Ozone forecast correction method and system based on machine learning model

Abstract

The invention provides an ozone forecast correction method and system based on a machine learning model, and relates to the technical field of ambient air forecast, and the ozone forecast correction method is mainly based on a nested grid air quality forecast mode system, a forecast result of a mesoscale weather forecast mode and observation data of an ambient air quality monitoring station. An air quality model is combined with a plurality of models in a machine learning method, and a correction model for near-surface ozone forecasting is established, so that the accuracy of ozone forecasting is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ambient air forecasting, and in particular to an ozone forecast correction method and system based on a machine learning model. Background Art

[0002] Severe ozone pollution is the main component of photochemical smog and greenhouse gases, which is harmful to human health, ecological balance and climate, and will have a negative impact on regional air quality. Therefore, providing accurate and timely ozone forecasts is of great significance to protecting public health, improving the ecological environment and air quality. The forecast effect of the numerical model currently used for air quality forecasting is affected by factors such as the initial meteorological field, boundary conditions, pollutant emission sources, physical and chemical parameterization schemes and atmospheric chemical data assimilation, as well as the uncertainty of ozone precursor VOCs emission inventory, anthropogenic and plant VOCs source inventory, and ozone generation and destruction process involving at least hundreds of pollutants in the gas phase, liquid phase and heterogeneous reactions. However, traditional numerical models are difficult to measure these uncertainty factors in an all-round way, which will lead to inaccurate ozone forecasts. Summary of the invention

[0003] The purpose of the present invention is to provide an ozone forecast correction method and system based on a machine learning model. The ozone forecast correction method is mainly based on a nested grid air quality forecast model system, the forecast results of a mesoscale weather forecast model and the observation data of ambient air quality monitoring stations. The air quality model is combined with multiple models in the machine learning method to establish a correction model for near-ground ozone forecast to improve the accuracy of ozone forecast.

[0004] In order to solve the above technical problems, the present invention adopts the following solutions: A method for correcting ozone forecast based on a machine learning model, the method comprising the following steps: S1, collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; S2. The forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are used as input feature variables of the correction model, and the site observation data K is used as the label variable of the correction model. The input feature variables and label variables constitute the input data set of the correction model, and the input data set is divided into training set data and test set data; S3, input the training set data into the revised model, perform sliding training on the revised model according to the forecast date, obtain the test results, and then obtain the optimal training days based on the test results and the site observation data K; S4, selecting different input feature variables to train the revised model with the optimal number of training days to obtain training results, and evaluating the training results through evaluation parameters to obtain the optimal input variables; S5. Correct the forecast data using the corrected model with the best training days and the best input variables.

[0005] Furthermore, in S1, the preprocessing specifically includes deletion of abnormal or missing values, data matching, and data normalization.

[0006] Further, in S1, the WRF forecast meteorological data is the WRF forecast meteorological data obtained by performing WRF forecast on a certain date at the forecast time, and the WRF forecast meteorological data is then marked with a label date indicating a certain date; the NAQPMS forecast pollution data is the NAQPMS forecast pollution data obtained by performing NAQPMS forecast on a certain date at the forecast time, and the NAQPMS forecast pollution data is then marked with a label date indicating a certain date, then the corresponding WRF forecast meteorological data and NAQPMS forecast pollution data are obtained by forecasting the same date at the same forecast time.

[0007] Furthermore, in S1, the site observation data K is ozone data monitored by monitoring equipment at the site in the target area.

[0008] Furthermore, the specific process in S3 is: Get the current preset total training days and the current preset training gradient days; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; According to the corresponding forecast date, the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are sorted respectively, and then sliding training is performed according to the preset training gradient days as the gradient to obtain the test results corresponding to the forecast date; According to the forecast date, the corresponding test result is compared with the site observation data K to obtain a comparison result, and then the error value corresponding to the current preset total number of training days is obtained according to the comparison result; The preset total number of training days with the smallest error value is selected as the optimal number of training days.

[0009] Furthermore, the specific process in S4 is: Get the current best number of training days for the revised model; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; Screening input feature variables in the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} to obtain the current input feature variables; The correction model is trained according to the current input feature variables and the current optimal training days to obtain the training results, and then the error evaluation of the training results is performed through the evaluation parameters to obtain the error evaluation value corresponding to the current input feature variables; The input feature variable with the smallest error evaluation value is selected as the optimal input feature variable.

