Hectometer-level greenhouse gas emission list generation method fused with vehicle-mounted underway observation

Abstract

The invention discloses a hectometer-level greenhouse gas emission list generation method fused with vehicle-mounted underway observation, which comprises the following steps of: acquiring data through vehicle-mounted underway observation, eliminating background influence in observation data, performing quality control and preprocessing, determining an optimal space average scale and characteristic variables, dividing data subsets based on a spatial clustering analysis method, and generating a hectometer-level greenhouse gas emission list. A model is constructed and evaluated, and the rationality of characteristic values is checked; and constructing a model to simulate regional concentration, and carrying out emission list downscaling on the basis of a concentration quantification weight coefficient. According to the method, the problems of light population misallocation, heterogeneous interpolation extrapolation distortion and lack of high-resolution input of machine learning in a hectometer method in the prior art are solved; a set of greenhouse gas emission list with high spatial resolution is generated by fusing the emission list with kilometer-level or even higher resolution with vehicle-mounted underway observation data, and hectometer-level greenhouse gas emission data can be provided for cities.

Description

Technical Field

[0001] This invention relates to a method for generating greenhouse gas emission inventories, specifically a method for generating greenhouse gas emission inventories at the 100-meter level by integrating vehicle-mounted mobile observation. Background Technology

[0002] Although cities account for only 3% of the global land area, they contribute 40% to 50% of anthropogenic carbon greenhouse gas emissions. Reliable greenhouse gas emission data is essential for effectively formulating emission reduction measures and scientifically assessing their benefits. However, current mainstream emission inventories (such as EDGAR and ODIAC) generally have spatial resolutions at the kilometer level or even coarser, making it difficult to precisely depict the highly heterogeneous spatiotemporal characteristics of emissions within cities. Therefore, there is an urgent need to construct emission data with high spatial resolution at the hundred-meter level to support precise urban greenhouse gas management and control.

[0003] Existing literature 1 (Zeng, C., Wu, S., Cheng, M., et al. High-resolution mapping of carbon dioxide emissions in Guizhou Province and its scale effects. Scientific Reports, 2024, 14(1), 20916) discloses a hundred-meter-level emission inventory, but it relies on nighttime light data and population data for spatial allocation of emissions, which easily leads to the allocation of industrial emissions to the central urban area with dense nighttime lights. Literature 2 (Bezyk Y, Sówka I, Górka M, et al. GIS-based approach to spatio-temporal interpolation of atmospheric CO2 concentrations in limited monitoring dataset[J]. Atmosphere, 2021, 12(3): 384) mentions that spatial interpolation is also a key tool for extending point and line observation data to area regions, but the strong heterogeneity of urban source and sink distribution poses a greater challenge to spatial interpolation, making it difficult to guarantee the prediction effect of extrapolated areas. Reference 3 (Liu, Z., Hu, YJ, Qiu, ZW, et al. Characteristics and prediction of traffic-related PMs and CO2 at the urban neighborhood scale. Atmospheric Pollution Research, 2024, 15(2), 101985) points out that while machine learning models are currently used to simulate the spatial distribution of greenhouse gas or pollutant concentrations and emissions, these studies require input factors such as meteorological elements and traffic volume, which are difficult to obtain in actual research and cannot achieve the high spatial resolution requirement of hundreds of meters. In summary, existing technologies have three main problems: first, the spatial allocation of emissions is unreasonable; second, the extrapolation effect of spatial interpolation is poor; and third, machine learning lacks high-resolution input, making it impossible to generate accurate emissions inventories at the hundreds of meters level. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating a greenhouse gas emission inventory at the 100-meter level by integrating vehicle-mounted mobile observation data. This method addresses the problems of misallocation of light and population, distortion in heterogeneous interpolation extrapolation, and lack of high-resolution input in existing 100-meter methods. Without refined meteorological and traffic data, this invention integrates emission inventories at the kilometer level or even higher resolution with vehicle-mounted mobile observation data to generate a high spatial resolution greenhouse gas emission inventory. This can provide cities with 100-meter-level greenhouse gas emission data, providing precise support for the formulation of emission reduction measures and the evaluation of emission reduction effects.

