Automatic monitoring system and distribution prediction method for greenhouse gas flux at water-air interface

Abstract

The embodiment of the application discloses a kind of water-gas interface greenhouse gas flux automatic monitoring system and distribution prediction method.Therein, system includes: wind speed sensor and air pressure temperature sensor, respectively for collecting the wind speed of air of monitoring point, and atmospheric pressure and Kelvin air temperature;Infrared sensor is used to collect the content of CO2, CH4 and N2O in the air of monitoring point;Acidity sensor and temperature sensor are used to collect the pH value and Celsius water temperature of water of monitoring point respectively;Carbon dioxide sensor is used to collect bicarbonate concentration in water of monitoring point;Methane sensor and nitrous oxide sensor are used to collect the content of dissolved greenhouse gas CH4 and N2O in water of monitoring point respectively;Computing module is used to calculate the flux of water-gas interface greenhouse gas.This embodiment can accurately monitor the water-gas interface flux of various greenhouse gases under actual conditions.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of gas flux monitoring, and in particular to an automatic monitoring system for greenhouse gas flux at a water-air interface and a distribution prediction method. Background Art

[0002] The monitoring and estimation of long-term changes in greenhouse gas fluxes in water bodies is not only of great significance to the study of the biogeochemical cycles of carbon and nitrogen and global climate change, but also provides effective data support for predicting the carbon sequestration capacity of water bodies on a longer time scale and promoting the ecological governance of the water environment.

[0003] At present, the observation equipment for water-air greenhouse gas flux in water bodies mostly focuses on carbon flux. For example, patent application CN101852723A provides an online automatic monitoring system for carbon flux at the water-air interface of inland water bodies and its monitoring method. This method lacks the flux monitoring of two important greenhouse gases in water bodies - methane and nitrous oxide, which have greater global warming potential, 25 times and 298 times that of carbon dioxide, respectively. At the same time, this method uses gas partial pressure to calculate gas flux. Since gas partial pressure refers to the pressure formed by a component in a gas mixture when the component occupies the same volume of the gas mixture at the same temperature, the gas flux calculated by gas partial pressure is assumed to be the gas flux under standard atmospheric pressure, not the gas flux at the actual pressure and temperature of the water-air interface, which is still different from the actual situation.

[0004] In addition, in the existing technology, most of the observation equipment for the flux of greenhouse gases carbon dioxide (CO2), methane (CH4) and nitrous oxide (N2O) in water bodies adopts the method of collecting water samples or extracting greenhouse gases from water bodies, and then using a gas analyzer for testing. For example, patent application CN221350771U provides a greenhouse gas emission sampling device, which changes the original water environment conditions, causing changes in the Henry equilibrium of gas-liquid exchange, destroying the inherent water-gas interface equilibrium pattern in the water body, and will cause huge errors between the measured data and the actual situation. Summary of the Invention

[0005] The embodiments of the present invention provide a water-air interface greenhouse gas flux automatic monitoring system and distribution prediction method to accurately monitor and predict the water-air interface flux of various greenhouse gases under actual conditions.

[0006] In a first aspect, an embodiment of the present invention provides a water-air interface greenhouse gas flux automatic monitoring system, which is applied to inland water bodies, and the system includes:

[0007] Wind speed sensor and air pressure and temperature sensor are used to collect wind speed of the air at the monitoring point , and atmospheric pressure and Kelvin temperature ;

[0008] Infrared sensor, used to collect the content of CO2, CH4 and N2O in the air at the monitoring point ;

[0009] Acidity sensor and temperature sensor, used to collect pH value of water at monitoring point and water temperature in Celsius ;

[0010] Carbon dioxide sensor, used to collect bicarbonate concentration in water at the monitoring point ;

[0011] Methane sensor and nitrous oxide sensor are used to collect the content of dissolved greenhouse gases CH4 and N2O in the water at the monitoring point respectively ;

[0012] The calculation module is used to calculate the flux of greenhouse gases at the water-air interface by:

