A method for inverting methane point source emissions from drone observations based on large eddy simulation and machine learning

Abstract

This paper proposes a method for inverting methane point source emissions from drone observations based on large eddy simulation (LES) and machine learning. By setting emission source environmental parameters during LES and performing multiple emission simulations, emission and drone sampling data are generated. Machine learning training is then used to derive a localized methane emission optimization conversion function. The resulting two-dimensional methane concentration distribution, reconstructed after optimization, is combined with a mass conservation algorithm to estimate actual methane emissions. This approach fully accounts for the differences in point source emission estimates caused by atmospheric turbulence in low-altitude environments and the different flight scheduling strategies of drones, thereby improving the accuracy of both drone sampling data and inversion results.

Description

Technical Field

[0001] The present invention belongs to the field of quantitative observation of greenhouse gases, and specifically relates to a drone observation inversion method for methane point source emissions based on large eddy simulation and machine learning. Background Art

[0002] Methane is a potent greenhouse gas with a significantly higher warming potential per unit mass than carbon dioxide. Over a 20-year timescale, its warming effect is approximately 86 times that of carbon dioxide. Its atmospheric lifetime is relatively short, at approximately 10 years. This makes methane a key target for climate change control. Compared to carbon dioxide, reducing methane emissions can produce a more rapid cooling effect. Consequently, methane emissions are receiving increasing attention in global climate governance and are widely considered an important and effective means of mitigating climate warming in the short term. Currently, the scientific community has developed a variety of observation and inversion techniques to quantify methane emissions, ranging from ground-based point source measurements to remote sensing inversion at regional and even global scales. These methods include fixed-site observations, mobile platform monitoring, and aerial and satellite remote sensing. These methods each have advantages at different spatial scales, but also face varying degrees of technical challenges and uncertainties. Furthermore, methane transport and concentration variations are most complex in the atmospheric boundary layer, from the surface to 100 meters above sea level, but current monitoring systems still have significant sampling gaps at this critical level.

[0003] In recent years, drones, as an emerging observation platform, have been playing an increasingly important role in greenhouse gas monitoring. Compared to traditional ground-based observation platforms and manned aircraft, drones offer greater flexibility and maneuverability, enabling rapid deployment and emergency response. They are particularly suitable for facility-scale monitoring of methane point sources, effectively addressing the sampling gap between the ground and 100-meter altitudes. In practical applications, drone observations primarily acquire three-dimensional methane concentration distribution information, which is used as a basis for inversion analysis to estimate point source emission intensity. A typical operation involves deploying a flight path downwind of a methane emission source. High-precision methane sensors acquire real-time concentration data, which are then combined with meteorological parameters to feed into the inversion model. The main inversion methods are the inverse Gaussian model and mass conservation. The former assumes that the methane concentration distribution follows an ideal Gaussian distribution and estimates methane emissions using this ideal model; the latter establishes an emission volume and uses wind field information to calculate the difference between the methane concentration entering and exiting the volume to estimate point source methane emissions.

[0004] However, the uncertainty of traditional drone observation inversion is still large, and its errors mainly come from two aspects. On the one hand, existing methods generally lack quantitative analysis of low-altitude atmospheric turbulence, which is a key factor affecting the diffusion and transmission process of methane plumes. The non-steady-state plume caused by turbulence will destroy the spatiotemporal continuity of the plume and intensify concentration pulsations, causing capture failure, statistical distortion, response lag and other interference to drone sampling, further leading to changes in sampling accuracy and inaccurate inversion results. The specific sampling accuracy changes are directly related to factors such as turbulence intensity, scale and dynamic characteristics of the sampling system; on the other hand, different flight strategies, such as the setting of parameters such as flight altitude, path layout, and sampling frequency, will also have a significant impact on the estimation results. These factors have not yet been included in the error analysis framework, resulting in large errors in drone sampling data and large errors in the inverted point source methane emissions results. Summary of the Invention

[0005] Figure 2 The red part in the middle represents the two-dimensional distribution of methane concentration. Figure 2 As can be seen from (a) and (b), due to the influence of low-altitude atmospheric turbulence and drone sampling strategies, there are large differences between the drone sampling two-dimensional matrix and the methane point source emission two-dimensional matrix, which leads to inaccurate inversion estimates.

[0006] The purpose of the present invention is to address the shortcomings of the existing technology and propose a drone observation inversion method for methane point source emissions based on large eddy simulation and machine learning. It fully considers the extremely important atmospheric turbulence information in the low-altitude environment and the differences in point source emission estimation results caused by different drone flight arrangement strategies, improves the accuracy of drone sampling data and the accuracy of inversion results, promotes research in the field of quantitative observation of greenhouse gases, and provides important technical support for the accurate quantification of facility-scale methane point source emissions.

[0007] First, a method for inverting methane point source emissions from drone observations based on large eddy simulation and machine learning is provided, including:

[0008] S1. Using the large eddy simulation method, set the emission source environmental parameters, conduct multiple emission simulation experiments, and generate a set of methane point source emission simulation data.

[0009] Specifically, the methane point source emission process is related to the height information of the emission source at the emission site and the meteorological environment information of the observation point. The emission source environmental parameters include the height information of the emission source and the meteorological environment information of the observation point. The emission source height information h (m) is the height of the emission source from the ground, and the meteorological environment information of the observation point includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri.

[0010] S2. Simulate the UAV in-situ sampling process in a large eddy simulation, set preliminary UAV sampling parameters, perform multiple simulated samplings, and generate a simulated UAV sampling methane concentration data set.

[0011] The preliminary UAV sampling parameters include: sampling time, preliminary sampling location, and preliminary sampling accuracy.

[0012] Multiple simulated samplings were carried out under the above-mentioned methane point source emission simulation data set using preliminary UAV sampling parameters to generate a simulated UAV sampling methane concentration data set.

[0013] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0014] S3. Based on the simulated two-dimensional concentration distribution of methane emissions and the corresponding simulated two-dimensional concentration distribution of methane sampled by drones, a convolutional neural network model is constructed. Using the machine learning algorithm, the localized methane emission optimization conversion function is trained.

[0015] S31: Through the three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with the preliminary drone sampling parameters, a large number of simulated two-dimensional concentration distributions of methane emissions and simulated two-dimensional concentration distributions of methane sampling are obtained.

[0016] Based on the three-dimensional dynamic methane concentration simulation data set, combined with the sampling position and sampling time in the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated to obtain the simulated two-dimensional concentration distribution of methane emissions.

[0017] Based on the simulated UAV sampling concentration data set and combined with preliminary UAV sampling parameters, the simulated sampling methane two-dimensional concentration distribution is calculated.

[0018] Through the above method, a large number of simulated sampled methane two-dimensional concentration distributions and simulated emitted methane two-dimensional concentration distributions are generated and collected and organized into data pairs.

[0019] S32: Data preprocessing to obtain a two-dimensional concentration distribution of simulated sampled methane after preprocessing.

[0020] The preprocessing includes outlier removal, normalization, filtering, etc.

[0021] S33: Construct a convolutional neural network model, set the conditional parameters as the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity, and use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

[0022] S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the actual drone-sampled methane two-dimensional concentration distribution. Use the localized methane emission optimization conversion function obtained in S3 to calculate the actual optimized and reconstructed methane two-dimensional concentration distribution.

[0023] S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the optimized two-dimensional methane concentration distribution obtained in S4.

[0024] Secondly, a method for inverting methane point source emissions from drone observations based on large eddy simulation and machine learning is provided, including:

[0025] S1. Using the large eddy simulation method, set the emission source environmental parameters, conduct multiple emission simulation experiments, and generate a set of methane point source emission simulation data.

