Ozone prediction method and system based on kriging coupled lightgbm

Abstract

This disclosure provides an ozone prediction method and system based on co-kriging coupled LightGBM, relating to the field of ozone concentration prediction technology. The method includes: acquiring and preprocessing hourly-scale O3-related multi-source data for the target site to extract basic features; inputting the basic features into a traditional LightGBM model to generate baseline prediction values; constructing co-kriging spatial prior features by combining O3 observations from other sites at the same hour, the spatial covariance weights of the spherical variogram function, and the similarity weights of auxiliary variables; constructing regression kriging residual features based on the model residuals; inputting the basic features, co-kriging spatial prior features, and regression kriging residual features into the LightGBM model for training, constructing a co-kriging coupled LightGBM model, and using the co-kriging coupled LightGBM model to predict the hourly O3 concentration at the target site. This disclosure improves the stability and accuracy of the final prediction results.

Description

Technical Field

[0001] This disclosure relates to the field of ozone concentration prediction technology, specifically to an ozone prediction method and system based on synergistic Kriging coupling LightGBM. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Ozone (O3) is a significant pollutant in near-surface atmospheric complex pollution. Its concentration variation is influenced by a variety of factors, including solar radiation, temperature, humidity, wind speed, boundary layer height, cloud cover, precipitation, and regional transport, exhibiting significant nonlinearity, temporal sequence, and spatial heterogeneity. Accurately obtaining hourly O3 concentrations at different monitoring stations or spatial locations is crucial for ozone pollution early warning, pollution process analysis, regional joint prevention and control, and pollution exposure assessment at the county or grid scale.

[0004] Existing methods for estimating ozone concentration mainly include statistical regression methods, time series models, spatial interpolation methods, and machine learning models. However, these methods often fail to adequately characterize the complex nonlinear relationships between meteorological conditions, pollutants, and O3, or they focus primarily on historical concentration variations, neglecting spatial correlation structures and multi-source meteorological drivers. Furthermore, while they can consider spatial distances between stations, they do not adequately address factors such as temperature, humidity, solar radiation, NO2, CO, and PM2.5. 2.5 The nonlinear relationship between auxiliary variables and O3 is not adequately expressed.

[0005] The LightGBM model, as a gradient boosting tree-based machine learning model, possesses strong non-linear fitting capabilities and feature interaction learning abilities, making it suitable for processing multi-source, high-dimensional, and complex environmental data. However, existing LightGBM models still have the following technical limitations: (1) If only the traditional LightGBM model is used to estimate O3 concentration, the model mainly relies on latitude and longitude, meteorological variables, contemporaneous pollutants and photochemical proxy variables, without explicitly introducing the spatial continuity information of O3 concentration at surrounding stations at the same time, which may lead to insufficient characterization of local spatial differences and regional transport characteristics.

[0006] (2) Co-kriging or regression kriging methods can estimate the concentration of target points by utilizing spatial correlation, but traditional kriging methods are usually difficult to make full use of auxiliary features such as multi-source meteorological, pollutant and satellite remote sensing at the same time, and the computational cost is high in large-scale hourly station data. Summary of the Invention

[0007] To address the aforementioned issues, this disclosure proposes an ozone prediction method and system based on co-kriging coupled with LightGBM. It utilizes O3 observations from other monitoring stations within the same hour, combined with spatial covariance weights of the spherical variogram and similarity weights of auxiliary variables, to construct co-kriging O3 spatial prior features. Furthermore, it constructs regression kriging residual features based on the residuals of the traditional LightGBM model. The basic features, co-kriging spatial prior features, and regression kriging residual features are input into the LightGBM model to construct a co-kriging coupled with LightGBM model. The co-kriging coupled with LightGBM model is then used to predict the hourly spatial estimation results of O3 concentration at the target station, thereby improving the spatial prediction accuracy of ozone concentration.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions: Ozone prediction methods based on co-kriging coupled with LightGBM include: Obtain O3-related multi-source data at the hourly scale for the target site and preprocess it; Basic features are extracted from preprocessed multi-source data; The basic features are input into the traditional LightGBM model, and the baseline predicted value of ozone concentration is output. Obtain O3 observations from other monitoring stations within the same hour, and construct co-kriging space prior features by combining the spatial covariance weights of the spherical variogram function and the similarity weights of auxiliary variables. Based on the residuals of the traditional LightGBM model, regression kriging residual features are constructed. The basic features, co-kriging spatial prior features, and regression kriging residual features are input into the LightGBM model to construct a co-kriging coupled LightGBM model. The hourly O3 concentration results of the target site are then predicted using the co-kriging coupled LightGBM model.