[0013] according to , calculate the solubility of CO2, CH4 and N2O in surface water , water-air exchange rate , and Henry's equilibrium constant for bicarbonate concentration in water ;

[0014] Calculate the concentrations of dissolved CO2, CH4 and N2O in the water at the monitoring point according to the following formula: :

[0015]

[0016] Calculate the concentrations of CO2, CH4 and N2O in the air on the surface of the water at the monitoring point according to the following formula: :

[0017]

[0018] in, is the correction coefficient of the ideal gas state equation, which is used to correct the ideal gas equation to the real gas equation;

[0019] According to the following formula, the CO2, CH4 and N2O fluxes per unit area and time at the water-air interface of the monitoring point are calculated. :

[0020] .

[0021] In a second aspect, an embodiment of the present invention provides a method for predicting greenhouse gas flux distribution at a water-air interface, comprising:

[0022] S210, dividing the water area to be monitored into a plurality of grids;

[0023] S220, using each grid as a monitoring point, and using the above-mentioned automatic monitoring system for greenhouse gas flux at the water-air interface to monitor the flux of any greenhouse gas at the water-air interface of each grid in real time; and using the flux of each grid at the same time as a sample to construct a finite sample set;

[0024] S230, averaging the flux distribution monitored over a period of time, taking multiple extreme value grids and large gradient grids in the averaged flux distribution as multiple initial grids; and constructing a neural network model that uses the flux of the multiple initial grids at the same time as input and uses the flux distribution of all grids at the same time as output;

[0025] S240, training and testing the neural network model using the finite sample set; if the model accuracy does not meet the requirements, clustering the grids whose flux prediction values ​​differ from the monitored values ​​by more than a set threshold in the test results by geographical location, and incorporating the cluster centers into the model input;

[0026] S250, returning to S240 according to the new neural network model, and repeating this cycle until the model accuracy meets the requirements;

[0027] S260. Using a genetic algorithm, select an optimal input grid combination from the input grid set of the final model that can meet the model accuracy requirements and has the least number of grids, and use the optimal model corresponding to the optimal input grid combination to predict the water-air interface flux distribution.

[0028] In summary, the embodiments of the present invention disclose an automatic monitoring system and distribution prediction method for greenhouse gas fluxes at a water-air interface. The system can achieve the following beneficial effects:

[0029] 1. Using advanced infrared sensors, methane sensors, and nitrous oxide sensors, the system monitors the levels of two important greenhouse gases, methane and nitrous oxide, in the air and water, providing comprehensive data support for monitoring greenhouse gas fluxes in inland waters.

[0030] 2. To address the issue that the gas flux calculated using gas partial pressure is not the actual gas flux at the water-air interface pressure and temperature, this embodiment calculates the flux of each greenhouse gas based on gas concentration. This calculation incorporates the correction coefficient of the ideal gas state equation, taking into account the effect of the actual pressure and temperature at the water interface on the gas flux, thereby obtaining more accurate gas flux data.

[0031] 3. The entire system does not collect water or gas samples, does not change the original water environment, maintains the water's inherent gas-liquid exchange equilibrium, and further ensures that the gas flux data is consistent with actual conditions;

[0032] 4. A flux distribution prediction method based on the water-air interface greenhouse gas flux automatic monitoring system combines partial point monitoring with overall water area prediction, using as few water-air interface greenhouse gas flux automatic monitoring devices as possible to obtain the greenhouse gas flux distribution of the entire water area, thereby reducing the number of devices used for long-term monitoring, improving equipment utilization, and reducing the monitoring cost of greenhouse gas fluxes. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 Schematic diagram of an automatic monitoring system for greenhouse gas flux at a water-air interface provided by an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the installation of an online automatic monitoring device provided by an embodiment of the present invention;

[0036] Figure 3 This is a flow chart of a method for predicting greenhouse gas flux distribution at a water-air interface provided by an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of grid division of a water area to be monitored provided by an embodiment of the present invention;

[0038] Reference numerals:

[0039] 1. Power supply device;

[0040] 2. Warning lights;