[0026] Specifically, the methane point source emission process is related to the height information of the emission source at the emission site and the meteorological environment information of the observation point. The emission source environmental parameters include the height information of the emission source and the meteorological environment information of the observation point. The emission source height information h (m) is the height of the emission source from the ground, and the meteorological environment information of the observation point includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri.

[0027] S2. Simulate the UAV in-situ sampling process in a large eddy simulation (LES), set preliminary UAV sampling parameters, perform multiple simulated samplings, and generate a simulated UAV sampling methane concentration data set.

[0028] Specifically, based on actual sampling experience, a drone sampling plan can be designed to determine the drone's flight trajectory, flight speed, flight duration, etc., and preliminary drone sampling parameters can be set based on factors such as the actual flight speed, flight duration, flight trajectory, and methane plume cross-section.

[0029] The preliminary UAV sampling parameters include: sampling time, preliminary sampling location, and preliminary sampling accuracy.

[0030] Furthermore, the preliminary UAV sampling parameters can be corrected in combination with environmental parameters to obtain corresponding corrected UAV sampling parameters. By correcting the UAV sampling parameters, multiple simulated samplings are performed under the above-mentioned methane point source emission simulation data set to generate a simulated UAV sampling methane concentration data set.

[0031] The environmental disturbance parameter includes turbulence intensity.

[0032] The modified simulated drone parameters include sampling time, modified sampling position, and modified sampling accuracy.

[0033] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0034] S3. Based on the simulated two-dimensional concentration distribution of methane emissions and the corresponding simulated two-dimensional concentration distribution of methane sampled by drones, a convolutional neural network model is constructed. Using the machine learning algorithm, the localized methane emission optimization conversion function is trained.

[0035] S31: Through the three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with the preliminary drone sampling parameters, a large number of simulated two-dimensional concentration distributions of methane emissions and simulated two-dimensional concentration distributions of methane sampling are obtained.

[0036] Based on the three-dimensional dynamic methane concentration simulation data set, combined with the sampling position and sampling time in the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated to obtain the simulated two-dimensional concentration distribution of emitted methane.

[0037] Based on the simulated UAV sampling concentration data set and combined with preliminary UAV sampling parameters, the simulated sampling methane two-dimensional concentration distribution is calculated.

[0038] Through the above method, a large number of simulated sampled methane two-dimensional concentration distributions and simulated emitted methane two-dimensional concentration distributions are generated and collected and organized into data pairs.

[0039] S32: Data preprocessing to obtain a two-dimensional concentration distribution of simulated sampled methane after preprocessing.

[0040] The preprocessing includes outlier removal, normalization, filtering, etc.

[0041] S33: Construct a convolutional neural network model, set the conditional parameters to the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity, and use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

[0042] S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the actual drone-sampled methane two-dimensional concentration distribution. Use the localized methane emission optimization conversion function obtained in S3 to calculate the actual optimized and reconstructed methane two-dimensional concentration distribution.

[0043] S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the optimized two-dimensional methane concentration distribution obtained in S4.

[0044] In a third aspect, a methane point source emission drone observation inversion system based on large eddy simulation and machine learning is provided, which is used to perform the methane point source emission drone observation inversion method based on large eddy simulation and machine learning as described in any one of the first and second aspects, including:

[0045] The emission simulation module performs high-resolution numerical simulation of methane point source emissions through large eddy simulation, and obtains the relative two-dimensional concentration distribution of simulated methane emissions under different emission source environmental parameters.

[0046] The simulation sampling module is used to simulate UAV sampling. Combined with the UAV sampling parameters, the UAV sampling process is simulated in the large eddy simulation data and the relevant concentration data is obtained.

[0047] The drone sampling parameters are preliminary sampling parameters or revised sampling parameters.

[0048] The first calculation module is used to calculate the plume cross-section in the simulated methane point source emission data set to obtain the simulated methane emission two-dimensional concentration distribution and the simulated drone sampling methane two-dimensional concentration distribution.

[0049] The second calculation module is used to train the localized methane emission optimization conversion function. The localized methane emission optimization conversion function is obtained by using a machine learning algorithm, and the optimized reconstructed two-dimensional methane concentration distribution is calculated based on the simulated two-dimensional concentration distribution of methane samples.

[0050] The measurement module is used to carry out actual UAV sampling experiments to obtain the two-dimensional concentration distribution of methane actually sampled by the UAV, and calculate the two-dimensional concentration distribution of methane reconstructed after actual optimization through the second calculation module.

[0051] The inversion module is used to invert point source methane emissions. It uses the two-dimensional methane concentration distribution reconstructed after actual optimization and combines it with the mass conservation algorithm to obtain the actual methane emission estimation results.

[0052] The beneficial effects of the present invention are:

[0053] Through the large eddy simulation process, the emission source environmental parameters are set, multiple emission simulations are carried out, and a correlation model between the emission source environmental parameters, flight strategy and inversion accuracy is established. The sampling scheme can be adaptively adjusted according to real-time meteorological data (such as wind speed and atmospheric stability) to avoid systematic errors caused by improper parameter settings.

[0054] Using large eddy simulations (LES), high-resolution modeling of low-altitude atmospheric turbulence was achieved, incorporating dynamic parameters such as turbulence intensity and scale into the inversion framework. This approach overcomes the sampling failures inherent in traditional methods due to the unsteady diffusion of the plume caused by turbulence. Machine learning algorithms were also used to extract features from data subjected to turbulent interference, enabling dynamic correction of sampled data and constructing a methane emissions optimization transfer function. This approach avoids statistical distortion and response lags caused by low-altitude atmospheric turbulence.

[0055] This invention fully considers the crucial atmospheric turbulence in low-altitude environments and the variations in point source emission estimates caused by different drone flight strategies, effectively improving the accuracy of drone sampling data and inversion results. This helps accurately quantify methane point source emissions in key regions and industries (such as oil and gas fields, gas stations, and coal mines), and represents a technological innovation in methane point source emission accounting. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A process diagram of a methane point source emission UAV observation inversion method based on large eddy simulation and machine learning provided by the present invention;

[0057] Figure 2 (a) is a schematic diagram of the results of the UAV sampling two-dimensional matrix;

[0058] Figure 2 (b) is a schematic diagram of the two-dimensional emission matrix of methane point sources;

[0059] Figure 3 This is a schematic diagram comparing the methane point source emissions obtained by the inversion method of the present invention with the actual results. DETAILED DESCRIPTION

[0060] The present invention will be further described below with reference to the following examples. The following examples are provided only to facilitate understanding of the present invention. It should be noted that, without departing from the principles of the present invention, it is possible for a person skilled in the art to make various modifications to the present invention, and such improvements and modifications fall within the scope of the claims of the present invention.

[0061] The embodiment of the present application provides a method for inverting drone observations of methane point source emissions based on large eddy simulation and machine learning. It fully considers the extremely important atmospheric turbulence information in low-altitude environments, utilizes high-precision methane point source emission simulation and high-simulation drone sampling simulation, and reduces the impact of different drone flight arrangement strategies on point source emission estimation results. It will help to achieve accurate quantification of methane point source emissions and provide important technical support for realizing methane point source emission observations.

[0062] Specifically, the above method includes:

[0063] S1. Using the large eddy simulation method, set the environmental parameters of the emission source and conduct multiple emission simulation experiments to generate a methane point source emission simulation data set.

[0064] Specifically, the methane point source emission process is related to the height information of the emission source at the emission site and the meteorological environment information of the observation point. The environmental parameters of the emission source include the height information of the emission source and the meteorological environment information of the observation point. Among them, the emission source height information h (m) is the height of the emission source from the ground, and the meteorological environment information of the observation point includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri.