[0005] To achieve the above objectives, this invention provides a method for generating a 100-meter-level greenhouse gas emission inventory by integrating vehicle-mounted mobile observations. This method includes: Step 1: Conduct mobile monitoring and collect concentration data; Step 2: Eliminate background effects in the concentration data to obtain the enhancement value ΔC; Step 3: Perform quality control and preprocessing on the enhancement value ΔC to obtain the preprocessed ΔC data; Step 4: Use the standard error to determine the optimal spatial average scale of the preprocessed ΔC data, and obtain the spatially averaged ΔC data. Step 5: Select the feature variables for simulating the ΔC data after spatial averaging, and use the recursive feature elimination method to filter and obtain the optimal feature value set; Step 6: After spatially averaging the ΔC data and matching it with the corresponding optimal feature set, divide the data into several subsets according to the spatial clustering analysis method. Step 7: Divide the spatially averaged ΔC data and the corresponding optimal feature value set into training set and test set, construct ΔC evaluation model and test simulation effect; Step 8: Perform a mechanistic interpretation and completeness test on the characteristic variables of the △C evaluation model; Step 9: Construct a △C model using the spatially averaged △C data and the corresponding optimal feature set, and generate a △C spatial distribution map with a resolution of 100m×100m for the region; Step 10: Based on the regional ΔC data, quantify the weighting factors, downscale the emission inventory, and generate a greenhouse gas emission inventory at the 100-meter level.

[0006] Preferably, in step one, the data collection involves mobile monitoring in different functional areas of the city, simultaneously collecting greenhouse gas concentrations, concentration observation location information, and nearby road conditions; in step two, if there is a fixed-point observation upwind of the study area, its value is used as the background value; otherwise, the fifth percentile of the concentration observed by the mobile monitoring is used as the background value; in step three, observation data with vehicle speeds less than 10 km / h and observation data inside tunnels are excluded; in step five, the characteristic variables include Normalized Difference Vegetation Index (NDVI), Landuse Data (LD), Building High (BH), Car Park Density (CP), Building Density (BD), Bus Station Density (BS), Road Density (RD), Point of Interest (POI), Population Density (POP), and urban function area (…). UFA and any two or more of the following: Emission Inventory (EI), Waste Disposal Emission (WD), Energy-Related Emission (EN), and Agricultural Emission (AG); Preferably, in step six, the method includes: using Lon, Lat, NDVI, UFA, BH, POI, and RD as clustering variables, KMeans clustering is performed and then divided into folds according to region %5 for cross-validation of the model; in step seven, using the optimal feature value as input and the spatially averaged ΔC data as the target variable, an evaluation model is constructed using the training set, and the simulation effect is evaluated based on indicators such as the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) of the test set.

[0007] Preferably, in step five, when the ΔC data is ΔCO2, its corresponding optimal feature set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, and UFA; when the ΔC data is ΔCH4, its corresponding optimal feature set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, UFA, LD, AG, and EN; in step seven, the partitioning is to divide the spatially averaged ΔC data and the corresponding optimal feature set into 80% training set and 20% test set, ensuring that each data subset contains both training and test sets.

[0008] Preferably, in step nine, the method includes: using the optimal feature set as the input variable and the spatially averaged ΔC data as the target variable, reconstructing the model using all ΔC data and the corresponding optimal feature set; in step eight, using the additivity feature interpretation method to interpret the model mechanism, using Shapley values ​​to quantify the marginal contribution of feature values, interpreting the contribution and direction of each feature value to the model output, and identifying key factors affecting the distribution of urban greenhouse gas concentrations; if the Shapley value of a feature variable is abnormal, supplementing other basic geographical factors related to the greenhouse gas emission, and iteratively optimizing.

[0009] Preferably, in step eight, the formula for identifying the key factors affecting the distribution of urban greenhouse gas concentrations is as follows (Lundberg SM, Lee S I. A unified approach to interpreting model predictions[C]. Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS 2017). Curran Associates, Inc., 2017, 4768-4777): (2); In equation (2), It is the Shapley value of eigenvalue i. F It is the set of all features. S It does not include i Feature subset, It is a set S Input the value of the feature in the input field. Using feature sets S The trained model It is the result of model prediction. In the feature set S Add eigenvalues ​​to the basei Post-trained model, It is the model's prediction result; Features i At the | S | +1 number of all permutations of the positions; Features i The remaining | F |-| S | - The number of ways a feature can be arranged; The number of all possible permutations; Features i Previous | S The number of ways each feature can be arranged; For feature subset S The number of features contained therein; For features not included i A subset of features.

[0010] Preferably, in step ten, the method includes: dividing a coarse grid and further subdividing it into fine grids; determining a weighting coefficient based on the relative magnitude of the ΔC data in each fine grid and the sum of ΔC in the coarse grid; and redistributing the calculated total greenhouse gas emissions in the coarse grid to the fine grids to generate an emissions inventory for the fine grids.