[0009] According to some embodiments, the present disclosure adopts the following technical solutions: An ozone prediction system based on co-kriging coupled LightGBM includes: The data acquisition module is used to acquire O3-related multi-source data at the hourly scale of the target site and preprocess it. The feature extraction module is used to extract basic features based on preprocessed multi-source data; The initial prediction module is used to input basic features into the traditional LightGBM model and output the baseline prediction value of ozone concentration. The prior feature construction module is used to obtain the O3 observation values ​​of other monitoring stations in the same hour, and to construct co-kriging space prior features by combining the spherical variogram space covariance weight and the auxiliary variable similarity weight. The final prediction module is used to construct regression kriging residual features based on the residuals of the traditional LightGBM model. The basic features, co-kriging spatial prior features and regression kriging residual features are input into the LightGBM model to construct a co-kriging coupled LightGBM model. The co-kriging coupled LightGBM model is then used to predict the hourly O3 concentration results of the target site.

[0010] According to some embodiments, the present disclosure adopts the following technical solutions: A computer program product includes a computer program that, when executed by a processor, implements the ozone prediction method based on co-kriging coupling LightGBM.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the ozone prediction method based on co-kriging coupling LightGBM.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the ozone prediction method based on co-kriging coupling LightGBM.

[0013] Compared with the prior art, the beneficial effects of this disclosure are as follows: The ozone prediction method based on co-kriging coupled with LightGBM disclosed herein combines co-kriging spatial priors with the LightGBM machine learning model. It explicitly introduces the spatial continuity information of O3 at surrounding sites within the same hour on the basis of traditional LightGBM, thereby improving the model's ability to express local spatial heterogeneity and regional transport characteristics.

[0014] The ozone prediction method based on co-kriging coupled with LightGBM disclosed herein considers not only the spatial distance relationship between the target site and neighboring sites when constructing the co-kriging spatial prior, but also constructs spatial covariance weights based on the spherical variogram, so that the spatial weights not only reflect distance attenuation, but also reflect the spatial correlation structure of ozone concentration.

[0015] This disclosure presents an ozone prediction method based on co-kriging coupled LightGBM, which incorporates NO2, CO, and PM2.5. 2.5 PM 10The similarity weights of auxiliary variables such as temperature, relative humidity, wind speed, solar radiation, boundary layer height, and total cloud cover enable the co-kriging spatial prior to comprehensively consider the similarity of chemical and meteorological states between neighboring stations and the target station, which is more physically reasonable than simple inverse distance interpolation or KNN neighborhood averaging.

[0016] This disclosure presents an ozone prediction method based on co-kriging coupled with LightGBM, which constructs a regression kriging residual correction mechanism. First, the nonlinear relationship between fundamental features and O3 concentration is learned using a traditional LightGBM model. Then, spatial co-kriging estimation is performed on the model residuals to compensate for local spatial biases that the machine learning model fails to explain, thereby improving the stability and accuracy of the final estimation results.

[0017] The ozone prediction method based on co-kriging coupled LightGBM disclosed herein integrates satellite ozone column concentration OMTO3e with ERA5 meteorological data and air quality station observation data to construct ozone column concentration proxy features, photochemical reaction proxy features, NO2 titration proxy features, and atmospheric diffusion stability features, thereby enhancing the model's comprehensive characterization ability of ozone generation, accumulation, diffusion, and consumption processes.