[0065] Among them, the emission source height information h (m) and the horizontal wind speed Ws (m / s) can be directly observed through monitoring sensors. The calculation formula of the atmospheric stability Ri is:

[0066] ;

[0067] In the formula, g is the gravitational acceleration (m / s2), T is the average temperature (K), is the vertical gradient of potential temperature (K / m), , are the vertical gradients of the meridional and zonal horizontal wind speeds respectively. When 0 < Ri < 0.25, it is weakly unstable or neutral; when Ri > 0.25, it is a stable stratification.

[0068] Select different environmental parameters to conduct multiple emission simulation experiments to generate a methane point source emission simulation data set .

[0069] S2. Simulate the in-situ sampling process of the unmanned aerial vehicle (UAV) in the large eddy simulation (LES), set the preliminary UAV sampling parameters, and conduct multiple simulated samplings to generate a simulated UAV sampling methane concentration data set.

[0070] Specifically, according to the actual sampling experience, design a UAV sampling plan to determine the UAV flight trajectory, flight speed, flight duration, etc. According to factors such as the actual flight speed, flight duration, flight trajectory, and methane plume cross-section situation, set the preliminary UAV sampling parameters. According to the sampling position and sampling time in the preliminary UAV sampling parameters, obtain the UAV's operation trajectory, deploy virtual sampling points in the LES computational domain, record the instantaneous methane concentration values at each point, and map them to the movement trajectory of the virtual UAV through numerical methods, thereby generating preliminary UAV simulated sampling data.

[0071] The preliminary UAV sampling parameters include sampling time, preliminary sampling position, and preliminary sampling accuracy.

[0072] According to the preliminary simulated UAV sampling parameter data set, obtain the methane point source emission simulation data set mentioned above A collection of simulated drone sampling concentration data under.

[0073] At this time, the simulated drone sampling concentration is: ,in is the UAV sampling time, They are the distance between the simulated sampling point and the emission point source in the preliminary sampling position, the horizontal value of the simulated sampling point, and the height value of the simulated sampling point. is the preliminary sampling accuracy corresponding to the preliminary sampling position.

[0074] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0075] Furthermore, the drone equipment model includes the sensor model, the drone flight trajectory includes the flight altitude, route pattern, sampling point spacing, etc., and the drone route pattern includes grid type, spiral type, zigzag type, etc.

[0076] Furthermore, preferably, the flight trajectory is zigzag, which can better cover the cross section of the methane plume and form a 200m 100m square section.

[0077] S3. Simulate the two-dimensional concentration distribution of methane emissions and the corresponding two-dimensional concentration distribution of methane sampled by simulated drones, build a convolutional neural network model, and use machine learning algorithms to train and obtain the localized methane emission optimization conversion function.

[0078] Furthermore, the localized methane emission optimization conversion function is a nonlinear optimization conversion function.

[0079] Among them, the three-dimensional dynamic methane concentration simulation data set and the simulated drone sampling concentration data set are the same methane point source emission simulation data set Get the corresponding data below.

[0080] Preferably, the two-dimensional concentration distribution of methane sampled by the simulated drone is used as input, and the two-dimensional concentration distribution of the simulated methane emissions is used as the target quantity. A convolutional neural network model is constructed, and a machine learning algorithm is used to train and obtain a localized methane emission conversion optimization conversion function.

[0081] S31: Through the same three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with preliminary drone sampling parameters, a large number of simulated emission methane two-dimensional concentration distribution and simulated sampling methane two-dimensional concentration distribution data pairs are obtained.

[0082] Based on the three-dimensional dynamic methane concentration simulation data set, combined with the sampling position and sampling time in the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated, and the generation time is obtained as follows: , the distance from the emission source is Simulated two-dimensional concentration distribution of methane emissions .

[0083] Based on the simulated UAV sampling concentration data set and combined with the preliminary UAV sampling parameters, the generation time of the simulated sampling is calculated as , the distance from the emission source is Simulated sampling of methane two-dimensional concentration distribution .

[0084] Among them, the simulated drone sampling concentration is: , the two-dimensional concentration distribution of methane sampled by simulated drones was obtained by calculation :

[0085] ;

[0086] Among them, k is the conversion function, emission source height h, ambient wind speed Ws, atmospheric stability Ri, sampling time , and the distance between the simulated sampling and the emission source , is the preliminary sampling accuracy corresponding to the preliminary sampling position.

[0087] Through the above method, a large number of simulated sampling methane two-dimensional concentration distributions are formed And simulate the two-dimensional concentration distribution of methane emissions , and collected and organized into data pairs.

[0088] S32: Data preprocessing, normalizing the data to obtain the two-dimensional concentration distribution of simulated sampled methane after preprocessing , and the corresponding two-dimensional concentration distribution of simulated methane emissions after preprocessing .

[0089] Specifically, the two-dimensional concentration distribution of methane was simulated and sampled. The maximum value and simulated two-dimensional concentration distribution of methane emissions The maximum value Normalize the corresponding two-dimensional matrix to obtain the two-dimensional concentration distribution of simulated sampled methane after preprocessing And the two-dimensional concentration distribution of simulated methane emissions after pretreatment ,in and The value range is (0, 1), and the simulation scale coefficient is obtained at the same time :

[0090] ;

[0091] Combined with the two-dimensional distribution of simulated methane emissions after pretreatment and simulation scale factor Forming the simulated methane emission concentration distribution after pretreatment .

[0092] S33: Construct a convolutional neural network model, set the conditional parameters to the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated sampling methane two-dimensional concentration distribution, and use the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity. Use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

[0093] Step S331: The pre-processed simulated sampled methane two-dimensional concentration distribution The simulated sampling image is used as input, multiple geometric features of the image are extracted, and display feature maps are output according to the multiple geometric features. The display feature maps are compared with the pre-processed simulated sampling methane two-dimensional concentration distribution. Splice together by channel to form a feature tensor .

[0094] Specifically, the geometric features of the image include color gradient features, edge features and texture features. The color gradient feature can be set to the methane concentration gradient, the edge feature can be set to the edge of the methane plume cross section, and the texture feature can be set to the isotropic diffusion of the methane plume.

[0095] The above multiple geometric features are output as display feature maps, recorded as color gradient features , edge features , texture features The display characteristic diagram is a two-dimensional characteristic diagram, and these display characteristic diagrams are compared with the two-dimensional concentration distribution Splice together by channel to form a feature tensor :

[0096] ;

[0097] in, is the modified simulated sampling methane two-dimensional concentration distribution, is the color gradient feature, is the edge feature, is a texture feature.

[0098] Step S332: The emission source environmental parameters and the distance between the simulated sampling and the emission point source are expanded through a multi-layer perceptron (MLP) and a fully connected layer to form a conditional feature vector, which is then fused with the feature vector to obtain the final input tensor.

[0099] Among them, the emission source environmental parameters are: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, and the distance between the simulated sampling and the emission point source is .

[0100] The above four parameters: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, the distance between the simulated sampling and the emission point source is , input a multi-layer perceptron (MLP), encoded into a low-dimensional vector , the above low-dimensional vector Expanded into feature tensor through the fully connected layer Conditional eigenvectors of the same spatial size , and then with the feature tensor Concatenate by channel and finally convert the feature tensor and the conditional eigenvector Fuse to get the final input tensor ,in Indicates the total number of channels after fusion.

[0101] Step S333: To improve the accuracy of the training simulation, the following function is selected as the loss function:

[0102] ;

[0103] in: The simulated methane emission concentration distribution after pretreatment , is the model output concentration distribution, and is the weighted coefficient of the two loss terms, For an A The matrix of B, It is a structural similarity index, which is a perceptual indicator used to measure the similarity between two images. Its specific formula on a sliding window is as follows:

[0104] ;

[0105] in: and is the average value of the image block, and is the variance of the image block, is the covariance, and is a stability constant.