[0011] Preferably, in step ten, the fine grid is configured with a spatial resolution of 100 m × 100 m; the formula for calculating the weighting coefficient is as follows: (3); (4); In equations (3) and (4), Indicates the position of the coarse grid. Indicates the location of the finer grid. Indicates the first In the coarse grid, the first The weighting factor of each fine grid, Indicates the first The total number of fine grids in a coarse grid, and , Indicates the first In the coarse grid, the first Simulated ΔC values ​​for a fine-grid array. Indicates the first Greenhouse gas emissions from a coarse grid Indicates the first In the coarse grid, the first Greenhouse gas emissions per fine grid.

[0012] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described herein.

[0013] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described herein.

[0014] This invention provides a method for generating a 100-meter-level greenhouse gas emission inventory by integrating vehicle-mounted mobile observations. This method addresses the problems of misallocation of light population, heterogeneous interpolation extrapolation distortion, and lack of high-resolution input for machine learning in existing 100-meter-level methods, and has the following advantages: 1. In the absence of refined meteorological and traffic data, this invention first obtains the true values ​​of ΔCO2 or ΔCH4 within a 100 m grid in urban areas through vehicle-mounted mobile high-frequency observations in steps one to three, ensuring that the observation data has the advantages of high spatial resolution and wide coverage. Steps four to nine construct a model based on the mobile concentrations: basic geographical factors are used to replace meteorological factors and traffic data as feature variables for modeling, achieving regional ΔC simulation without additional meteorological or traffic flow parameters. The recursive feature elimination method and the additivity feature interpretation method are used to optimize the feature values, and spatial clustering analysis is used to ensure the spatial independence of data during cross-validation. In step ten, the kilometer-level emission inventory is combined with the ΔC simulation results, and weighting coefficients are calculated based on the simulation results to generate a hundred-meter-level emission inventory, achieving a "coarse grid → hundred-meter" scale conversion. This provides direct data support for accurately identifying high-emission blocks in cities and quantifying emission reduction potential, overcoming the shortcomings of existing technologies that struggle to assess the effectiveness of measures due to the lack of high-resolution input.

[0015] 2. The method of this invention effectively overcomes the spatial misalignment of industrial emissions caused by reliance on nighttime light and population data in traditional technologies, as well as the spatial interpolation distortion caused by urban heterogeneity. The generated emission data at the hundred-meter level can accurately characterize the spatiotemporal heterogeneity of emissions within the city, providing a reliable data foundation for urban greenhouse gas control and governance. The final results can directly serve the precise formulation of urban emission reduction measures and the dynamic scientific evaluation of their implementation effects, significantly improving the precision and efficiency of carbon management at the city level, and providing important support for achieving low-carbon and sustainable urban development. Attached Figure Description

[0016] Figure 1 This is a map showing the functional zone distribution in Hangzhou and the route for near-surface atmospheric mobile observations according to the present invention.

[0017] Figure 2 This is a spatial distribution map of ΔCO2 and ΔCH4 in Hangzhou based on vehicle-mounted mobile observations, as presented in this invention.

[0018] Figure 3 This invention presents SEM images of the standard errors of ΔCO2 and ΔCH4 under different spatial resolution conditions in Hangzhou.

[0019] Figure 4 This is a graph showing the screening results of the optimal characteristic parameters of ΔCO2 and ΔCH4 in Hangzhou according to the present invention.

[0020] Figure 5 The diagram shows the effect of using data subsets formed by different methods of the present invention for modeling.

[0021] Figure 6 This is a simulation result diagram of the spatial distribution of ΔCO2 and ΔCH4 in Hangzhou according to the present invention.

[0022] Figure 7 This is a diagram showing the influencing factors of Hangzhou ΔCO2 analyzed by the additive characteristic interpretation method of this invention.

[0023] Figure 8 The diagram shows the influencing factors of Hangzhou △CH4 before and after adding WD, EN, and AG to the present invention, based on the additivity feature interpretation method.

[0024] Figure 9 This is a CO2 emission inventory diagram at different scales according to the present invention.

[0025] Figure 10 This is a CH4 emission inventory diagram at different scales according to the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 A method for generating a 100-meter-level greenhouse gas emission inventory by integrating vehicle-mounted mobile observation, the method comprising: Step 1: Conduct mobile monitoring and collect concentration data.