[0018] This disclosed ozone prediction method based on co-kriging coupled LightGBM uses K-fold cross-validation within the training set to generate out-of-sample baseline predictions for traditional LightGBM, which are then used to construct the training set residuals, reducing the impact of overfitting of the training residuals on the correction of the regression Kriging residuals. It can simultaneously output evaluation results for traditional LightGBM, co-kriging O3 prior, regression Kriging prior, and co-kriging coupled LightGBM models, facilitating a quantitative comparison of model performance before and after spatial prior enhancement. Attached Figure Description

[0019] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0020] Figure 1 This is a flowchart of the ozone prediction method based on co-kriging coupled LightGBM according to an embodiment of the present disclosure; Figure 2 This is a comparison chart of the accuracy of the traditional LightGBM, co-kriging prior, regression kriging prior, and co-kriging coupled LightGBM models according to embodiments of this disclosure. Detailed Implementation

[0021] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Example 1 One embodiment of this disclosure provides an ozone prediction method based on co-kriging coupled LightGBM, the method steps of which include: Step 1: Obtain O3-related multi-source data at the hourly scale for the target site and preprocess it; Step 2: Extract basic features based on the preprocessed multi-source data; Step 3: Input the basic features into the traditional LightGBM model and output the baseline predicted value of ozone concentration; Step 4: Obtain O3 observations from other monitoring stations in the same hour, and construct co-kriging space prior features by combining the spherical variogram space covariance weights and auxiliary variable similarity weights; Step 5: Construct regression kriging residual features based on the residuals of the traditional LightGBM model. Input the basic features, co-kriging spatial prior features, and regression kriging residual features into the LightGBM model to construct a co-kriging coupled LightGBM model. Use the co-kriging coupled LightGBM model to predict the hourly O3 concentration results of the target site.

[0025] As one embodiment, the ozone prediction method based on co-kriging coupled with LightGBM disclosed herein uses a traditional LightGBM model to obtain a baseline O3 concentration prediction value; further, it uses O3 observation values ​​from other monitoring stations in the same hour, combined with the spatial covariance weights of the spherical variogram function and the similarity weights of auxiliary variables, to construct co-kriging spatial prior features; simultaneously, it constructs regression kriging residual features based on the residuals of the traditional LightGBM model; finally, it inputs the basic features, co-kriging spatial prior features, and regression kriging residual features into the LightGBM model to construct a co-kriging coupled with LightGBM model, and uses the co-kriging coupled with LightGBM model to predict the spatial estimation results of the hourly O3 concentration at the target station.

[0026] It should be noted that the method described in this disclosure is mainly applicable to scenarios such as site space estimation, spatial interpolation, or county-level spatial estimation. When predicting the O3 concentration at a certain target site for a certain hour, co-kriging spatial prior only uses the O3 information of other sites at the same hour, and strictly excludes the O3 concentration of the target site itself to avoid the direct participation of the observed value of the target site itself in the prediction of the target value. The specific implementation process of this disclosure is as follows: Step 1: Obtain the multi-source data related to O3 at the target site on the hourly scale and preprocess it; Specifically, obtain the multi-source data related to O3 at the target site on the hourly scale. The multi-source data related to O3 includes air quality observation data, site longitude and latitude data, ERA5 meteorological data, and satellite data. The air quality observation data includes site number, monitoring time, O3, PM 2.5 、PM 10 、NO2, SO2, and CO pollutant concentration data, etc.; the site longitude and latitude data includes site number, longitude, and latitude; the ERA5 meteorological data includes variables such as 2m air temperature, 2m dew point temperature, solar radiation, total precipitation, boundary layer height, total cloud cover, and near-surface wind speed; the satellite data is used to extract OMTO3e satellite ozone column concentration data.

[0027] Furthermore, the process of preprocessing the collected multi-source data includes: 1. Unify the site number field as station_id (monitoring site number); 2. Unify the time field as time (time); 3. Unify the longitude and latitude fields as longitude (longitude) and latitude (latitude); 4. Unify the ozone field as O3; 5. Unify the naming of pollutant fields such as PM 2.5 、PM 10 、NO2, SO2, and CO; 6. Combine the date and hour fields into a complete hourly timestamp; 7. Convert longitude and latitude, pollutant concentrations, and meteorological variables into numerical variables; 8. Delete records with empty site number, time, longitude or latitude; 9. Remove duplicate observation records according to site number and time; 10. Perform multi-encoding reading and abnormal web page file identification on the table file to be compatible with different formats of CSV, TXT, Excel, or Parquet files; 11. Convert the common wide table format of national air quality to a long table format, so that different station columns are converted into a unified structure of station_id, time, pollutant (pollutant type) and value (pollutant concentration value); 12. Associate the station observation data with the station latitude and longitude metadata to obtain hourly station data containing station number, time, longitude, latitude and pollutant concentration.