[0041] 3. Industrial computer;

[0042] 4. Signal transmitter;

[0043] 5. Wind speed sensor;

[0044] 6. Air pressure and temperature sensor;

[0045] 7. Infrared sensor;

[0046] 8. Acidity sensor;

[0047] 9. Temperature sensor;

[0048] 10. Carbon dioxide sensor;

[0049] 11. Methane sensor;

[0050] 12. Nitrous oxide sensor. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0052] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0053] In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood broadly. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0054] Figure 1 FIG. 1 is a schematic diagram of an automatic monitoring system for greenhouse gas flux at a water-air interface provided by an embodiment of the present invention. Figure 1 As shown, the system includes an online automatic monitoring device and a calculation module.

[0055] Figure 2 This is a schematic diagram of the installation of an online automatic monitoring device provided by an embodiment of the present invention. Figure 2As shown, the online automatic monitoring device mainly includes a floating box, a power supply device 1, a series of sensors and an industrial computer 3. The power supply device 1 mainly provides power energy for the online automatic monitoring device through the photovoltaic panels above and the batteries under the panels. The warning light 2 is used to prevent collisions with water vehicles.

[0056] During use, the floating box can be fixed to the location of the water body to be monitored using reinforced concrete anchor points in the field. The float-type automatic regulator of the floating box allows the probes of the acidity sensor 8, temperature sensor 9, carbon dioxide sensor 10, methane sensor 11 and nitrous oxide sensor 12 to be located 5-10 cm below the water surface. Photovoltaic solar panels are used to power the online automatic monitoring device. The monitoring principle of greenhouse gas flux is based on the water chemical ion balance relationship and Henry's law in the water body. The sensor signals are collected and calculated by the data acquisition module of the industrial computer 3, and the result data is then output by the signal transmitter 4 of the remote communication module (wired RS485 communication interface and wireless GPRS transmission).

[0057] Before installing the online automatic monitoring device in the field, the acidity sensor 8 was first calibrated with pH buffer solutions (2.00, 4.01, 7.00, 9.21, and 10.00) to accurately measure the pH value (unit, dimensionless) of the surface water environment parameter. Then, the carbon dioxide sensor 10 was calibrated with ion electrode standards using NaHCO3 standard solutions (0.1mM, 0.5mM, 1.0mM, 5.0mM, and 10.0mM, where mM represents millimoles per liter) to accurately measure the bicarbonate content (unit, mM) in the surface water.

[0058] The online automatic monitoring device uses the precise signal (±0.01°) transmitted by the temperature sensor 9 to automatically perform temperature compensation correction on the signals of the acidity sensor 8 and the carbon dioxide sensor 10. The signal transmitted by the wind speed sensor 5 (model Pro10) The wind speed of the air at the monitoring point (unit: m / s, meters per second), the signal transmitted by the air pressure and temperature sensor 6 (model BME280) are the atmospheric pressure (unit, atm) and the Kelvin temperature (unit, ° K ), the signal transmitted by infrared sensor 7 (model SGA-900) The following are the contents of CO2, CH4 and N2O in the air at the monitoring point (unit: ppm), and the signal transmitted by the acidity sensor 8 (model: Orion 9815JP pH) The pH value of the water at the monitoring point (unit, dimensionless), the signal transmitted by the temperature sensor 9 (model PT100) is the water temperature in degrees Celsius at the monitoring point (unit: °C), and the signal transmitted by the carbon dioxide sensor 10 (model: Orion 9502BNWP) is the bicarbonate concentration in water at the monitoring point (unit, mM), the signal transmitted by the methane sensor 11 (model Franatech METS) The signal transmitted by the nitrous oxide sensor 12 (model Unisense full electrode) is the content of the greenhouse gas CH4 dissolved in water at the monitoring point (unit, ppm). It is the content of the greenhouse gas N2O dissolved in the water at the monitoring point (unit: ppm).