[0028] Mobile observations were conducted in different functional areas of Hangzhou's main urban area, including residential areas, commercial and business areas, comprehensive service areas, ecological protection areas, industrial development areas, rural areas, and farmland. Observations were conducted in clear weather, during the day (approximately 10:00-13:00) and at night (approximately 20:00-22:30). Four observation routes were used, covering various functional areas of the study region. A high-precision greenhouse gas analyzer (model 915-0011, Los Gatos Research, Mountain View, CA, USA) was used to collect greenhouse gas concentrations along the routes at a frequency of 1Hz. The sampling tubes were approximately 2.5m above the ground. GPS was used to record location information, and a vehicle recorder was used to record real-time road conditions. The obtained concentration data (…) C mobile ).

[0029] The route of navigation is as follows Figure 1 As shown, this invention presents a map of the functional zones in Hangzhou and the route of its near-surface atmospheric mobile observation, where (a) is a map of the city's functional zones; and (b) is the mobile observation route and background value location. Figure 1 It can be seen that the concentrations of CO2 and CH4 in the near-surface atmosphere of the city can be obtained.

[0030] Step 2: Eliminate the background influence in the concentration data to obtain the enhancement value ΔC.

[0031] To eliminate the influence of temporal variations in background concentration on the results, concentration data from mobile observations were used ( C mobile Subtract background concentration () C back The enhancement value ΔC obtained reflects the spatial variation of greenhouse gas concentration (ΔC). C = C mobile -C back If there is a fixed observation station upwind of the study area, the concentration value of that observation is used as the background value for the mobile observation; otherwise, the fifth percentile of the mobile observation is used as the background value. Here, the fixed observation station marked with "☆" is selected as the background value. The CO2 and CH4 concentrations obtained from the mobile observation are subtracted from the CO2 and CH4 concentrations of the background station, and the difference is used as the CO2 and CH4 concentration enhancement values ​​(ΔCO2 and ΔCH4).

[0032] Step 3: Perform quality control and preprocessing on the enhancement value ΔC to obtain the preprocessed ΔC data.

[0033] To reduce the impact of vehicle exhaust emissions at low speeds on the observation results, data from vehicle speeds less than 10 km / h were excluded. To eliminate the influence of sampling tube length on observation time, the actual observation time was obtained by subtracting the sampling lag time from the instrument recording time. The sampling lag time is caused by the sampling tube length and instrument flow rate. The time from exhaled gas to the start of gas concentration display was recorded before the experiment began, and this time is the sampling lag time. Since the changes in greenhouse gas concentration in the tunnel originate from vehicle emissions and are largely unaffected by geographical factors, observation data from the tunnel were excluded to ensure that the vehicle-mounted mobile observation data can represent the spatial variation characteristics of near-surface atmospheric CO2 and CH4 concentrations.

[0034] Step 4: Use the standard error to determine the optimal spatial average scale of the preprocessed ΔC data, and obtain the spatially averaged ΔC data.

[0035] The spatially averaged ΔC data are the ΔC data after 100 m spatial average processing, including spatially averaged ΔCO2 and ΔCH4 data. To effectively smooth the disturbances of micro-scale horizontal advection within the city, the standard error (SEM) of observations under different spatial resolutions was calculated at 10 m intervals. Specifically, the study area was divided into grids of 10 m × 10 m, 20 m × 20 m, 30 m × 30 m, ..., 200 m × 200 m. The SEM of observations within different grid sizes was calculated, and the optimal spatial average scale was determined by the inflection point of SEM decreasing with resolution. The SEM formula is as follows: (1); In equation (1), σ and n are the standard deviation and number of samples for each grid, respectively.

[0036] like Figure 3 As shown, this invention presents SEM images of the standard errors of ΔCO2 and ΔCH4 under different spatial resolution conditions in Hangzhou. (From...) Figure 3 It can be seen that the sample SEM decreases as the grid resolution decreases. When the resolution reaches 100 m, the attenuation trend tends to stabilize. Therefore, 100 m is determined to be the optimal spatial averaging scheme, and spatial averaging of ΔCO2 and ΔCH4 is performed within a range of 100 m.

[0037] like Figure 2 As shown, this invention presents spatial distribution maps of ΔCO2 and ΔCH4 in Hangzhou based on vehicle-mounted mobile observations, where (a) represents ΔCO2 and (b) represents ΔCH4. Figure 2 It can be seen that the ΔCO2 and ΔCH4 values ​​obtained from the mobile observation of Hangzhou's main urban area after 100 m spatial averaging show obvious spatial heterogeneity. The high ΔCO2 values ​​are mainly located in the city center, showing a trend of low values ​​in the east and west regions and high values ​​in the middle region. ΔCH4, on the other hand, shows a trend of high values ​​in the east region and low values ​​in the west region.