[0028] Furthermore, when processing ERA5 meteorological data, UTC time is converted to Beijing time, and meteorological variables are extracted from the nearest meteorological grid points based on the station's latitude and longitude. Relative humidity is calculated based on 2m air temperature and 2m dew point temperature; meteorological driving features required for subsequent models are constructed based on the meteorological variables. When processing OMTO3e satellite ozone column concentration data, ozone column concentrations for the corresponding location and date are extracted based on the station's latitude, longitude, and date, and matched to the station's hourly samples. If satellite data is missing for certain dates or locations, preset background values ​​are used to fill in the gaps.

[0029] Step 2: Extract basic features based on the preprocessed multi-source data; Specifically, basic features are constructed based on preprocessed multi-source data. These basic features include spatial location features, hourly cycle features, meteorological features, pollutant features, satellite ozone column concentration proxy features, photochemical proxy features, and diurnal statistical features. The specific extraction process is as follows: (1) Spatial location features include: longitude, latitude, lon_lat_interaction.

[0030] lon_lat_interaction = longitude × latitude Where longitude is longitude; latitude is latitude; and lon_lat_interactio is the interaction term between longitude and latitude.

[0031] (2) Hourly periodic characteristics include: hour, hour_sin, hour_cos.

[0032]

[0033]

[0034] Where hour is the hour of the day; hour_sin is the sine periodic term of the hour; and hour_cos is the cosine periodic term of the hour.

[0035] (3) Meteorological characteristics include: temperature_c (2-meter temperature), relative humidity, wind speed, solar radiation, total precipitation, boundary layer height, and total cloud cover.

[0036] (4) Pollutant characteristics include: PM 2.5 PM 10 NO2, SO2, CO.

[0037] (5) Satellite ozone column concentration proxy features include: omto3e_col_o3, omto3e_uv_proxy, omto3e_no2_interaction, omto3e_surface_proxy.

[0038] Wherein, omto3e_col_o3 represents the satellite ozone column concentration; omto3e_uv_proxy represents the ultraviolet radiation proxy feature constructed based on ozone column concentration, solar radiation, and particulate matter attenuation; omto3e_no2_interaction represents the interaction feature between ozone column concentration, NO2, and temperature; and omto3e_surface_proxy represents the surface proxy feature based on the boundary layer height for calculating the ozone column concentration.

[0039] (6) Photochemical and atmospheric process proxy characteristics include: temp_sq (temperature squared), rad_sq (solar radiation squared), no2_sq (NO2 squared), o3_formation_potential (O3 formation potential proxy), temp_radiation (temperature × radiation), hot_dry_index (hot and dry index), ventilation_index (ventilation index), cloud_adjusted_radiation (cloud adjusted radiation), photo_chem_index (photochemical index), precipitation_humidity_index (precipitation humidity index), o3_photolysis_proxy (O3 photolysis proxy), voc_nox_proxy (VOCs / NOx proxy), no_titration_proxy (NOx titration proxy), thermal_wind_index (thermal-wind speed index), no2_t_interaction (NO2-temperature interaction), rad_no2_ratio (radiation / NO2 ratio), no2_co_ratio (NO2 / CO ratio), pm25_no2_ratio (PM25 / NO2 ratio). 2.5 (NO2 ratio), temp_rh_ratio (temperature / humidity ratio), o3_prod_rate_proxy (O3 generation rate proxy), atmospheric_stability (atmospheric stability proxy), solar_uv_proxy (effective ultraviolet radiation proxy).

[0040] in:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] Where T represents air temperature, R represents solar radiation, RH represents relative humidity, WS represents wind speed, BLH represents boundary layer height, C represents total cloud cover, and P represents precipitation.

[0059] (7) Daily-scale statistical features include: daily_temp_mean (daily average temperature), daily_temp_max (daily maximum temperature), daily_temp_range (daily temperature range), daily_radiation_mean (daily average solar radiation), daily_radiation_sum (daily cumulative solar radiation), daily_wind_mean (daily average wind speed), daily_precip_sum (daily cumulative precipitation), daily_humidity_mean (daily average humidity), and daily_pm25_mean (daily average PM25). 2.5 ), daily_no2_mean (daily average NO2).