[0010] Furthermore, the WRF forecast meteorological data includes the WRF forecast RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data, and the NAQPMS forecast pollution data includes O3 and O3_8h hourly concentration data. In S2, the O3 and O3_8h hourly concentration data and the RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data are used as input characteristic variables of the correction model.

[0011] Further, in S4, the evaluation parameters include root mean square error, mean absolute error, correlation coefficient, root mean square error reduction rate, and mean absolute error reduction rate.

[0012] Furthermore, the revised model is trained and constructed by machine learning models RF, LightGBM, and XGBoost.

[0013] An ozone forecast correction system based on a machine learning model, applying the ozone forecast correction method based on a machine learning model, comprising: Data collection module: used to collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; Model building module: used to take the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} as the input feature variables of the correction model, take the site observation data K as the label variable of the correction model, and use the input feature variables and label variables to form the input data set of the correction model, and divide the input data set into training set data and test set data; The module for obtaining the optimal number of training days is used to input the training set data into the correction model, perform sliding training on the correction model according to the forecast date, obtain the test results, and then obtain the optimal number of training days based on the test results and the site observation data K; The module for obtaining the best input variables is used to select different input feature variables to train the correction model with the best training days, obtain the training results, evaluate the training results through evaluation parameters, and obtain the best input variables; Correction module: Correct the forecast data using the corrected model with the best training days and input variables.

[0014] Beneficial effects of the present invention: The present invention provides an ozone forecast correction method and system based on a machine learning model. The ozone forecast correction method mainly corrects the ozone numerical forecast results based on the machine learning model to solve the problem of inaccurate forecasts of traditional numerical models. Data is collected by nesting a grid air quality forecast model system, forecast results of a mesoscale weather forecast model, and observation data from ambient air quality monitoring stations. The correction model is subjected to sliding training using a forecast meteorological data set and a forecast pollution data set, and the optimal training days and the optimal input variables are obtained. The correction model is constructed using the optimal training days and the optimal input variables to ensure the accuracy of the correction model and improve the accuracy of the ozone forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the process of the ozone forecast correction method in Example 1 of the present invention; Figure 2This is a schematic diagram of inputting characteristic variables in Example 1 of the present invention; Figure 3 Schematic diagram of evaluation parameters and formulas in Example 1 of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0018] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0019] Additionally, descriptions of well-known structures, functions, and configurations may be omitted for clarity and conciseness.One of ordinary skill in the art will recognize that various changes and modifications may be made to the examples described herein without departing from the spirit and scope of the present disclosure.

[0020] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.

[0021] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0022] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments: Example 1 In this embodiment, a method for correcting ozone forecast based on a machine learning model is provided. The method for correcting ozone forecast is mainly based on a machine learning model to correct the ozone numerical forecast results to solve the problem of inaccurate forecasts of traditional numerical models, such as Figure 1 As shown, the specific steps include: S1, collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; S2. The forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are used as input feature variables of the correction model, and the site observation data K is used as the label variable of the correction model. The input feature variables and label variables constitute the input data set of the correction model, and the input data set is divided into training set data and test set data; S3, input the training set data into the revised model, perform sliding training on the revised model according to the forecast date, obtain the test results, and then obtain the optimal training days based on the test results and the site observation data K; S4, selecting different input feature variables to train the revised model with the optimal number of training days to obtain training results, and evaluating the training results through evaluation parameters to obtain the optimal input variables; S5. Correct the forecast data using the corrected model with the best training days and the best input variables.

[0023] Specifically, the NAQPMS pollution forecast data is the forecast result obtained through the nested grid air quality forecast model system (NAQPMS), the WRF meteorological forecast data is the forecast result obtained through the mesoscale weather forecast model (WRF), and the site observation data K is the ozone observation data monitored by the ambient air quality monitoring station.

[0024] When NAQPMS forecast pollution data and WRF forecast meteorological data are obtained through the above-mentioned nested grid air quality forecast model system and mesoscale weather forecast model, since the nested grid air quality forecast model system and the mesoscale weather forecast model can both forecast the forecast data for the next n days, and the nested grid air quality forecast model system and the mesoscale weather forecast model are both forecasted based on recent weather factors, it can be obtained that when the same date is forecasted on different dates, the two forecast results for the same date are not the same, and the closer to the forecast date, the more accurate the forecast result. For example, the forecast result for January 3, 2024 is obtained on January 1, 2024 and January 2, 2024 respectively. The two forecast results are compared with the actual ozone observation data of the day, and the forecast result obtained on January 2, 2024 is more accurate.