[0038] Step 5: Select the feature variables for simulating the ΔC data after spatial averaging, and use the recursive feature elimination method to filter and obtain the optimal feature value set.

[0039] The selection of eigenvalues ​​is crucial in determining the extension of point and line observation data to a surface region. Too few eigenvalues ​​fail to reflect the spatial distribution characteristics of emission sources, resulting in suboptimal simulation performance; too many eigenvalues ​​may contain redundant information, further worsening the simulation. Therefore, it is necessary to utilize recursive feature elimination for eigenvalue selection in ΔCO2 and ΔCH4 simulations and additivity feature interpretation for eigenvalue determination in ΔCO2 and ΔCH4 simulations.

[0040] First, 11 indicators were selected to construct a feature set for simulating the spatially averaged ΔC data of the main urban area of ​​Hangzhou within a 100m radius. These indicators include, but are not limited to, Normalized Difference Vegetation Index (NDVI), Landuse Data (LD), Building High (BH), Car Park Density (CP), Building Density (BD), Bus Station Density (BS), Road Density (RD), Point of Interest Density (POI), Population Density (POP), Urban Function Area (UFA), and Emission Inventory (EI). To eliminate potentially redundant information among the feature data, the Recursive Feature Elimination (RFE-RF) method of Random Forest was used for feature value selection. Random Forest models were constructed based on different combinations of feature values, and the root mean square error (RMSE) of the corresponding 10-fold cross-validation was calculated to determine the optimal number of feature values ​​and the optimal feature set.

[0041] like Figure 4 As shown in the figure, the screening results of the optimal characteristic parameters of ΔCO2 and ΔCH4 in Hangzhou according to the present invention are illustrated, where (a) is the root mean square error (RMSE) of ΔCO2; (b) is the optimal eigenvalue set of ΔCO2; (c) is the root mean square error (RMSE) of ΔCH4; and (d) is the optimal eigenvalue set of ΔCH4. Figure 4It can be seen that the RMSE gradually decreases as the number of eigenvalues ​​increases; when the number of eigenvalues ​​is ≥9, the RMSE tends to stabilize; the RMSE value is lowest when the number of eigenvalues ​​is 10; and the RMSE increases slightly after adding the eigenvalue LD. Therefore, the optimal number of eigenvalues ​​for ΔCO2 is 10, and the corresponding optimal eigenvalue set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, and UFA. LD is at the end of the eigenvalue importance ranking, and its role in the spatial distribution of ΔCO2 is relatively weak or its role can be replaced by other variables, so this eigenvalue is removed.

[0042] For ΔCH4, combined Figure 8 The results show that when the number of eigenvalues ​​WD, EN, and AG are added, the RMSE value is the lowest. When the eigenvalue WD is added, the RMSE value increases slightly. Therefore, the optimal number of eigenvalues ​​for ΔCH4 is 13, and the corresponding optimal eigenvalue set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, UFA, LD, AG, and EN.

[0043] Step 6: After spatially averaging the ΔC data, perform spatial matching with the optimal eigenvalue set and then divide the data into several subsets based on spatial clustering analysis.

[0044] Since the H2O AutoML platform uses random partitioning to generate cross-validation data subsets (folds) by default, this method often ignores spatial correlations between data when processing data with spatial characteristics, thus affecting the model's simulation capabilities. Therefore, this invention uses spatial clustering analysis to partition folds, improving the spatial independence of data during cross-validation, mitigating the interference of spatial autocorrelation on model performance evaluation, and enhancing the model's generalization ability. Therefore, seven variables—Lon, Lat, NDVI, UFA, BH, POI, and RD—are selected as clustering features. After data standardization, the KMeans algorithm (K=7) is used for clustering. The clustering results serve as region identifiers. Subsequently, the data are assigned fold numbers based on the region: fold = region % 5, which is used in the fold_column parameter in H2O AutoML.

[0045] Step 7: Divide the spatially averaged ΔC data and the corresponding optimal feature value set into training set and test set, construct ΔC evaluation model and test simulation effect.

[0046] The spatially averaged ΔCO2 and ΔCH4 observation data and their corresponding optimal eigenvalue sets were divided into an 80% training set and a 20% test set to ensure that each data subset contained both training and test sets. The optimal eigenvalues ​​were used as input variables, and the spatially averaged ΔCO2 and ΔCH4 data were used as target variables. The ΔCO2 and ΔCH4 evaluation models were constructed using the training set. The simulation performance of the ΔCO2 and ΔCH4 evaluation models was evaluated based on the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) of the test set.