[0060] The above-mentioned daily-scale statistical features are aggregated and calculated for hourly meteorological variables and pollutant variables according to station number and date, and are used to characterize the overall meteorological background and pollution background of the target station on that day.

[0061] Step 3: Input the basic features into the traditional LightGBM model and output the baseline predicted value of ozone concentration; This disclosure uses the constructed basic features as input variables and the hourly O3 concentration at the station as the target variable to train a traditional LightGBM model. The traditional LightGBM model does not incorporate spatial prior features from other stations in the same hour; it only uses basic features such as latitude and longitude, meteorological data, concurrent pollutants, satellite ozone column concentration proxy features, and photochemical proxy variables. Within the training set, K-fold cross-validation is used to generate an out-of-sample baseline prediction value, base_lgbm_pred (the baseline ozone concentration prediction value output by the traditional LightGBM model).

[0062] Specifically, the training samples are divided into K subsets, and K subsets are used each time. A LightGBM model is trained on one subset of the dataset, and predictions are made on the remaining subset to obtain the out-of-sample baseline prediction for each training sample. This process reduces the overfitting effect caused by directly using the full sample fitted values ​​from the training set residuals. For the test set or samples to be estimated, a traditional LightGBM model trained on the complete training set is used to generate base_lgbm_pred.

[0063] Using the trained traditional LightGBM model, the basic features are input into the traditional LightGBM model, and the baseline predicted value of ozone concentration is output.

[0064] Step 4: Obtain O3 observations from other monitoring stations in the same hour, and construct co-kriging space prior features by combining the spherical variogram space covariance weights and auxiliary variable similarity weights; For the target site i At any moment t O3 concentration prediction and estimation, selecting the same time point t O3 observations from other monitoring stations were used as spatial reference information, and the target station was strictly excluded. i O3 observation value.

[0065] First, convert the station's latitude and longitude to approximate planar coordinates:

[0066]

[0067] in, lon Indicates longitude. lat Indicates latitude, Indicates the reference latitude.

[0068] Target site i With neighboring stations j The spatial distance between them is:

[0069] Then, the spatial weights are calculated using the spatial covariance weights corresponding to the spherical variogram function. Let the range be *a* and the nugget ratio be *r*. η ,when dij < a When the semivariance function is:

[0070] when ,

[0071] The spatial covariance weights are:

[0072] In one embodiment, variable range a The target site has a search radius of 220 km, a nugget ratio η of 0.05, a maximum of 30 neighboring stations, a search radius of 350 km, and a minimum of 3 neighboring stations. Furthermore, to characterize the similarity of chemical and meteorological states between the target site and its neighboring stations, auxiliary variables were selected to construct co-weighting. These auxiliary variables included NO2, CO, and PM2.5. 2.5 PM 10 , temperature_c, relative_humidity, wind_speed, solar_radiation, boundary_layer_height, total_cloud_cover.

[0073] For the target site i With neighboring stations j Calculate the standardized differences of auxiliary variables and construct auxiliary similarity weights:

[0074] in Daux This represents the average absolute value of the standardized differences of the auxiliary variables. α This represents the auxiliary similarity weight strength parameter. In one embodiment, α Take 0.35.

[0075] The final co-Kriging approximation weights are:

[0076] Based on weight wj Calculate target site i At any moment t Co-kriging space priors:

[0077] Where O3(j,t) represents other stations at the same time. j The observed O3 values.

[0078] Furthermore, co-kriging uncertainty and neighborhood statistical characteristics are calculated simultaneously, including: ck_o3_variance, ck_neighbor_mean, ck_neighbor_std, ck_neighbor_min_distance, ck_neighbor_count, and ck_neighbor_weight_sum. Here, ck_o3_variance represents the uncertainty index constructed based on the weighted variance of neighboring sites, the nugget ratio, and the number of valid samples; ck_neighbor_mean represents the mean O3 of neighboring sites; ck_neighbor_std represents the standard deviation of O3 of neighboring sites; ck_neighbor_min_distance represents the distance between the target site and the nearest valid neighboring site; ck_neighbor_count represents the number of neighboring sites involved in the estimation; and ck_neighbor_weight_sum represents the sum of the weights of neighboring sites. If the number of valid neighboring sites is insufficient at a certain moment, the mean O3 of other sites in the same hour is used as a fallback prior; if still missing, the traditional LightGBM baseline prediction is used to fill in the gaps.

