is no overfitting. From the perspective of the weight of each variable, the three variables DT1, DT2, and DT3 that reflect the change of low-level temperature respectively represent the sum of the temperature difference between 900hpa and the ground at 14:00 and 17:00, the product of the temperature difference between the ground temperature at 14:00 and 05:00 and the temperature at 17:00, and the temperature difference between the 975hpa layer at 14:00 and 08:00, which represent the heating of the lower atmosphere after the surface receives solar radiation, and account for the largest proportion, accounting for 42.7% in total; followed by O3_1, which represents the accumulation of ozone concentration in the early stage, accounting for 26.7%; and V95, which represents the characteristics of the low-level wind field. , DV variables, representing V wind at 950hpa14 o'clock, wind speed difference between 900hpa and 975hpa at 08 o'clock, accounting for 14.6% in total; other variables have a lower proportion, DRH is the 24-hour change of ground relative humidity at 14 o'clock, accounting for 8.5%, KEZHAO and Vmon6 variables represent the difference in solar radiation climate, accounting for 7.9%; the main factor of the change in ozone concentration is the lower atmospheric temperature change factor, followed by the previous pollutant concentration, and then the low-level wind field factor, which is consistent with the actual experience of ozone forecasting.

[0063] S5, model evaluation: test and evaluate the trained perceptron learning model, and compare and evaluate it with the linear regression model and radial basis function model; evaluate the perceptron learning model trained in S4, use the linear regression method and the radial basis function machine learning method to build a model, and use the same predictor as the input feature to train the linear regression model and the radial basis function model respectively, and compare and evaluate the performance indicators of the three models.

[0064] Using the prediction factors in Table 1, the perceptron machine learning method was used to model 40 ozone prediction models with different contribution degrees of the factors. The maximum, minimum, upper quartile, lower quartile, median and other model prediction values ​​of the daily forecast data were obtained. These values ​​were plotted into a time series diagram with the optimal model forecast and the actual situation, as shown in Figure 1. Figure 3 As shown in the figure, from the simulation of a city in the first half of June 2023, it can be seen that the overall trend of the ensemble forecast is basically consistent with the actual situation, and the forecast of the optimal model is basically consistent with the actual situation. It can be seen that the forecast factor combination in Table 1 can simulate the changing trend of ozone concentration in a certain city well at different contribution levels, and the forecast factor contribution combination of the optimal model performs best.

[0065] S6. Ozone pollution model prediction: By selecting historical ozone concentration data under different time periods and weather conditions, and filtering the results according to the S3 factor, the input features used for prediction are determined, and then the trained perceptron learning model is loaded into the prediction system. Based on the real-time monitoring data or short-term weather forecast data, the corresponding input features are extracted, and the extracted input features are input into the perceptron learning model to obtain the predicted value of ozone concentration.

[0066] That is, using the same prediction factors, the linear regression method and the radius basis function machine learning method are used to build models, and the samples of Hengyang City from 2015 to 2022 are forecasted. The three models are evaluated using the accuracy coefficient (correlation coefficient square R2), mean absolute error, mean square error, root mean square error, maximum error, and minimum error, as shown in Table 2:

[0067] Table 2 Evaluation parameters of three models

[0068]

[0069] The average absolute error of the perceptron learning model is 11.04, the precision coefficient (R2) is 0.8766, and the root mean square error is 14.29. From the data comparison, it can be seen that the perceptron learning model has the advantages of smaller error, higher precision, smaller (larger) maximum (minimum) error, etc. compared with the models established by the other two methods. It is a more suitable model for ozone concentration forecast in Hengyang City, followed by the linear regression model. The radius basis function model has a larger error than the other two models.

[0070] From the above, we can see that the forecast factor combination is obtained through stepwise regression, and the optimal ozone forecast model is obtained through multi-layer perceptron machine learning, and a large number of tests and adjustments are carried out; and the change variables of lower atmospheric meteorological factors are innovatively introduced into the ozone forecast factors, and the importance of the forecast factors is analyzed; at the same time, the preferred ozone model is tested and it is found that: the ozone model forecast values ​​and actual observation values ​​both have the characteristics of periodic changes with high in summer and autumn and low in winter and spring, and are very similar in overall trends; the model prediction accuracy is high, and the residual distribution is uniform; the forecast factors can accurately reflect the changes in ozone in different years, and the model forecast stability is good; it can better simulate the changes in ozone concentration before and after the ozone pollution process, the model has a high degree of explanatory power, the upward and downward trends are reliable, and the model forecast has high credibility; finally, the three modeling methods are compared, and it is found that the perceptron machine learning model has the advantages of smaller error, higher accuracy, smaller (larger) maximum (minimum) error, etc. compared with the linear regression model and the radius basis function model.