[0047] like Figure 5 The diagram shows the effects of using data subsets formed by different methods in this invention for modeling, where (a) is the random method and (b) is the clustering result. Figure 5 It can be seen that the simulation performance of using spatial clustering results to divide the fold is slightly better than that of random partitioning. This improves the spatial independence of the data during cross-validation and enhances the generalization ability of the model to a certain extent.

[0048] Step 8: Perform a mechanistic interpretation and completeness test on the characteristic variables of the △C evaluation model.

[0049] Using ΔCO2 data with a spatial average of 100 m as the target variable and the selected optimal eigenvalues ​​as the input variables, ΔCO2 was simulated using H2O AutoML. The simulation results were interpreted using the additivity eigenvalue interpretation method. The Shapley value was used to quantify the marginal contribution of the eigenvalues, explaining the contribution and direction of each eigenvalue to the model output. Referring to existing techniques (Lundberg SM, Lee S I. A unified approach to interpreting model predictions[C]. Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS 2017). Curran Associates, Inc., 2017, 4768-4777), key factors influencing urban greenhouse gas concentration distribution were identified, and the formula is as follows: (2); In equation (2), It is the Shapley value of eigenvalue i. F It is the set of all features. S It does not include i Feature subset, It is a set S Input the value of the feature in the input field. Using feature sets S The trained model It is the result of model prediction. In the feature set S Add eigenvalues ​​to the base i Post-trained model, It is the model's prediction result; Features i At the | S | +1 number of all permutations of the positions; Features i The remaining | F |-| S | - The number of ways a feature can be arranged; The number of all possible permutations; Features i Previous | S The number of ways each feature can be arranged; For feature subset S The number of features contained therein; For features not included i A subset of features.

[0050] If the optimal model selected by H2O AutoML is an ensemble model (i.e., an integrated model constructed from multiple machine learning models), its interpretability is relatively weak. Therefore, a single model with simulation performance second only to the ensemble model needs to be selected for mechanistic analysis. If the magnitude of the Shapley value of a certain feature deviates from that of other feature values, it indicates that an important feature variable is missing, and the model cannot reflect the spatial variation of ΔC. In this case, other relevant data that can reflect the distribution of urban greenhouse gas sources and sinks need to be added as input feature variables, and steps five to ten above should be repeated.

[0051] like Figure 7 As shown, this invention analyzes the influencing factors of Hangzhou's ΔCO2 based on the additivity characteristic interpretation method. Figure 7 It can be seen that the SHAP values ​​corresponding to various eigenvalues ​​in the ΔCO2 model for Hangzhou's main urban area do not differ significantly in magnitude. During the day, the main factors affecting urban CO2 concentration are population density, building density, and road density, all of which have a positive impact; that is, as the eigenvalue increases, the CO2 concentration increases. At night, building density and parking lot density have a positive impact. These results are consistent with the characteristics of CO2 concentration changes, indicating that the simulation results are reasonable.

[0052] Similarly, the same eigenvalues ​​were selected for ΔCH4 simulation, and the SHAP values ​​corresponding to each eigenvalue in the ΔCH4 model for the main urban area of ​​Hangzhou were calculated. Figure 8 ).

[0053] like Figure 8 As shown, the influencing factors of Hangzhou △CH4 before and after adding WD, EN, and AG are analyzed using the additivity feature interpretation method, where (a) is the influencing factor diagram before adding WD, EN, and AG; and (b) is the influencing factor diagram after adding WD, EN, and AG. Figure 8 It is evident that when NDVI, LD, BH, CP, BD, BS, RD, POI, POP, UFA, and EI are introduced as eigenvalues, the SHAP value of RD is significantly larger than that of the other eigenvalues, indicating a substantial overestimation of its contribution and resulting in unreasonable simulation results. However, by adding three more eigenvalues—waste disposal emission (WD), energy-related emission (EN), and agricultural emission (AG)—the contributions of different eigenvalues ​​are balanced, their SHAP magnitudes become closer, and the simulation results become more reasonable.

[0054] Step 9: Construct a △C model using the spatially averaged △C data and the corresponding optimal feature set, and generate a △C spatial distribution map with a resolution of 100m×100m for the region.

[0055] Using the optimal feature set as the input variable and the spatially averaged ΔCO2 and ΔCH4 data as the target variables, the ΔCO2 and ΔCH4 models are reconstructed using all ΔCO2 and ΔCH4 observation data and the corresponding optimal feature set, generating ΔCO2 spatial distribution maps and ΔCH4 spatial distribution maps with a resolution of 100m×100m for the region, respectively.