[0079] Step 5: Construct regression Kriging residual features based on the residuals of the traditional LightGBM model; Based on the baseline predictions of the traditional LightGBM model, calculate the baseline residual for each sample:

[0080] Where O3(i,t) represents the station i At any moment t The O3 observations, base_lgbm_pred(i,t) represent the baseline prediction of the traditional LightGBM model for this sample. For the target site i At any moment t The residual estimation uses the residual values ​​of other stations at the same time, strictly excluding the residuals of the target station itself. The co-kriging residuals are obtained using the same spherical variogram space covariance weights and auxiliary variable similarity weights as in step four.

[0081] Further, the regression Kriging O3 prior value was obtained:

[0082] If the residual estimation result is less than 0, then rk_o3 is truncated to a non-negative value.

[0083] In this way, the traditional LightGBM model is responsible for learning the nonlinear relationship between O3 concentration and basic meteorological, pollutant, satellite and photochemical characteristics; while the regression Kriging residual correction is responsible for compensating for the spatial structural biases that the baseline model fails to explain.

[0084] Step 6: Couple the co-kriging space prior with the LightGBM model to achieve the final prediction of O3 concentration.

[0085] The input features of the co-kriging space prior features, basic features, and regression kriging residual features are combined to form the input features of the co-kriging coupled LightGBM model. The input features of the co-kriging coupled LightGBM model include the following components.

[0086] The basic features include: longitude, latitude, lon_lat_interaction, temperature_c, relative_humidity, wind_speed, solar_radiation, total_precipitation, boundary_layer_height, total_cloud_cover, hour, hour_sin, hour_cos, temp_sq, rad_sq, no2_sq, o3_formation_potential, omto3e_col_o3, omto3e_uv_proxy, omto3e_no2_interaction, omto3e_surface_proxy, temp_radiation, hot_dry_index, ventilation_index, cloud_adjusted_radiation, photo_chem_index, precip_humidity_index, o3_photolysis_proxy, voc_nox_proxy, no_titration_proxy, thermal_wind_index, no2_t_interaction, rad_no2_ratio, no2_co_ratio, pm25_no2_ratio, temp_rh_ratio, o3_prod_rate_proxy, atmospheric_stability, solar_uv_proxy, daily_temp_mean, daily_temp_max, daily_temp_range, daily_radiation_mean, daily_radiation_sum, daily_wind_mean, daily_precip_sum, daily_humidity_mean, daily_pm25_mean, daily_no2_mean, PM 2.5 , PM 10 , NO2, SO2, CO.

[0087] The co-kriging prior features and regression kriging residual features include: ck_o3 (co-kriging O3 spatial prior value), ck_o3_variance, ck_neighbor_mean, ck_neighbor_std, ck_neighbor_min_distance, ck_neighbor_count, ck_neighbor_weight_sum, ck_residual (regression kriging residual correction value), ck_residual_variance (uncertainty index of residual co-kriging estimation), and rk_o3 (regression kriging O3 prior value). After inputting all features into the LightGBM model, a co-kriging coupled LightGBM model is constructed, and trained using the hourly O3 concentration at the site as the target variable. The co-kriging coupled LightGBM model automatically learns the nonlinear relationship between the basic features, co-kriging O3 prior, neighborhood statistics, and regression kriging residual correction through machine learning, ultimately outputting the predicted hourly O3 concentration at the target site.

[0088] in, Xbas e represents the set of basic features. Xck Represents the set of prior features in the co-kriging space. Xrk This represents the set of features corrected for regression Kriging residuals.

[0089] Step 7: Model Evaluation and Result Output Calculate the evaluation metrics for the following four types of models or prior results: 1. Traditional_LightGBM: The traditional LightGBM model; 2. Cokriging_O3_prior_only: Only use cokriging O3 space priors; 3. Regression_Kriging_prior: Regression Kriging O3 prior; 4. Improved_LightGBM_Cokriging: Co-kriging LightGBM model.

[0090] Evaluation metrics include the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).

[0091] The formula for calculating R² is:

[0092] The formula for calculating RMSE is:

[0093] The formula for calculating MAE is:

[0094] in, yi This indicates the observed O3 concentration. i This indicates an estimate of the O3 concentration. This represents the average observed O3 concentration. n Indicates the number of samples.