[0106] Step S334: After training, a localized methane emission optimization conversion function is obtained.

[0107] The two-dimensional concentration distribution of methane in the simulated samples after preprocessing As input parameters, the simulated emission methane concentration distribution after preprocessing As the goal, the training model converges and the localized methane emission optimization conversion function is obtained. .

[0108] The localized methane emission optimization conversion function The following relationship is satisfied:

[0109] ;

[0110] Two-dimensional distribution of methane concentration in simulated emissions The calculation formula is:

[0111] ;

[0112] in, is the simulated emission methane concentration distribution after preprocessing, is the proportionality coefficient, is the maximum value of the two-dimensional distribution of simulated sampled methane concentration.

[0113] S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the two-dimensional methane concentration distribution actually sampled by the drone. Use the localized methane emission optimization conversion function obtained in S3 to calculate the optimized and reconstructed two-dimensional methane concentration distribution.

[0114] Furthermore, the emission source height information h, horizontal wind speed Ws and atmospheric stability Ri during the actual observation experiment were recorded, and the distance between the simulated sampling and the emission point source was , conduct drone sampling experiments and obtain the actual drone sampling concentration distribution , and further obtain the actual UAV sampling two-dimensional concentration distribution , perform normalization processing to obtain the normalized actual UAV sampling concentration distribution , combined with the localized methane emission optimization conversion function obtained in step S3 , the two-dimensional emission matrix reconstructed after actual optimization is calculated :

[0115] ;

[0116] ;

[0117] S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the optimized two-dimensional methane concentration distribution obtained in S4.

[0118] Further:

[0119] S51. Two-dimensional methane emission matrix reconstructed using actual optimization under the framework of mass conservation physics , and obtain the methane point source emissions estimated by a single drone observation experiment , the unit is (kg s-1):

[0120] ;

[0121] in, is the molar mass of methane, The area represented by a single sampling point can be further selected according to the sampling point parameters set in S2, for example, a single sampling point represents 40 square meters. is the universal gas constant, 、 and Represent the sampled air pressure, temperature and wind speed respectively.

[0122] After multiple drone experiments, the average estimated point source methane emissions is obtained as the final estimated point source methane emissions Q:

[0123] ;

[0124] Where m is the number of drone experiments.

[0125] Based on Example 1, Example 2 of the present application provides a more specific method for inversion of methane point source emission drone observation based on large eddy simulation and machine learning, specifically including:

[0126] S1. Using the large eddy simulation method, set the emission source environmental parameters, conduct multiple emission simulation experiments, and generate a set of methane point source emission simulation data.

[0127] Specifically, the methane point source emission process is related to the height information of the emission source at the emission site and the meteorological environment information of the observation point. The emission source environmental parameters include the height information of the emission source and the meteorological environment information of the observation point. The emission source height information h (m) is the height of the emission source from the ground, and the meteorological environment information of the observation point includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri.

[0128] Among them, the emission source height information h (m) and horizontal wind speed Ws (m / s) can be directly observed through monitoring sensors. The calculation formula of atmospheric stability Ri is:

[0129] ;

[0130] where \(g\) is the acceleration due to gravity (\(m / s^2\)), \(T\) is the average temperature (\(K\)), is the vertical gradient of potential temperature (\(K / m\)), , are the vertical gradients of the meridional and zonal horizontal wind speeds respectively. When \(0 < Ri < 0.25\), it is weakly unstable or neutral; when \(Ri > 0.25\), it is a stable stratification.

[0131] Select different environmental parameters to conduct multiple emission simulation experiments, and generate a methane point source emission simulation data set .

[0132] S2. Simulate the in-situ sampling process of the unmanned aerial vehicle (UAV) in large-eddy simulation (LES), set preliminary UAV sampling parameters, conduct multiple simulation samplings, and generate a simulated UAV sampling methane concentration data set.

[0133] Specifically, according to actual sampling experience, design a UAV sampling plan to determine the UAV flight trajectory, flight speed, flight duration, etc. Set preliminary UAV sampling parameters according to factors such as actual flight speed, flight duration, flight trajectory, and the cross-section of the methane plume.

[0134] According to the sampling position and sampling time in the preliminary UAV sampling parameters, obtain the operation trajectory of the UAV, deploy virtual sampling points in the LES calculation domain, record the instantaneous methane concentration values at each point, and map them to the movement trajectory of the virtual UAV through numerical methods, thereby generating preliminary UAV simulated sampling data.

[0135] However, during the UAV sampling process, in addition to the sampling position and sampling time related to the flight trajectory, the sampling data is also related to the UAV device model, sensor comprehensive accuracy, flight speed, flight time, environmental parameters, flight error, total uncertainty, etc. For example, the same sensor of the UAV has different sampling accuracies at different flight speeds, environmental parameters, and sampling positions, and different environmental disturbances and different sensor comprehensive accuracies will also result in different actual sampling data for the same flight trajectory.

[0136] Therefore, in order to further improve the accuracy of UAV simulated sampling and reduce the impact of flight trajectory and environmental interference on the accuracy of sampling data, reasonable sampling accuracy needs to be set in the sampling parameters.

[0137] According to the parameter sets of various UAV devices used in actual cases, set the corresponding simulated UAV parameters, obtain a preliminary simulated UAV sampling parameter data set, and simulate different UAV sampling parameters.

[0138] The preliminary UAV sampling parameters include sampling time, preliminary sampling position, and preliminary sampling accuracy.

[0139] However, during methane sampling, environmental parameters can significantly impact the accuracy of drone sampling data. Environmental interference parameters include turbulence characteristics, wind speed, wind direction, temperature, humidity, air pressure, inversion layer, atmospheric composition interference, electromagnetic and physical interference, mechanical vibration, and airflow disturbances. In addition to common characteristics such as wind speed and wind direction, which can affect the accuracy of sensors and sampling locations, the unsteady plume induced by turbulence can disrupt the spatiotemporal continuity of the plume and exacerbate concentration fluctuations, causing serious interference with drone sampling, such as capture failure, statistical distortion, and response lag, leading to changes in sampling accuracy. Specific changes in sampling accuracy are directly related to the turbulence intensity, scale, and dynamic characteristics of the sampling system. Different turbulence intensities can affect sampling accuracy and sampling location. To reduce sampling errors caused by turbulence, during simulated sampling, drone sampling parameters such as sampling time, sampling location, and sampling accuracy are selected to optimize the sampling scheme.

[0140] Furthermore, the preliminary UAV sampling parameters can be corrected in combination with environmental parameters to obtain corresponding corrected UAV sampling parameters. By correcting the UAV sampling parameters, multiple simulated samplings are performed under the above-mentioned methane point source emission simulation data set to generate a simulated UAV sampling methane concentration data set.

[0141] The modified simulated drone parameters include sampling time, modified sampling position, and modified sampling accuracy.

[0142] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0143] Preferably, the environmental parameters include turbulence characteristics, and the three-dimensional dynamic methane concentration simulation data obtained by large eddy simulation is subjected to turbulence filtering to separate large-scale concentration changes and turbulence pulsations, and output turbulence statistical characteristics. Preferably, high-frequency components related to turbulence intensity are extracted through high-pass filtering.

[0144] Preferably, the sampling position deviation caused by turbulence is corrected by combining the real-time attitude data of the UAV (such as angular velocity and acceleration) to obtain the corrected sampling position. .

[0145] Preferably, the corrected sampling position can be calculated by the following formula: :

[0146] ;

[0147] in, is the initial sampling location, is the change in the UAV acceleration caused by turbulence, The angular velocity change of the UAV caused by turbulence is calculated from the start time to the sampling time. The acceleration and angular velocity changes between , and the corrected sampling position is obtained; among them, The sampling time of the drone.