[0056] like Figure 6 The figure shows the simulation results of the spatial distribution of ΔCO2 and ΔCH4 in Hangzhou according to the present invention, where (a) represents ΔCO2 and (b) represents ΔCH4. Figure 6 It can be seen that ΔCO2 and ΔCH4 exhibit high spatial heterogeneity in cities. High ΔCO2 values ​​are mainly located in the city center, where the population density and traffic volume are high, resulting in high greenhouse gas emissions. ΔCH4, on the other hand, shows a generally higher trend in the eastern region.

[0057] Step 10: Based on the regional ΔC data, quantify the weighting factors, downscale the emission inventory, and generate a greenhouse gas emission inventory at the 100-meter level.

[0058] A coarse grid is created, with a spatial resolution ≥ that of the existing greenhouse gas emission inventory requiring downscaling. This coarse grid is then further subdivided into finer grids of 100 m × 100 m. The minimum value is subtracted from the ΔCO2 (or ΔCH4) data of all fine grids to eliminate the influence of negative bias. Then, the weighting coefficients are determined by the relative magnitude of the corrected ΔCO2 (or ΔCH4) data of the fine grids and the sum of ΔCO2 (or ΔCH4) data of the coarse grids. The total greenhouse gas emissions within the coarse grid are calculated using the existing emission inventory. The emissions from the coarse grid are then allocated according to the weighting coefficients to generate a ΔCO2 (or ΔCH4) emission inventory with a spatial resolution of 100 m × 100 m. (3); (4); In equations (3) and (4), Indicates the position of the coarse grid. Indicates the location of the finer grid. Indicates the first In the coarse grid, the first The weighting factor of each fine grid, Indicates the first The total number of fine grids in a coarse grid, and , Indicates the first In the coarse grid, the first Simulated ΔC values ​​for a fine-grid array. Indicates the first Greenhouse gas emissions from a coarse grid Indicates the first In the coarse grid, the first Greenhouse gas emissions per fine grid.

[0059] Since the spatial allocation of a 1km×1km CO2 emission inventory has low uncertainty, this invention divides the space into coarse grids with a spatial resolution of 1km×1km, thus downscaling the 1km×1km emission inventory to 100m×100m.

[0060] like Figure 9 As shown, the present invention presents CO2 emission inventories at different scales, where (a) is the CO2 emission inventory at 1km × 1km before downscaling; and (b) is the CO2 emission inventory after downscaling to 100m × 100m. Figure 9 It can be seen that, compared with the CO2 emission inventory of 1km×1km before downscaling, the inventory with 100m resolution can more accurately reflect the local emission characteristics.

[0061] Due to the significant uncertainty in the spatial allocation of the CH4 emission inventory, this invention divides the entire study area into a coarse grid and redistributes the CH4 emission inventory spatially.

[0062] like Figure 10 As shown, the CH4 emission inventory diagrams at different scales of the present invention are shown, where (a) is the CH4 emission inventory at 0.1°×0.1° before downscaling; and (b) is the CH4 emission inventory after downscaling to 100m×100m. (From...) Figure 10 It is evident that, compared to an emission inventory with a resolution of 0.1°, an emission inventory with a resolution of 100m can reflect the emission situation in more localized areas such as city centers and major transportation routes, providing more refined data support for the formulation of emission reduction policies.

[0063] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for generating a 100-meter-level greenhouse gas emission inventory by integrating vehicle-mounted mobile observation, characterized in that, The method includes: Step 1: Conduct mobile monitoring and collect concentration data; Step 2: Eliminate background effects in the concentration data to obtain the enhancement value ΔC; Step 3: Perform quality control and preprocessing on the enhancement value ΔC to obtain the preprocessed ΔC data; Step 4: Use the standard error to determine the optimal spatial average scale of the preprocessed ΔC data, and obtain the spatially averaged ΔC data. Step 5: Select the feature variables for simulating the ΔC data after spatial averaging, and use the recursive feature elimination method to filter and obtain the optimal feature value set; Step 6: After spatially averaging the ΔC data and matching it with the corresponding optimal feature set, divide the data into several subsets according to the spatial clustering analysis method. Step 7: Divide the spatially averaged ΔC data and the corresponding optimal feature value set into training set and test set, construct ΔC evaluation model and test simulation effect; Step 8: Perform a mechanistic interpretation and completeness test on the characteristic variables of the △C evaluation model; Step 9: Construct a △C model using the spatially averaged △C data and the corresponding optimal feature set, and generate a △C spatial distribution map with a resolution of 100m×100m for the region; Step 10: Based on the regional ΔC data, quantify the weighting factors, downscale the emission inventory, and generate a greenhouse gas emission inventory at the 100-meter level.