[0025] Based on the above situation, this embodiment mainly considers the problem that different forecast results generated by forecasting the same date on different dates are not the same. The grid air quality forecast model system and the mesoscale weather forecast model are nested on different forecast dates to obtain the forecast meteorological data set Q{q1, q2, ..., qn} and the forecast pollution data set R{r1, r2, ..., rn}. Specifically, the NAQPMS forecast pollution data and WRF forecast meteorological data for the next n days are obtained by nesting the grid air quality forecast model system and the mesoscale weather forecast model every day, and the WRF forecast meteorological data and the NAQPMS forecast pollution data are labeled according to the date. The forecast results of different forecast times can be obtained through the labeled date, which is convenient for data analysis.

[0026] When obtaining WRF forecast meteorological data, NAQPMS forecast pollution data and ozone observation data, the data can be preprocessed, and the preprocessing specifically includes deletion of abnormal or missing values, data matching, and data normalization, as follows: (1) Deletion of abnormal or missing values: abnormal equipment damage and other reasons may cause the monitoring data to be too large or the equipment to be insensitive, which may cause the monitoring data to be continuously zero for a period of time. This part of the ozone station monitoring data is deleted, and the data rows at the corresponding time in the forecast data are deleted at the same time; (2) Data matching: Pollutant data and meteorological data are derived from two adjacent stations. After the anomalies and missing values ​​are removed, the pollutant data and meteorological data are matched according to time. The matched data have a unified time series. (3) Data normalization: Data normalization is a basic work of machine learning. The purpose of normalization is to remove the irregularity of data. The values ​​of different input variables after normalization can be directly compared, eliminating the influence of the dimensions and dimensional units of different input variables, so that the influence of each variable on the model output value is consistent. In addition, normalization allows each input variable to be scaled to a smaller numerical range, and smaller values ​​can also improve computational efficiency. The study used the min-max normalization method to normalize the collected data. After normalization, the original data was scaled proportionally to between 0 and 1, but the distribution of the data did not change.

[0027] Then, in this embodiment, the correction model is trained and constructed based on the machine learning models RF, LightGBM, and XGBoost. The forecast meteorological data set Q{q1, q2,…, qn} and the forecast pollution data set R{r1, r2,…, rn} are mainly used as the input feature variables of the correction model, and the site observation data K is used as the label variable of the correction model. The input feature variables and label variables constitute the input data set of the correction model, and the input data set is divided into training set data and test set data.

[0028] Specifically, the WRF forecast meteorological data includes the WRF forecast RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data, and the NAQPMS forecast pollution data includes O3 and O3_8h hourly concentration data. The O3 and O3_8h hourly concentration data forecasted by NAQPMS and the WRF forecast RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data are used as input feature variables of the correction model, and the site observation data K is used as the label variable of the correction model.

[0029] Then, as described in S3, the training set data is input into the revised model, and the revised model is subjected to sliding training according to the forecast date to determine the optimal number of training days for the revised model, wherein the specific steps are as follows: Get the current preset total training days and the current preset training gradient days; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; According to the corresponding forecast date, the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are sorted respectively, and then sliding training is performed according to the preset training gradient days as the gradient to obtain the test results corresponding to the forecast date; According to the forecast date, the corresponding test result is compared with the site observation data K to obtain a comparison result, and then the error value corresponding to the current preset total number of training days is obtained according to the comparison result; The preset total number of training days with the smallest error value is selected as the optimal number of training days.

[0030] As a specific scheme of the above embodiment, the test time is selected as July 10, 2022 to August 8, 2022. The current preset total training days are the sliding training duration of 30 days before the test time. The current preset training gradient days are 5 days, and sliding training is performed with a gradient of every 5 days. The input feature variables include the hourly concentration data of O3 and O3_8h predicted by NAQPMS and the meteorological hourly data of RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH predicted by WRF. The training days with the smallest error in the test results are selected as the optimal training days.