[0148] Furthermore, the influence of turbulence intensity on UAV sampling, as well as the angular velocity and acceleration of the UAV's real-time attitude data, are combined to obtain the corrected UAV sampling accuracy. .

[0149] Specifically, the corresponding turbulence statistics obtained by large eddy simulation are turbulence intensity , the UAV sampling accuracy is corrected to obtain the UAV sampling accuracy corresponding to the turbulence intensity ,in is the UAV sampling time, They are the distance between the simulated sampling point and the emission point source in the preliminary sampling position, the horizontal value of the simulated sampling point, and the height value of the simulated sampling point. They are respectively The corresponding corrected sampling position, is the UAV sampling accuracy corresponding to the turbulence intensity.

[0150] On this basis, the UAV correction sampling accuracy is obtained :

[0151] ;

[0152] in, 、 is the weight, is the initial sampling accuracy, is the turbulence intensity Affects the accuracy of drone sampling.

[0153] Based on the preliminary simulation of the drone sampling parameter data set, the above methane point source emission simulation data set was obtained. Corrected simulated drone sampling parameters under , including drone sampling time , Corrected sampling position , Corrected sampling accuracy , and perform simulated sampling based on the modified simulated drone sampling parameters to generate a simulated drone sampling concentration data set.

[0154] At this time, the simulated drone sampling concentration is , and satisfy the following relationship:

[0155] ;

[0156] in, is the UAV sampling time, They are the distance between the simulated sampling point and the emission point source in the preliminary sampling position, the horizontal value of the simulated sampling point, and the height value of the simulated sampling point. is the preliminary sampling accuracy corresponding to the preliminary sampling position, They are respectively The corresponding corrected sampling position, is the corrected sampling accuracy.

[0157] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0158] Furthermore, the drone equipment model includes the sensor model, the drone flight trajectory includes the flight altitude, route pattern, sampling point spacing, etc., and the drone route pattern includes grid type, spiral type, zigzag type, etc.

[0159] Preferably, the flight trajectory is zigzag, which can better cover the cross section of the methane plume and form a 200m 100m square section.

[0160] S3. Simulate the two-dimensional concentration distribution of methane emissions and the corresponding two-dimensional concentration distribution of methane sampled by simulated drones, build a convolutional neural network model, and use machine learning algorithms to train and obtain the localized methane emission optimization conversion function.

[0161] Furthermore, the localized methane emission optimization conversion function is a nonlinear optimization conversion function.

[0162] Among them, the three-dimensional dynamic methane concentration simulation data set and the simulated drone sampling concentration data set are the same methane point source emission simulation data set Get the corresponding data below.

[0163] Preferably, the two-dimensional concentration distribution of methane sampled by the simulated drone is used as input, and the two-dimensional concentration distribution of simulated methane emissions is used as the target quantity. A convolutional neural network model is constructed, and a machine learning algorithm is used to train and obtain a localized methane emission conversion optimization conversion function to eliminate the interference of non-steady-state plumes caused by turbulence on sampling observations.

[0164] S31: Through the same three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with preliminary drone sampling parameters, a large number of simulated emission methane two-dimensional concentration distribution and simulated sampling methane two-dimensional concentration distribution data pairs are obtained.

[0165] Based on the three-dimensional dynamic methane concentration simulation data set, combined with the sampling position and sampling time in the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated, and the generation time is obtained as follows: , the distance between the simulated sampling and the emission source is Simulated two-dimensional concentration distribution of methane emissions .

[0166] Based on the simulated UAV sampling concentration data set and combined with the preliminary UAV sampling parameters, the generation time of the simulated sampling is calculated as , the distance from the emission source is Simulated sampling of methane two-dimensional concentration distribution .

[0167] Among them, the simulated drone sampling concentration is , the two-dimensional concentration distribution of methane sampled by simulated drones was obtained by calculation :

[0168] ;

[0169] Among them, k is the conversion function, emission source height h, ambient wind speed Ws, atmospheric stability Ri, sampling time , and the distance between the simulated sampling and the emission source , is the preliminary sampling accuracy corresponding to the preliminary sampling position.

[0170] Through the above method, a large number of simulated sampling methane two-dimensional concentration distributions are formed And simulate the two-dimensional concentration distribution of methane emissions , and collected and organized into data pairs.

[0171] S32: Data preprocessing, normalizing the data to obtain the two-dimensional concentration distribution of simulated sampled methane after preprocessing , and the corresponding two-dimensional concentration distribution of simulated methane emissions after preprocessing .

[0172] Specifically, the two-dimensional concentration distribution of methane was simulated and sampled. The maximum value and simulated two-dimensional concentration distribution of methane emissions The maximum value Normalize the corresponding two-dimensional matrix to obtain the two-dimensional concentration distribution of simulated sampled methane after preprocessing And the two-dimensional concentration distribution of simulated methane emissions after pretreatment ,in and The value range is (0, 1), and the simulation scale coefficient is obtained at the same time :

[0173] ;

[0174] Combined with the two-dimensional distribution of simulated methane emissions after pretreatment and simulation scale factor Forming the simulated methane emission concentration distribution after pretreatment .

[0175] S33: Construct a convolutional neural network model, set the conditional parameters to the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated sampling methane two-dimensional concentration distribution, and use the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity. Use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

[0176] Step S331: The pre-processed simulated sampled methane two-dimensional concentration distribution The simulated sampling image is used as input, multiple geometric features of the image are extracted, and display feature maps are output according to the multiple geometric features. The display feature maps are compared with the pre-processed simulated sampling methane two-dimensional concentration distribution. Splice together by channel to form a feature tensor .

[0177] Specifically, the geometric features of the image include color gradient features, edge features and texture features. The color gradient feature can be set to the methane concentration gradient, the edge feature can be set to the edge of the methane plume cross section, and the texture feature can be set to the isotropic diffusion of the methane plume.

[0178] The above multiple geometric features are output as display feature maps, recorded as color gradient features , edge features , texture features The display characteristic diagram is a two-dimensional characteristic diagram, and these display characteristic diagrams are compared with the two-dimensional concentration distribution Splice together by channel to form a feature tensor :

[0179] ;

[0180] in, is the modified simulated sampling methane two-dimensional concentration distribution, is the color gradient feature, is the edge feature, is a texture feature.

[0181] Step S332: The emission source environmental parameters and the distance between the simulated sampling and the emission point source are expanded through a multi-layer perceptron (MLP) and a fully connected layer to form a conditional feature vector, which is then fused with the feature vector to obtain the final input tensor.

[0182] Among them, the emission source environmental parameters are: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, and the distance between the simulated sampling and the emission point source is .

[0183] The above four parameters: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, the distance between the simulated sampling and the emission point source is , input a multi-layer perceptron (MLP), encoded into a low-dimensional vector , the above low-dimensional vector Expanded into feature tensor through the fully connected layer Conditional eigenvectors of the same spatial size , and then with the feature tensor Concatenate by channel and finally convert the feature tensor and the conditional eigenvector Fuse to get the final input tensor ,in, Indicates the total number of channels after fusion.

[0184] Step S333: To improve the accuracy of the training simulation, the following function is selected as the loss function:

[0185] ;

[0186] in: The simulated methane emission concentration distribution after pretreatment , is the model output concentration distribution, and is the weighted coefficient of the two loss terms, For an A The matrix of B, It is a structural similarity index, which is a perceptual indicator used to measure the similarity between two images. Its specific formula on a sliding window is as follows:

[0187] ;

[0188] in: and is the average value of the image block, and is the variance of the image block, is the covariance, and is a stability constant.