[0059] The calculation module can be set up on the user's computer. After receiving the real-time online automatic monitoring data sent by the communication base station, the calculation module gradually calculates the greenhouse gas flux per unit time and per unit area at the water-air interface of the water body through the following method:

[0060] S110. Calculate the solubility of CO2, CH4 and N2O in surface water (Unit, mole·L -1 ·atm -1 ), water-air exchange rate (Unit, cm·h -1 , cm / h) and Henry's equilibrium constant for bicarbonate concentration in water (Unit, atm‧L‧mol -1 ).

[0061] Specifically, the signals collected from the wind speed sensor 5 and the temperature sensor 9 are calculated to obtain the changes in the CO2, CH4 and N2O fluxes at the water-air interface caused by different physical and chemical effects. The calculation formula is:

[0062]

[0063] Where, Kelvin temperature (unit: ° K ),pass Calculated. .

[0064] S120, calculate the concentrations of dissolved CO2, CH4 and N2O in the water at the monitoring point (Unit, μM, micromolar):

[0065]

[0066] S130, calculate the concentrations of CO2, CH4 and N2O in the air on the water surface at the monitoring point (Unit, μM):

[0067]

[0068] wherein, is a correction coefficient of the ideal gas state equation, used for correcting the ideal gas equation to the actual gas equation;

[0069] S140, calculating the water-gas interface greenhouse gas flux of the water area per unit area per unit time (unit, mg·m -2 ·h -1 ):

[0070]

[0071] In summary, the embodiment provides a water-gas interface greenhouse gas flux automatic monitoring system, which can achieve the following beneficial effects:

[0072] 1. By using advanced infrared sensors, methane sensors and nitrous oxide sensors, the concentrations of two important greenhouse gases, methane and nitrous oxide, in the air and water body are monitored, providing comprehensive data support for the monitoring of inland water body greenhouse gas flux;

[0073] 2. In view of the problem that the gas flux calculated by using gas partial pressure in patent application CN101852723A “Inland water body water-gas interface carbon flux online automatic monitoring system and monitoring method thereof” is not the gas flux under the actual air pressure and temperature at the water-gas interface, the embodiment calculates the flux of each greenhouse gas by using gas concentration, introduces a correction coefficient of the ideal gas state equation when calculating the gas concentration, considers the influence of the actual air pressure and temperature at the water body interface on the gas flux, and thus obtains more accurate gas flux data;

[0074] 3. The whole system does not collect water or gas samples, does not change the original water environment, maintains the gas-liquid exchange balance mode of the water body itself, and further ensures that the gas flux data is consistent with the actual situation.

[0075] Based on the above water-gas interface greenhouse gas flux automatic monitoring system, Figure 3 is a flowchart of a water-gas interface greenhouse gas flux distribution prediction method, which can predict the greenhouse gas flux distribution of the whole water area by as few monitoring points as possible, and reduce the number of in-service equipment in long-term monitoring. As shown in Figure 3 , the method specifically includes:

[0076] S210, dividing the water area plane to be monitored into a plurality of grids.

[0077] Figure 4Taking a regular rectangle as an example, a top-down view of a water area to be monitored is shown. As shown in the figure, the plane can be divided into multiple grids for zoned monitoring. For irregularly shaped water areas, the division method is similar.

[0078] S220. Each grid is used as a monitoring point, and the above-mentioned automatic monitoring system for greenhouse gas flux at the water-air interface is installed respectively to monitor the flux of any greenhouse gas at the water-air interface of each grid in real time; and the flux of each grid at the same time is used as a sample to construct a finite sample set.

[0079] Combine Figure 2 , in each grid according to Figure 2 In the illustrated method, online automatic monitoring devices are installed separately to transmit the collected sensor data in real time. Each online automatic monitoring device can share a computing module to calculate the greenhouse gas flux of each grid at each moment in real time.