[0031] For example, the forecast meteorological dataset Q predicted on ten consecutive forecast dates is obtained through the nested grid air quality forecast model system. Each forecast meteorological dataset Q includes the WRF forecast meteorological data for the next 10 days. Then, the forecast meteorological dataset Q1{q1,q2,…,q10}, the forecast meteorological dataset Q2{q2,q3,…,q11}, the forecast meteorological dataset Q3{q3,q4,…,q12}, and so on can be obtained. Among them, q1,q2,…,q10 represent different label dates respectively. It can be seen that different forecast meteorological datasets have the same q2, that is, the same future day is forecasted on different forecast dates. In view of the above situation, the ten forecast meteorological datasets are substituted into the correction model for sliding training. That is, sliding training is performed on ten forecast meteorological data sets with a gradient of every 5 days. Specifically, the correction model is firstly subjected to the first sliding training, that is, training is performed with {Q1{q1,q2,…,q10},Q2{q2,q3,…,q11},…, Q5{q5,q6,…,q14}} to obtain the test result corresponding to q15, which represents the result of correcting the date corresponding to the label date q15 by the forecast data. Then, the test result is compared with the actual station observation data K obtained on the date corresponding to the label date q15 to obtain the error value of the first sliding training. Then, the correction model is subjected to the second sliding training, that is, training is performed with {Q2{q2,q3,…,q11},Q3{q3,q4,…,q12},…, Q6{q6,q7,…,q15}}, and so on. By repeated single training with a large amount of forecast data, more accurate results can be obtained.

[0032] In this embodiment, when the optimal training days are 30 days, the error between the ozone concentration predicted by NAQMPS and the observed value after correction by the three models is reduced by about 35.8% to 37.7%. Therefore, 30 days is selected as the optimal training days for the machine learning correction model for ozone forecasting.

[0033] Then, as described in S4, different input feature variables are screened to train the revised model with the optimal number of training days to determine the optimal input variables of the revised model, wherein the specific steps are as follows: Get the current best number of training days for the revised model; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; Screening input feature variables in the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} to obtain the current input feature variables; The correction model is trained according to the current input feature variables and the current optimal training days to obtain the training results, and then the error evaluation of the training results is performed through the evaluation parameters to obtain the error evaluation value corresponding to the current input feature variables; The input feature variable with the smallest error evaluation value is selected as the optimal input feature variable.

[0034] As a specific solution of the above embodiment, when the current optimal training days are 30 days, Figure 2 As shown, in this embodiment, four different types of input feature variables are first set, namely input_1, input_2, input_3, and input_4. The revised model is trained with four different types of input feature variables. The results show that when input_4 is selected as the input variable for model training, the RMSE decrease rate after correction is the largest. Therefore, meteorological variables RH_2, T_2, U_10, V_10, PSFC, RAIN, PBLH and conventional pollutant variables O3 and O3_8h are selected as the best input feature variables for model training.

[0035] In order to evaluate the accuracy of the correction results, in this embodiment, the evaluation parameters include root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), root mean square error reduction rate (reduction rate of RMSE), mean absolute error reduction rate (reduction rate of MAE) and other statistical indicators for evaluation and inspection. The definition of each evaluation parameter and its calculation formula are as follows: Figure 3As shown in the formula, is the simulated concentration value of the model, is the average value of the simulated concentration, is the site observation data K, is the average value of the site observation data K, ML represents the machine learning algorithm, and NAQPMS represents numerical forecast.

[0036] Example 2 An ozone forecast correction system based on a machine learning model, applying the ozone forecast correction method based on a machine learning model, comprising: Data collection module: used to collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; Model building module: used to take the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} as the input feature variables of the correction model, take the site observation data K as the label variable of the correction model, and use the input feature variables and label variables to form the input data set of the correction model, and divide the input data set into training set data and test set data; The module for obtaining the optimal number of training days is used to input the training set data into the correction model, perform sliding training on the correction model according to the forecast date, obtain the test results, and then obtain the optimal number of training days based on the test results and the site observation data K; The module for obtaining the best input variables is used to select different input feature variables to train the correction model with the best training days, obtain the training results, evaluate the training results through evaluation parameters, and obtain the best input variables; Correction module: Input the test set data into the correction model, test the correction model with the optimal number of training days and the optimal input variables, and obtain the correction results.

[0037] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. According to the technical essence of the present invention, within the spirit and principles of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for correcting ozone forecast based on a machine learning model, characterized in that: The method comprises the following steps: S1, collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; S2. The forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are used as input feature variables of the correction model, and the site observation data K is used as the label variable of the correction model. The input feature variables and label variables constitute the input data set of the correction model, and the input data set is divided into training set data and test set data; S3, input the training set data into the revised model, perform sliding training on the revised model according to the forecast date, obtain the test results, and then obtain the optimal training days based on the test results and the site observation data K; S4, selecting different input feature variables to train the revised model with the optimal number of training days to obtain training results, and evaluating the training results through evaluation parameters to obtain the optimal input variables; S5. Correct the forecast data using the corrected model with the best training days and the best input variables.

2. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: In S1, the preprocessing specifically includes deletion of abnormal or missing values, data matching, and data normalization.

3. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: In S1, the WRF forecast meteorological data is the WRF forecast meteorological data obtained by performing a WRF forecast on a certain date at the forecast time, and then the WRF forecast meteorological data is marked with a label date used to represent a certain date; the NAQPMS forecast pollution data is the NAQPMS forecast pollution data obtained by performing a NAQPMS forecast on a certain date at the forecast time, and then the NAQPMS forecast pollution data is marked with a label date used to represent a certain date, then the corresponding WRF forecast meteorological data and NAQPMS forecast pollution data are obtained by forecasting the same date at the same forecast time.

4. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: In S1, the site observation data K is the ozone data monitored by the monitoring equipment at the site in the target area.

5. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: The specific process in S3 is: Get the current preset total training days and the current preset training gradient days; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; According to the corresponding forecast date, the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} are sorted respectively, and then sliding training is performed according to the preset training gradient days as the gradient to obtain the test results corresponding to the forecast date; According to the forecast date, the corresponding test result is compared with the site observation data K to obtain a comparison result, and then the error value corresponding to the current preset total number of training days is obtained according to the comparison result; The preset total number of training days with the smallest error value is selected as the optimal number of training days.

6. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: The specific process in S4 is: Get the current best number of training days for the revised model; Input the forecast meteorological data set Q{q1,q2,…,qn}, the forecast pollution data set R{r1,r2,…,rn} and the site observation data K with forecast dates within the current preset total training days into the revised model; Screening input feature variables in the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} to obtain the current input feature variables; The correction model is trained according to the current input feature variables and the current optimal training days to obtain the training results, and then the error evaluation of the training results is performed through the evaluation parameters to obtain the error evaluation value corresponding to the current input feature variables; The input feature variable with the smallest error evaluation value is selected as the optimal input feature variable.

7. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: The WRF forecast meteorological data include the WRF forecast RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data, and the NAQPMS forecast pollution data include O3 and O3_8h hourly concentration data. In S2, the O3 and O3_8h hourly concentration data and the RH_2, T_2, U_10, V_10, PSFC, RAIN, and PBLH meteorological hourly data are used as input characteristic variables of the correction model.

8. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: In S4, the evaluation parameters include root mean square error, mean absolute error, correlation coefficient, root mean square error reduction rate, and mean absolute error reduction rate.

9. The ozone forecast correction method based on a machine learning model according to claim 1, characterized in that: The revised model is trained and constructed by machine learning models RF, LightGBM, and XGBoost.

10. An ozone forecast correction system based on a machine learning model, characterized in that: The ozone forecast correction method based on a machine learning model as described in any one of claims 1 to 9 is applied, comprising: Data collection module: used to collect multiple forecast meteorological data sets Q{q1,q2,…,qn} and multiple forecast pollution data sets R{r1,r2,…,rn} of the target area obtained on different forecast dates, and collect multiple station observation data K of the target area obtained on different label dates, and preprocess the data; Among them, "q1, q2, ..., qn" and "r1, r2, ..., rn" represent multiple WRF forecast meteorological data and NAQPMS forecast pollution data sorted by label date, and the "q1, q2, ..., qn" and "r1, r2, ..., rn" predicted at the same forecast time correspond one to one based on the label date; Model building module: used to take the forecast meteorological data set Q{q1,q2,…,qn} and the forecast pollution data set R{r1,r2,…,rn} as the input feature variables of the correction model, take the site observation data K as the label variable of the correction model, and use the input feature variables and label variables to form the input data set of the correction model, and divide the input data set into training set data and test set data; The module for obtaining the optimal number of training days is used to input the training set data into the correction model, perform sliding training on the correction model according to the forecast date, obtain the test results, and then obtain the optimal number of training days based on the test results and the site observation data K; The module for obtaining the best input variables is used to select different input feature variables to train the correction model with the best training days, obtain the training results, evaluate the training results through evaluation parameters, and obtain the best input variables; Correction module: Correct the forecast data using the corrected model with the best training days and input variables.

Images