2. The method according to claim 1, characterized in that, In step one, the data collection involves mobile observations conducted in different functional areas of the city, simultaneously collecting greenhouse gas concentrations, the location of the concentration observations, and nearby road conditions. In step two, if there is a fixed-point observation upwind of the study area, its value is used as the background value; otherwise, the fifth percentile of the concentration observed by the mobile observation is used as the background value. In step three, observation data with vehicle speeds less than 10 km / h and observation data inside tunnels are excluded. In step five, the characteristic variables include any two or more of the following: normalized vegetation index, land use data, building height, parking density, building density, bus stop density, road density, point of interest density, population density, urban functional areas, emission inventory data, waste treatment emissions, energy activity emissions, and agricultural emissions.

3. The method according to claim 1, characterized in that, In step six, the method includes: using Lon, Lat, NDVI, UFA, BH, POI, and RD as clustering variables, KMeans clustering is performed, and the data subsets are divided into folds according to region %5 for cross-validation in modeling; In step seven, the optimal eigenvalue set is used as the input variable and the spatially averaged ΔC data is used as the target variable. The ΔC evaluation model is constructed using the training set data, and the simulation effect is evaluated based on the determination coefficient, mean absolute error, and root mean square error of the test set.

4. The method according to claim 2, characterized in that, In step five, when the ΔC data is ΔCO2, its corresponding optimal feature set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, and UFA; when the ΔC data is ΔCH4, its corresponding optimal feature set is BD, POP, RD, POI, BH, EI, CP, BS, NDVI, UFA, LD, AG, and EN; in step seven, the partitioning is to divide the spatially averaged ΔC data and the corresponding optimal feature set into 80% training set and 20% test set, ensuring that each data subset contains both training and test sets.

5. The method according to claim 2, characterized in that, In step nine, the method includes: using the optimal feature set as the input variable and the spatially averaged ΔC data as the target variable, and reconstructing the ΔC model using all ΔC observation data and the corresponding optimal feature set; In step eight, the model mechanism is explained using the additivity feature interpretation method, the marginal contribution of the feature values ​​is quantified using Shapley values, the contribution and direction of each feature value to the model output are explained, and the key factors affecting the distribution of urban greenhouse gas concentrations are identified. If the Shapley value of a feature variable is abnormal, other basic geographical factors related to the greenhouse gas emission are added, and the process is iteratively optimized.

6. The method according to claim 5, characterized in that, In step eight, the formula for identifying the key factors affecting the distribution of urban greenhouse gas concentrations is as follows: (2); In equation (2), It is an eigenvalue i Shapley value, F It is the set of all features. S It does not include i Feature subset, It is a set S Input the value of the feature in the input field. Using feature sets S The trained model, It is the result of model prediction. In the feature set S Add eigenvalues ​​to the base i Post-trained model, It is the model's prediction result; Features i At the | S | +1 number of all permutations of the positions; Features i The remaining | F |-| S | - The number of ways a feature can be arranged; The number of all possible permutations; Features i Previous | S The number of ways each feature can be arranged; For feature subset S The number of features contained therein; For features not included i A subset of features.

7. The method according to claim 2, characterized in that, In step ten, the method includes: dividing a coarse grid and then further subdividing it into fine grids; determining a weighting coefficient based on the relative magnitude of the ΔC data in each fine grid and the sum of ΔC in the coarse grid; and redistributing the calculated total greenhouse gas emissions in the coarse grid to the fine grids to generate an emissions inventory for the fine grids.

8. The method according to claim 7, characterized in that, In step ten, the fine mesh is a fine mesh with a spatial resolution of 100 m × 100 m; the formula for calculating the weighting coefficient is as follows: (3); （4）； In equations (3) and (4), Indicates the position of the coarse grid. Indicates the location of the finer grid. Indicates the first In the coarse grid, the first The weighting factor of each fine grid, Indicates the first The total number of fine grids in a coarse grid, and , Indicates the first In the coarse grid, the first Simulated ΔC values ​​for a fine-grid array. Indicates the first Greenhouse gas emissions from a coarse grid Indicates the first In the coarse grid, the first Greenhouse gas emissions per fine grid.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images