[0080] After monitoring for a period of time, for any greenhouse gas, the greenhouse gas flux of each grid at the same time in the monitoring data is taken as a sample. Multiple samples can be obtained from the data of multiple historical moments to form a finite sample set. The finiteness of the sample set is emphasized here because, for waters with larger areas, the cost of installing a set of online automatic monitoring devices for each grid is very high. Therefore, this embodiment can only achieve sample collection for a limited time while ensuring that all grids have equipment. In subsequent steps, a prediction model for the greenhouse gas flux distribution will be trained using limited samples, and the flux distribution of the entire water area will be predicted by the monitoring flux of some grids, thereby reducing the number of equipment in service for a long time and reducing the flux monitoring cost of the entire water area.

[0081] S230. Average the flux distribution monitored over a period of time, and take multiple extreme value grids and large gradient grids in the averaged flux distribution as multiple initial grids; and construct a neural network model that uses the flux of the multiple initial grids at the same time as input and the flux distribution of all grids at the same time as output.

[0082] The time period can be selected as a period of time when the wind speed, temperature, etc. are relatively stable. The flux distribution at each moment in the period is taken as a two-dimensional matrix arranged according to the geographical location of the grid (in irregular waters, the grid elements corresponding to the land position are permanently set to 0). The two-dimensional matrices of multiple moments are added and then divided by the number of moments in the period to obtain the average flux distribution in the period. From the average flux distribution, the grids where the extreme points are located and the grids where the gradient of change exceeds the set threshold are selected to form the initial set of input grids of the neural network model. For ease of description, the grids in these initial sets are referred to as initial grids in this embodiment. These grids are selected because they often represent the key positions where turning points or mutations occur in the flux distribution of the entire water area, and can provide key information for the later prediction of the flux distribution of the entire water area.

[0083] A neural network model is constructed based on this initial set. This embodiment trains the model so that, after inputting the greenhouse gas fluxes of these initial grids at the same moment, it can output the greenhouse gas flux distribution for all grids at that moment. Optionally, the model can utilize a multi-layer convolutional neural network, followed by a fully connected layer to generate results of a specific dimension. The dimensions of the input and output layers can be adjusted based on the amount of input and output data. The number of convolutional layers can be greater than seven, and the convolution kernel can use a 3x3 configuration, expanding or extracting data features layer by layer. The fully connected layer is used to bridge the dimensions of the final convolutional layer and the output layer.

[0084] S240. The neural network model is trained and tested using the finite sample set. If the model accuracy does not meet the requirements, the grids whose flux prediction values ​​differ from the monitored values ​​by more than a set threshold in the test results are clustered according to their geographical locations, and the cluster centers are included in the model input.

[0085] As mentioned above, the model needs to learn the data relationship between some grid fluxes and all grid fluxes. Due to the large amount of missing grid data, the model learning difficulty is very high. When the number of samples is limited, the accuracy of the trained model is likely to fall short of the ideal requirement. Therefore, this embodiment increases the number of input grids when the model accuracy is poor, providing more data information to the model and gradually reducing the learning difficulty.

[0086] In a specific implementation, in order to make full use of limited sample data, model training can be completed through the following steps:

[0087] Step 1: Divide the finite sample set into a first training set, a second training set, and a test set. The first training set and the second training set are used to correspond to and distinguish the two subsequent training stages. The samples in the two training sets can be the same, different, or partially the same, and this embodiment does not impose any specific restrictions.

[0088] Step 2: Use the first training set to train the neural network model once, and update the model parameters by minimizing the flux changes of adjacent grids during the training. This step corresponds to the first training stage of the model, which uses the rules followed by the input data itself to update the model parameters so that the model output has a certain regularity. It can be seen from the above system embodiment that the parameters such as gas concentration and temperature environment used to calculate the flux are continuously changing in space, so the change of greenhouse gas flux along the water-air interface should also be relatively uniform. Therefore, this embodiment updates the model parameters by constraining the flux changes of adjacent grids to be minimal, so as to minimize random changes without any rules. Optionally, the following loss function can be constructed:

[0089]

[0090] in, represents the loss function value, and Represents the grid index along the horizontal and vertical directions respectively, Indicates the Rank The flux prediction value of the column grid, represents the summation of all grids at the water-air interface (excluding land grids where the flux is permanently 0), Indicates taking the absolute value. The model parameters are updated by minimizing the loss function.