[0095] If the R² of the co-kriging coupled LightGBM model is higher than that of the traditional LightGBM model, and the RMSE and MAE are lower, then the co-kriging spatial prior enhancement mechanism can effectively improve the spatial estimation accuracy of ozone concentration. Finally, save the results file, including the prediction results file and the model evaluation index file. The prediction results file may include: station_id, time, longitude, latitude, O3, base_lgbm_pred, ck_o3, ck_o3_variance, ck_neighbor_mean, ck_neighbor_std, ck_neighbor_min_distance, ck_neighbor_count, ck_neighbor_weight_sum, ck_residual, ck_residual_variance, rk_o3, improved_lgbm_pred, PM 2.5 PM 10 , NO2, SO2, CO, temperature_c, relative_humidity, wind_speed, solar_radiation, boundary_layer_height.

[0096] Simulation Experiment (1) Data Acquisition: Hourly O3 observation data from Chinese ambient air quality monitoring stations (https: / / air.cnemc.cn:18014) and meteorological data from the fifth-generation global atmospheric reanalysis dataset - ERA5 developed by the European Centre for Medium-Range Weather Forecasts (https: / / cds.climate.copernicus.eu / datasets). Ozone column concentration data were obtained from the OMI (Ozone Monitoring Instrument) sensor carried by NASA Aura satellite (https: / / www.earthdata.nasa.gov / ).

[0097] (2) Using this disclosure, the O3 data of a certain site in Beijing in the past were predicted, and the model was evaluated and compared.

[0098] (3) Comparison diagram as follows Figure 2 The R-squared of the co-kriging coupled LightGBM model disclosed herein. 2 The R² value is 0.94, the RMSE value is 12.38, and the MAE value is 8.98. The traditional LightGBM model has a lower R² value. 2 The R value is 0.81, RMSE is 21.48, and MAE is 16.56. The R value of O3 for co-kriging interpolation is... 2 The R² value is 0.89, the RMSE value is 16.20, and the MAE value is 11.63. The R² value of O3 in regression Kriging interpolation is... 2 The value is 0.92, the RMSE value is 13.96, and the MAE value is 10.33.

[0099] Example 2 One embodiment of this disclosure provides an ozone prediction system based on co-kriging coupled LightGBM, comprising: The data acquisition module is used to acquire O3-related multi-source data at the hourly scale of the target site and preprocess it. The feature extraction module is used to extract basic features based on preprocessed multi-source data; The initial prediction module is used to input basic features into the traditional LightGBM model and output the baseline prediction value of ozone concentration. The prior feature construction module is used to obtain the O3 observation values ​​of other monitoring stations in the same hour, and to construct co-kriging space prior features by combining the spherical variogram space covariance weight and the auxiliary variable similarity weight. The final prediction module is used to construct regression kriging residual features based on the residuals of the traditional LightGBM model. It inputs the basic features, baseline predicted values, co-kriging spatial prior features, and regression kriging residual features into the LightGBM model to construct a co-kriging coupled LightGBM model. The co-kriging coupled LightGBM model is then used to predict the hourly O3 concentration results at the target site.

[0100] Example 3 One embodiment of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the ozone prediction method based on co-kriging coupling LightGBM.

[0101] Example 4 One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the ozone prediction method based on co-kriging coupling LightGBM.

[0102] Example 5 One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the ozone prediction method based on co-kriging coupling LightGBM.

[0103] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. An ozone prediction method based on co-kriging coupled with LightGBM, characterized in that, include: Obtain O3-related multi-source data at the hourly scale for the target site and preprocess it; Basic features are extracted from preprocessed multi-source data; The basic features are input into the traditional LightGBM model, and the baseline predicted value of ozone concentration is output. Obtain O3 observations from other monitoring stations within the same hour, and construct co-kriging space prior features by combining the spatial covariance weights of the spherical variogram function and the similarity weights of auxiliary variables. Based on the residuals of the traditional LightGBM model, regression kriging residual features are constructed. The basic features, co-kriging spatial prior features, and regression kriging residual features are input into the LightGBM model to construct a co-kriging coupled LightGBM model. The hourly O3 concentration results of the target site are then predicted using the co-kriging coupled LightGBM model.