[0189] Step S334: After training, a localized methane emission optimization conversion function is obtained.

[0190] The two-dimensional concentration distribution of methane in the simulated samples after preprocessing As input parameters, the simulated emission methane concentration distribution after preprocessing As the goal, the training model converges and the localized methane emission optimization conversion function is obtained. .

[0191] The localized methane emission optimization conversion function The following relationship is satisfied:

[0192] ;

[0193] Two-dimensional distribution of methane concentration in simulated emissions The calculation formula is:

[0194] ;

[0195] in: is the simulated emission methane concentration distribution after preprocessing, is the proportionality coefficient, is the maximum value of the two-dimensional distribution of simulated sampled methane concentration.

[0196] S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the two-dimensional methane concentration distribution actually sampled by the drone. Use the localized methane emission optimization conversion function obtained in S3 to calculate the optimized and reconstructed two-dimensional methane concentration distribution.

[0197] Furthermore, the emission source height information h, horizontal wind speed Ws and atmospheric stability Ri during the actual observation experiment were recorded, and the distance between the simulated sampling and the emission point source was , conduct drone sampling experiments and obtain the actual drone sampling concentration distribution , and further obtain the actual UAV sampling two-dimensional concentration distribution , perform normalization processing to obtain the normalized actual drone sampling concentration distribution ; Combined with the localized methane emission optimization conversion function obtained in step S3 , calculate the two-dimensional emission matrix reconstructed after actual optimization :

[0198] ;

[0199] ;

[0200] S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the optimized two-dimensional methane concentration distribution obtained in S4.

[0201] Further:

[0202] S51. Two-dimensional methane emission matrix reconstructed using actual optimization under the framework of mass conservation physics , and obtain the methane point source emissions estimated by a single drone observation experiment , the unit is (kg s-1):

[0203] ;

[0204] in: is the molar mass of methane, The area represented by a single sampling point can be further selected according to the sampling point parameters set in S2, for example, a single sampling point represents 40 square meters. is the universal gas constant, 、 and Represent the sampled air pressure, temperature and wind speed respectively.

[0205] After multiple drone experiments, the average estimated point source methane emissions is obtained as the final estimated point source methane emissions Q:

[0206] ;

[0207] Where m is the number of drone experiments.

[0208] It should be noted that the parts in this embodiment that are the same or similar to those in Example 1 can be referenced to each other and will not be described in detail in this application.

[0209] Based on Examples 1-2, Example 3 of the present application provides a more specific method for inversion of methane point source emissions by drone observation based on large eddy simulation and machine learning, specifically including:

[0210] S1. Using the large eddy simulation method, set the emission source environmental parameters, conduct multiple emission simulation experiments, and generate a set of methane point source emission simulation data.

[0211] Specifically, the methane point source emission process is related to the height information of the emission source at the emission site and the meteorological environment information of the observation point. The emission source environmental parameters include the height information of the emission source and the meteorological environment information of the observation point. The emission source height information h (m) is the height of the emission source from the ground, and the meteorological environment information of the observation point includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri.

[0212] Among them, the emission source height information h (m) and the horizontal wind speed Ws (m / s) can be directly observed through monitoring sensors. The calculation formula for the atmospheric stability Ri is as follows:

[0213] ;

[0214] In the formula, g is the acceleration due to gravity (m / s²), T is the average temperature (K), is the vertical gradient of potential temperature (K / m), , are the vertical gradients of the meridional and zonal horizontal wind speeds respectively. When 0 < Ri < 0.25, it is weakly unstable or neutral; when Ri > 0.25, it is a stable stratification.

[0215] Select different environmental parameters to conduct multiple emission simulation experiments to generate a methane point source emission simulation data set .

[0216] S2. Simulate the in-situ sampling process of the drone in the large eddy simulation (LES), set the preliminary drone sampling parameters, conduct multiple simulation samplings, and generate a simulated drone sampling methane concentration data set.

[0217] Specifically, according to the actual sampling experience, design a drone sampling plan, determine the drone flight trajectory, flight speed, flight duration, etc. According to factors such as the actual flight speed, flight duration, flight trajectory, and the cross-section of the methane plume, set the preliminary drone sampling parameters. According to the sampling position and sampling time in the preliminary drone sampling parameters, obtain the drone's operation trajectory, arrange virtual sampling points within the LES calculation domain, record the instantaneous methane concentration values at each point, and map them to the motion trajectory of the virtual drone through numerical methods, thereby generating preliminary drone simulation sampling data.

[0218] However, during the drone sampling process, in addition to the sampling position and sampling time related to the flight trajectory, the sampling data is also related to the drone device model, sensor comprehensive precision, flight speed, flight time, environmental parameters, flight error, total uncertainty, etc. For example, the same sensor of the drone has different sampling precisions at different flight speeds, environmental parameters, and sampling positions, and different environmental disturbances and different sensor comprehensive precisions will also result in different actual sampling data for the same flight trajectory.

[0219] Therefore, in order to further improve the accuracy of drone simulation sampling and reduce the impact of flight trajectory and environmental interference on the accuracy of sampling data, reasonable sampling accuracy needs to be set in the sampling parameters.

[0220] According to the various drone equipment parameter sets used in actual cases, the corresponding simulated drone parameters can be set to obtain a preliminary simulated drone sampling parameter data set to simulate different drone sampling parameters.

[0221] The preliminary UAV sampling parameters include sampling time, preliminary sampling location, and preliminary sampling accuracy.

[0222] However, during methane sampling, environmental parameters can significantly impact the accuracy of drone sampling data. Environmental interference parameters include turbulence characteristics, wind speed, wind direction, temperature, humidity, air pressure, inversion layer, atmospheric composition interference, electromagnetic and physical interference, mechanical vibration, and airflow disturbances. In addition to common characteristics such as wind speed and wind direction, which can affect the accuracy of sensors and sampling locations, the unsteady plume induced by turbulence can disrupt the spatiotemporal continuity of the plume and exacerbate concentration fluctuations, causing serious interference with drone sampling, such as capture failure, statistical distortion, and response lag, leading to changes in sampling accuracy. Specific changes in sampling accuracy are directly related to the turbulence intensity, scale, and dynamic characteristics of the sampling system. Different turbulence intensities can affect sampling accuracy and sampling location. To reduce sampling errors caused by turbulence, during simulated sampling, drone sampling parameters such as sampling time, sampling location, and sampling accuracy are selected to optimize the sampling scheme.

[0223] Furthermore, the preliminary UAV sampling parameters can be corrected in combination with environmental parameters to obtain corresponding corrected UAV sampling parameters. By correcting the UAV sampling parameters, multiple simulated samplings are performed under the above-mentioned methane point source emission simulation data set to generate a simulated UAV sampling methane concentration data set.

[0224] The modified simulated drone parameters include sampling time, modified sampling position, and modified sampling accuracy.

[0225] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0226] Preferably, the environmental parameters include turbulence characteristics. Turbulence filtering is performed on the three-dimensional dynamic methane concentration data obtained from the large eddy simulation to separate large-scale concentration variations from turbulent fluctuations and output turbulence statistical characteristics. Preferably, high-pass filtering is used to extract high-frequency components related to turbulence intensity.

[0227] Preferably, the sampling position deviation caused by turbulence is corrected by combining the real-time attitude data of the UAV (such as angular velocity and acceleration) to obtain the corrected sampling position. .

[0228] Preferably, the corrected sampling position can be calculated by the following formula: :

[0229] ;

[0230] in: is the initial sampling location, The change in the drone's acceleration caused by turbulence, The angular velocity change of the UAV caused by turbulence is calculated from the start time to the sampling time. The acceleration and angular velocity changes between the two are used to obtain the corrected sampling position, where The sampling time of the drone.