[0091] Step 3: Use the second training set to perform a second training on the model after the first training. In the second training, the model parameters are updated by minimizing the difference between the flux prediction value and the monitored value of each grid. This step corresponds to the second training phase of the model. In this phase, the actual monitored flux data is used as the true value. By constraining the model output to continuously approach the true value, the model after the first training is supervised. Optionally, the following loss function can be constructed:

[0092]

[0093] in, represents the loss function value, Indicates the Rank True value of flux for the column grid.

[0094] Furthermore, this embodiment also takes into account the situation where some grids have no true value data due to equipment failure or communication failure. In this case, although the sample data is missing data for some grids, it can still be used for secondary training of the model. It is only necessary to constrain the flux change between the grid and the adjacent grids in the loss function to minimize the flux change, and to minimize the difference between the flux prediction value and the monitored value of other grids. Optionally, the following loss function can be constructed for this type of sample:

[0095]

[0096] in, represents the loss function value, i 1 and j 1 represents the two-dimensional index of the complete grid of data, i 2 and j 2 represents the two-dimensional index of the data missing grid, Represent weight coefficients respectively. In this way, for samples with complete data, we can use Update the network parameters. For samples with missing data, you can use Update network parameters to further avoid waste of sample data.

[0097] The above two training stages first allow the model to learn the rules of the input data itself, and then perform one-to-one supervised training through the true value of each grid. Compared with directly using true value data for supervised training, the training method of this embodiment can accelerate the convergence of the model and achieve the best possible training effect using a limited sample set.

[0098] Step 4: After training, the retrained model is tested using the test set. If the model accuracy still does not meet the requirements, grids with flux predictions that differ from the monitored values ​​by more than a set threshold are clustered by location, and the centers of each cluster are included in the model input. This step is used to select grids with large prediction errors from the prediction results of each test sample if good model accuracy is still not achieved after two training phases.

[0099] Optionally, if some grids exhibit large errors across multiple test samples, these grids are clustered by geographic location to obtain multiple spatial clusters. Each spatial cluster represents a relatively close area where grids within this area repeatedly exhibit large measurement errors, indicating that the model is unable to effectively learn the data patterns of this area from the existing input grids. Therefore, it is necessary to select one or more grids from this area as model input to provide the model with effective information about this area. Optionally, the DBSCAN clustering method can be used, ultimately selecting the cluster center of each cluster as the model input.

[0100] Furthermore, when adding new model inputs, you can add convolution kernels connected to the newly added network and subsequent calculation layers in the first convolution layer based on the original model; or add a fully connected layer between the input layer and the first convolution layer to transform the new input layer dimension to the dimension of the original hidden layer; the rest of the structure in the model remains unchanged and inherits the original parameters as the initial parameters.

[0101] S250. Return to S240 based on the new neural network model, and repeat this process until the model accuracy meets the requirements.

[0102] Specifically, the new neural network model is trained and tested again using the finite sample set. If the new model accuracy still does not meet the requirements, the grids in the test results where the difference between the flux prediction value and the monitoring value exceeds the set threshold are clustered according to their geographical location, and the centers of each cluster are included in the model input to obtain an updated neural network, and then return to S240. This cycle is repeated until the final model accuracy meets the required indicators. At this point, the neural network model has sufficient valid grid data. Based on this data and the change pattern of the flux itself, it can more accurately predict the flux distribution of the entire water surface. For the sake of distinction and description, the neural network model obtained at this time is called the full model.

[0103] S260. Using a genetic algorithm, select an optimal input grid combination from the input grid set of the full model that can ensure the model accuracy meets the requirements and has the least number of grids, and use the optimal model corresponding to the optimal input grid combination to predict the water-air interface flux distribution.

[0104] As described above, during the multiple iterations of S240-S250, the model accuracy is improved by continuously increasing the number of input grids. To prevent redundancy in the final input grids, this step uses a genetic algorithm to filter the input grids of the full model to remove redundant grids. This allows the greenhouse gas flux automatic monitoring system to achieve flux distribution prediction for the entire water surface with minimal model input.