[0071] It is important to note that other substitutions, modifications, changes and omissions may be made in the design, operating conditions and arrangement of the exemplary embodiments without departing from the scope of the invention. Therefore, the present invention is not limited to a particular embodiment, but extends to a variety of modifications that still fall within the scope of the appended claims.

[0072] Additionally, in order to provide a concise description of exemplary embodiments, all features of an actual embodiment (ie, those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention) may not be described.

[0073] It should be understood that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but for those of ordinary skill having the benefit of this disclosure, without undue experimentation, the development effort will be a routine task of design, fabrication, and production.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An ozone forecasting method based on multidimensional meteorological factors and variables, characterized in that: The following steps are involved: S1. Collection and collation of multidimensional meteorological factors: Collect 3-hour interval meteorological data on the ground and ERA5 grid points and daily data on ozone concentration, and calculate the time variables of the meteorological data at 3h, 6h, and 24h intervals, as well as the vertical change variables of each factor; S2. Standardized forecast factors: All processed multidimensional meteorological factors, bimonthly classification, sunshine hours, and daily ozone values ​​of the previous day are used as forecast factors and standardized; S3, stepwise regression screening factors: all processed prediction factors are screened by stepwise regression method to obtain the optimal combination of prediction factors; S4, perceptron machine learning model training: Based on the screened predictive factors, a multi-layer perceptron machine learning method is used to train the model on historical data; S5. Model evaluation: Test and evaluate the trained perceptron learning model, and compare and evaluate it with the linear regression model and radial basis function model; S6. Ozone pollution model prediction: Select ozone pollution processes in different time periods and under different weather backgrounds for prediction.

2. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S1, by extracting the time series characteristics and vertical structure characteristics of the meteorological data, the meteorological data is calculated for 3h, 6h, and 24h time variables, and the time variable calculations are respectively expressed as: ΔX 3h =X(t)-X(t-3h) ΔX 6h =X(t)-X(t-6h) ΔX 24h =X(t)-X(t-24h) Among them, X(t) represents the meteorological element value at time t, ΔX 3h , ΔX 6h , ΔX 24h It means that for each meteorological element, the difference between two consecutive moments, i.e., the intervals of 3h, 6h and 24h, is calculated respectively; For the vertical change, by calculating the difference of meteorological elements at different altitudes, the vertical change variable is expressed as: ΔX Z =X(Z2)-X(Z1) Among them, Z1 and Z2 represent data of two different altitude layers, and X(Z1) and X(Z2) represent the meteorological element values ​​at altitudes Z1 and Z2 respectively.

3. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S2, the variable standardization process is to use the sorted bimonthly categorical factors as independent variables and all other factors as covariates and rescale them. The standardized expression of the rescaled covariates is: Among them, x represents the original data point, μ represents the mean of the data, and σ represents the standard deviation of the data.

4. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S3, after all predictors are standardized, forward stepwise regression is used. There are no variables in the model at the beginning. A variable is added to the model in each iteration. The variable has the largest correlation with the response variable among the remaining variables. Each variable in the model is checked. If an existing variable becomes insignificant after adding the new variable, it is removed from the model. The above steps are repeated until no more variables can be added to the model and the iteration is stopped.

5. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S4, the multilayer perceptron neural network structure includes an input layer, a hidden layer and an output layer, wherein the input layer is composed of independent variables and covariates, with a total of 14 input nodes, each of which corresponds to a screened predictor; the hidden layer contains 7 neurons, each of which is a function of the sum of an input weight, and a hyperbolic tangent activation function is used for nonlinear transformation; the output layer is used to generate the final prediction result, i.e., ozone concentration; The model training process is as follows: S401, the initial weights and biases of the network are randomly initialized by the system; S402, using mean square error as a loss function and adopting an adjusted conjugate gradient method as an optimization algorithm; S403, calculating the gradient of weights and biases according to the loss function, and updating the parameters using a gradient descent algorithm; S404. Use a hyperbolic tangent activation function in the hidden layer to introduce nonlinear characteristics.

6. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S5, the perceptron learning model trained in S4 is evaluated, and a linear regression method and a radial basis function machine learning method are used to build a model. By using the same predictor as an input feature, the linear regression model and the radial basis function model are trained respectively, and the performance indicators of the three models are compared and evaluated.

7. The ozone forecasting method based on multidimensional meteorological factors and variables as claimed in claim 1, characterized in that: In S6, historical ozone concentration data under different time periods and weather conditions are selected, and the input features used for prediction are determined according to the S3 factor screening results. The trained perceptron learning model is then loaded into the prediction system, and the corresponding input features are extracted according to the real-time monitoring data or short-term weather forecast data. The extracted input features are input into the perceptron learning model to obtain the predicted value of ozone concentration.

Images