[0231] Furthermore, the influence of turbulence intensity on UAV sampling, as well as the angular velocity and acceleration of the UAV's real-time attitude data, are combined to obtain the corrected UAV sampling accuracy. .

[0232] Specifically, the corresponding turbulence statistics obtained by large eddy simulation are turbulence intensity , the UAV sampling accuracy is corrected to obtain the UAV sampling accuracy corresponding to the turbulence intensity ;in is the UAV sampling time, They are the distance between the simulated sampling point and the emission point source in the preliminary sampling position, the horizontal value of the simulated sampling point, and the height value of the simulated sampling point. They are respectively The corresponding corrected sampling position, is the UAV sampling accuracy corresponding to the turbulence intensity.

[0233] On this basis, the UAV correction sampling accuracy is obtained :

[0234] ;

[0235] in: 、 is the weight, is the initial sampling accuracy, is the turbulence intensity Affects the accuracy of drone sampling.

[0236] Based on the preliminary simulation of the drone sampling parameter data set, the above methane point source emission simulation data set was obtained. Corrected simulated drone sampling parameters under , including drone sampling time , Corrected sampling position , corrected sampling accuracy , and perform simulated sampling based on the modified simulated drone sampling parameters to generate a simulated drone sampling concentration data set.

[0237] At this time, the simulated drone sampling concentration is: , and satisfy the following relationship:

[0238] ;

[0239] in: is the UAV sampling time, They are the distance between the simulated sampling point and the emission point source in the preliminary sampling position, the horizontal value of the simulated sampling point, and the height value of the simulated sampling point. is the preliminary sampling accuracy corresponding to the preliminary sampling position, They are respectively The corresponding corrected sampling position, is the corrected sampling accuracy.

[0240] Taking the emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities, a three-dimensional dynamic methane concentration simulation data set is formed through large eddy simulation.

[0241] Furthermore, the drone equipment model includes the sensor model, the drone flight trajectory includes the flight altitude, route pattern, sampling point spacing, etc., and the drone route pattern includes grid type, spiral type, zigzag type, etc.

[0242] Preferably, the flight trajectory is zigzag, which can better cover the cross section of the methane plume and form a 200m 100m square section.

[0243] S3. Simulate the two-dimensional concentration distribution of methane emissions and the corresponding two-dimensional concentration distribution of methane sampled by simulated drones, build a convolutional neural network model, and use machine learning algorithms to train and obtain the localized methane emission optimization conversion function.

[0244] Furthermore, the localized methane emission optimization conversion function is a nonlinear optimization conversion function.

[0245] Among them, the three-dimensional dynamic methane concentration simulation data set and the simulated drone sampling concentration data set are the same methane point source emission simulation data set Get the corresponding data below.

[0246] Preferably, the two-dimensional concentration distribution of methane sampled by the simulated drone is used as input, and the two-dimensional concentration distribution of simulated methane emissions is used as the target quantity. A convolutional neural network model is constructed, and a machine learning algorithm is used to train and obtain a localized methane emission conversion optimization conversion function to eliminate the interference of non-steady-state plumes caused by turbulence on sampling observations.

[0247] S31: Through the same three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with preliminary drone sampling parameters, a large number of simulated emission methane two-dimensional concentration distribution and simulated sampling methane two-dimensional concentration distribution data pairs are obtained.

[0248] Based on the three-dimensional dynamic methane concentration simulation data set, combined with the sampling position and sampling time in the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated, and the generation time is obtained as follows: , the distance between the simulated sampling and the emission source is Simulated two-dimensional concentration distribution of methane emissions .

[0249] Based on the simulated UAV sampling concentration data set and combined with the preliminary UAV sampling parameters, the generation time of the simulated sampling is calculated as , the distance from the emission source is Simulated sampling of methane two-dimensional concentration distribution .

[0250] Among them, the simulated drone sampling concentration is: , the two-dimensional concentration distribution of methane sampled by simulated drones was obtained by calculation :

[0251] ;

[0252] Among them, k is the conversion function, emission source height h, ambient wind speed Ws, atmospheric stability Ri, sampling time , and the distance between the simulated sampling and the emission source , is the preliminary sampling accuracy corresponding to the preliminary sampling position.

[0253] Through the above method, a large number of simulated sampling methane two-dimensional concentration distributions are formed And simulate the two-dimensional concentration distribution of methane emissions , and collected and organized into data pairs.

[0254] S32: Data preprocessing, removing outliers from the data to obtain the two-dimensional concentration distribution of simulated sampled methane after preprocessing , and the corresponding two-dimensional concentration distribution of simulated methane emissions after preprocessing .

[0255] S33: Construct a convolutional neural network model, set the conditional parameters to the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated sampling methane two-dimensional concentration distribution, and use the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity. Use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

[0256] Step S331: The pre-processed simulated sampled methane two-dimensional concentration distribution The simulated sampling image is used as input, multiple geometric features of the image are extracted, and display feature maps are output according to the multiple geometric features. The display feature maps are compared with the pre-processed simulated sampling methane two-dimensional concentration distribution. Splice together by channel to form a feature tensor .

[0257] Specifically, the geometric features of the image include color gradient features, edge features and texture features. The color gradient feature can be set to the methane concentration gradient, the edge feature can be set to the edge of the methane plume cross section, and the texture feature can be set to the isotropic diffusion of the methane plume.

[0258] The above multiple geometric features are output as display feature maps, recorded as color gradient features , edge features , texture features The display characteristic diagram is a two-dimensional characteristic diagram, and these display characteristic diagrams are compared with the two-dimensional concentration distribution Splice together by channel to form a feature tensor :

[0259] ;

[0260] in: is the modified simulated sampling methane two-dimensional concentration distribution, is the color gradient feature, is the edge feature, is a texture feature.

[0261] Step S332: The emission source environmental parameters and the distance between the simulated sampling and the emission point source are expanded through a multi-layer perceptron (MLP) and a fully connected layer to form a conditional feature vector, which is then fused with the feature vector to obtain the final input tensor.

[0262] Among them, the emission source environmental parameters are: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, and the distance between the simulated sampling and the emission point source is .

[0263] The above four parameters: emission source height information h, horizontal wind speed Ws and atmospheric stability Ri, the distance between the simulated sampling and the emission point source is , input a multi-layer perceptron (MLP), encoded into a low-dimensional vector , the above low-dimensional vector Expanded into feature tensor through the fully connected layer Conditional eigenvectors of the same spatial size , and then with the feature tensor Concatenate by channel and finally convert the feature tensor and the conditional eigenvector Fuse to get the final input tensor ,in Indicates the total number of channels after fusion.

[0264] Step S333: To improve the accuracy of the training simulation, the following function is selected as the loss function:

[0265] ;

[0266] in: The simulated methane emission concentration distribution after pretreatment , is the model output concentration distribution, and is the weighted coefficient of the two loss terms, For an A The matrix of B, It is a structural similarity index, which is a perceptual indicator used to measure the similarity between two images. Its specific formula on a sliding window is as follows:

[0267] ;

[0268] in: and is the average value of the image block, and is the variance of the image block, is the covariance, and is a stability constant.

[0269] Step S334: After training, a localized methane emission optimization conversion function is obtained.

[0270] The two-dimensional concentration distribution of methane in the simulated samples after preprocessing As input parameters, the corresponding simulated methane emission concentration distribution after pretreatment As the goal, the training model converges and the localized methane emission optimization conversion function is obtained. .

[0271] The localized methane emission optimization conversion function The following relationship is satisfied:

[0272] .

[0273] S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the two-dimensional methane concentration distribution actually sampled by the drone. Use the localized methane emission optimization conversion function obtained in S3 to calculate the two-dimensional methane concentration distribution reconstructed after actual optimization.