2. The ozone prediction method based on co-kriging coupled LightGBM as described in claim 1, characterized in that, The acquisition of hourly-scale O3-related multi-source data for the target station includes air quality observation data, station latitude and longitude data, meteorological data, and satellite data. The air quality observation data includes station number, monitoring time, O3, and PM2.

5. 2.5 PM 10 Concentration data for pollutants NO2, SO2, and CO; station latitude and longitude data including station number, longitude, and latitude; meteorological data including 2m air temperature, 2m dew point temperature, solar radiation, total precipitation, boundary layer height, total cloud cover, and near-surface wind speed; satellite data used to extract ozone column concentration data.

3. The ozone prediction method based on co-kriging coupled LightGBM as described in claim 1, characterized in that, The basic features extracted from the preprocessed multi-source data include spatial location features, hourly cycle features, meteorological features, pollutant features, satellite ozone column concentration proxy features, photochemical proxy features, and daily-scale statistical features. The daily-scale statistical features aggregate and calculate hourly meteorological variables and pollutant variables according to the station number and date to characterize the overall meteorological and pollution background of the target station on that day.

4. The ozone prediction method based on co-kriging coupled LightGBM as described in claim 1, characterized in that, The process involves obtaining O3 observations from other monitoring stations within the same hour, and constructing co-kriging space prior features by combining the spatial covariance weights of the spherical variogram function and the similarity weights of auxiliary variables, including: Select the same time t The O3 observations from other monitoring stations are used as spatial reference information, while the O3 observations of the target station itself are excluded; First, convert the latitude and longitude of the stations into approximate planar coordinates and calculate the spatial distance between the target station and neighboring stations; Spatial weights are calculated using the spatial covariance weights corresponding to the spherical variogram function; Select auxiliary variables, calculate the standardized differences of the auxiliary variables for the target site and neighboring sites, and construct the similarity weights of the auxiliary variables; The co-kriging approximation weight is calculated based on spatial covariance weight and auxiliary similarity weight. The target station at time [time value] is then calculated based on the co-kriging approximation weight. t The co-kriging space prior values.

5. The ozone prediction method based on co-kriging coupled LightGBM as described in claim 1, characterized in that, The construction of regression Kriging residual features based on the residuals of the traditional LightGBM model includes: The baseline residual for each sample is calculated based on the baseline predictions of the traditional LightGBM model. For the target site at time t The residual estimation uses the residual values ​​of other stations at the same time and strictly excludes the residuals of the target station itself. The regression Kriging residual features are calculated by using the spherical variogram space covariance weight and the auxiliary variable similarity weight.

6. The ozone prediction method based on co-kriging coupled LightGBM as described in claim 1, characterized in that, The co-kriging coupled LightGBM model automatically learns the nonlinear relationship between basic features, co-kriging spatial prior features, baseline predicted values, and regression kriging residual features through machine learning, and finally outputs the hourly O3 concentration estimate for the target site: in, Xbase Represents the set of basic features. Xck Represents the set of prior features in the same Kriging space. Xrk This represents the set of features of the regression Kriging residuals.

7. An ozone prediction system based on co-kriging coupled LightGBM, characterized in that, include: The data acquisition module is used to acquire O3-related multi-source data at the hourly scale of the target site and preprocess it. The feature extraction module is used to extract basic features based on preprocessed multi-source data; The initial prediction module is used to input basic features into the traditional LightGBM model and output the baseline prediction value of ozone concentration. The prior feature construction module is used to obtain the O3 observation values ​​of other monitoring stations in the same hour, and to construct co-kriging space prior features by combining the spherical variogram space covariance weight and the auxiliary variable similarity weight. The final prediction module is used to construct regression kriging residual features based on the residuals of the traditional LightGBM model. The basic features, co-kriging spatial prior features and regression kriging residual features are input into the LightGBM model to construct a co-kriging coupled LightGBM model. The co-kriging coupled LightGBM model is then used to predict the hourly O3 concentration results of the target site.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the ozone prediction method based on co-kriging coupled LightGBM as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the ozone prediction method based on co-kriging coupled LightGBM as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the ozone prediction method based on co-kriging coupling LightGBM as described in any one of claims 1-6.

Images