[0274] Furthermore, the emission source height information h, horizontal wind speed Ws and atmospheric stability Ri during the actual observation experiment were recorded, and the distance between the simulated sampling and the emission point source was , conduct drone sampling experiments and obtain the actual drone sampling concentration distribution , and further obtain the actual UAV sampling two-dimensional concentration distribution , perform outlier processing and obtain the actual drone sampling concentration distribution after preprocessing , combined with the localized methane emission optimization conversion function obtained in step S3 , calculate the two-dimensional emission matrix reconstructed after actual optimization :

[0275] .

[0276] S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the optimized two-dimensional methane concentration distribution obtained in S4, significantly improving the physical consistency and generalization ability of the model.

[0277] Further:

[0278] S51. Two-dimensional methane emission matrix reconstructed using actual optimization under the framework of mass conservation physics , and obtain the methane point source emissions estimated by a single drone observation experiment , the unit is (kg s-1):

[0279] ;

[0280] in, is the molar mass of methane, The area represented by a single sampling point can be further selected according to the sampling point parameters set in S2, for example, a single sampling point represents 40 square meters. is the universal gas constant, 、 and Represent the sampled air pressure, temperature and wind speed respectively.

[0281] After multiple drone experiments, the average estimated point source methane emissions is obtained as the final estimated point source methane emissions Q:

[0282] ;

[0283] Where m is the number of drone experiments.

[0284] Figure 3 The methane emission estimation results of the same methane point source emission drone observation inversion experiment are shown in Figure 2. The black dotted line represents the methane point source emission result, and the blue and red lines represent the optimized drone observation inversion results obtained by the method given in Example 3 of the present invention and the drone observation inversion results obtained by the traditional method at different distances from the emission point. Figure 3 It can be seen that the methane emission estimation result obtained by the present invention is closer to the actual emission result and has a better effect.

[0285] Furthermore, a computer-readable storage medium is provided, which includes a stored executable program, wherein when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the aforementioned methane point source emission drone observation inversion method based on large eddy simulation and machine learning.

[0286] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0287] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0288] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements that are inherent to such process, method, article, or terminal device. In the absence of further restrictions, an element defined by the phrase "comprises a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

Claims

1. A method for inverting methane point source emissions from drone observations based on large eddy simulation and machine learning, characterized in that: include: S1. Using the large eddy simulation method, set the emission source environmental parameters, conduct multiple emission simulation experiments, and generate a set of methane point source emission simulation data; S2. Simulate the UAV in-situ sampling process in a large eddy simulation, set preliminary UAV sampling parameters, perform multiple simulated samplings, and generate a simulated UAV sampling methane concentration data set; S3. Based on the simulated two-dimensional concentration distribution of methane emissions and the corresponding two-dimensional concentration distribution of methane sampled by the simulated drone, a convolutional neural network model is constructed. Using a machine learning algorithm, a training algorithm is used to obtain a localized methane emission optimization conversion function. S4. Conduct an actual drone observation experiment of methane point source emissions at the observation point to obtain the actual drone-sampled methane two-dimensional concentration distribution. Use the localized methane emission optimization conversion function obtained in S3 to calculate the actual optimized and reconstructed methane two-dimensional concentration distribution. S5. Combining the mass conservation physics framework, the methane point source emissions are inverted based on the two-dimensional methane concentration distribution reconstructed after actual optimization obtained in S4; The emission source environmental parameters include the emission source height information and the observation point meteorological environment information, wherein the emission source height information h (m) is the height of the emission source from the ground, and the observation point meteorological environment information includes the horizontal wind speed Ws (m / s) and the atmospheric stability Ri; The calculation formula of the atmospheric stability Ri is: ； where \(g\) is the acceleration due to gravity (\(m / s^2\)), \(T\) is the average temperature (\(K\)), is the vertical gradient of potential temperature (\(K / m\)), , are the vertical gradients of the meridional and zonal horizontal wind speeds respectively. When \(0 < Ri < 0.25\), it is weakly unstable or neutral; when \(Ri>0.25\), it is a stable stratification.

2. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to claim 1 is characterized in that: The method of constructing a convolutional neural network model based on the simulated two-dimensional concentration distribution of methane emissions and the corresponding two-dimensional concentration distribution of methane sampled by a simulated drone, and using a machine learning algorithm to train and obtain a localized methane emission optimization conversion function also includes: S31: Using the three-dimensional dynamic methane concentration simulation data set and the corresponding simulated drone sampling concentration data set, combined with preliminary drone sampling parameters, a large number of simulated methane emission two-dimensional concentration distribution and simulated sampling methane two-dimensional concentration distribution data pairs are obtained; S32: Data preprocessing to obtain a two-dimensional concentration distribution of simulated sampled methane after preprocessing; S33: Construct a convolutional neural network model, set the conditional parameters as the emission source environmental parameters and the distance value between the simulated sampling and the emission point source, input the preprocessed simulated emission methane two-dimensional concentration distribution as the target quantity, and use the machine learning algorithm to train and obtain the localized methane emission optimization conversion function.

3. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to claim 2 is characterized in that: include: The three-dimensional dynamic methane concentration simulation data set is a simulation data set obtained through large eddy simulation using emission source environmental parameters and preliminary drone sampling parameters as characteristic quantities; Based on the three-dimensional dynamic methane concentration simulation data set and the preliminary UAV sampling parameters, the average methane plume vertical cross-section at the corresponding set time and set distance from the emission source is calculated to obtain the generation time: , the distance from the emission source is Simulated two-dimensional concentration distribution of methane emissions .

4. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to claim 2 is characterized in that: include: Based on the simulated UAV sampling concentration data set and combined with the preliminary UAV sampling parameters, the generation time of the simulated sampling is calculated as , the distance from the emission source is Simulated sampling of methane two-dimensional concentration distribution .

5. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to claim 2 is characterized in that: include: The preprocessing includes any one or a combination of outlier removal, normalization, and filtering.

6. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to claim 2, characterized in that: The construction of the convolutional neural network model includes: Set the loss function: ； in: The simulated methane emission concentration distribution after pretreatment , is the model output concentration distribution, and is the weighted coefficient of the two loss terms, is an A*B matrix, is the structural similarity index, which is a perceptual indicator used to measure the similarity between two images.

7. The method for inversion of methane point source emissions from drone observation based on large eddy simulation and machine learning according to any one of claims 1 to 6, characterized in that: include: The emission simulation module uses large eddy simulation to perform high-resolution numerical simulation of methane point source emissions, and obtains the relative concentration distribution of simulated methane emissions under different emission source environmental parameters; The simulation sampling module is used to simulate drone sampling. It combines drone sampling parameters to simulate the drone sampling process in large eddy simulation data and obtain the relevant simulated drone sampling methane concentration data set; The first calculation module is used to calculate the plume cross section in the simulated methane point source emission data set to obtain the simulated methane emission two-dimensional concentration distribution and the simulated UAV sampling methane two-dimensional concentration distribution; The second computing module is used to train the localized methane emission optimization conversion function. Using a machine learning algorithm, the localized methane emission optimization conversion function is obtained, and the optimized reconstructed two-dimensional methane concentration distribution is calculated based on the simulated two-dimensional concentration distribution of methane samples. The measurement module is used to conduct actual UAV sampling experiments to obtain the two-dimensional concentration distribution of methane actually sampled by the UAV, and calculate the actual optimized and reconstructed two-dimensional methane concentration distribution through the second calculation module; The inversion module is used to invert point source methane emissions. It uses the two-dimensional methane concentration distribution reconstructed after actual optimization and combines it with the mass conservation algorithm to obtain the actual methane emission estimation results.

Images