[0105] In a specific implementation, the above screening process can be completed in the following manner:

[0106] Step 1: Construct a chromosome sequence structure of equal length based on the number of grids in the full model input grid set. For example, if the full model input grids have 15, a vector of length 15 can be constructed as the chromosome sequence structure in the genetic algorithm, where each bit in the vector corresponds to a grid in the input grid set; a value of 1 indicates that the corresponding grid is retained as input to the neural network model, and a value of 0 indicates that the corresponding grid is deleted from the neural network model input grid.

[0107] Step 2: Generate the current population of the genetic algorithm based on the chromosome sequence structure. For example, a random initialization method can be used to generate multiple chromosomes to form the current population.

[0108] Step 3: Prune the full model based on each chromosome in the current population, and train the pruned model three times using the finite sample set. Specifically, for a particular chromosome, remove the grid corresponding to the bit with a value of 0 in that chromosome from the input grid of the full model, and remove all branches in the first hidden layer of the model that connect to that input grid, to obtain a new model. This new model is then trained using the finite sample set. In this embodiment, this training is referred to as three training cycles.

[0109] Step 4: Determine whether the model after three trainings meets the optimal goal. If not, select, crossover, and mutate the current population using a genetic algorithm. Select excellent chromosomes from the mutated current population to form a new current population. Return to step 3, prune the full model based on the chromosomes in the new current population, and train the pruned model three more times using the limited sample set. This cycle repeats until the model after three trainings meets the optimal goal. The optimal goal is: the model accuracy after three trainings meets the requirements, and the model with the minimum number of input grids has been obtained in the three trainings so far.

[0110] Optionally, each time a chromosome that meets the model accuracy requirements is obtained, the optimal chromosome with the smallest number of input grids can be selected from all the chromosomes that meet the model accuracy requirements. If the optimal chromosome remains unchanged after three consecutive training cycles, the model with the smallest number of input grids is considered to have been obtained. Of course, it is also possible to iterate directly to the maximum number of iterations, and after the iterations terminate, select the model that meets the model accuracy requirements and has the smallest number of input grids as the optimal model.

[0111] Furthermore, when selecting excellent chromosomes, a fitness function can be constructed based on the variation of the predicted values ​​of the air flux between each input grid in the chromosome; based on the fitness function, excellent chromosomes are selected from the mutated current population to construct a new current population. Optionally, the following fitness function can be constructed:

[0112]

[0113] in, is the fitness value, Indicates the number of extreme value grids in the flux distribution predicted by the model after three trainings, where the input grid with the current chromosome value of 1 is the extreme value grid. is the two-dimensional index of the input grid with the value 1 for the current chromosome, Indicates the number of grids with a value of 0 in the current chromosome; Represent weight coefficients respectively. The larger the fitness value, the better the chromosome.

[0114] Through the above fitness function, chromosomes with a small number of input grids and input grids covering more extreme grids and large gradient grids (these grids represent key morphological changes in flux distribution) can be preferentially selected to evolve the new current population, which helps to obtain the optimal chromosome and optimal model as quickly as possible.

[0115] Finally, after obtaining the optimal input grid combination, only the water-air interface greenhouse gas flux automatic monitoring systems for each grid in that combination are retained, while the water-air interface greenhouse gas flux automatic monitoring systems for the remaining grids are removed. The remaining systems are used to monitor the air flux of each target grid in real time and input the optimal model to obtain the real-time flux distribution for all grids. The removed equipment systems can be used for flux monitoring ensemble prediction in other water areas, improving equipment utilization and reducing greenhouse gas flux monitoring costs.

[0116] Furthermore, the above method uses a certain greenhouse gas as an example to illustrate a specific method for predicting the greenhouse gas flux distribution of an entire water area using as few monitoring points as possible. In practical applications, operations S220-S260 can be performed for each greenhouse gas. Specifically, when using the optimal model corresponding to the optimal input grid combination to predict the water-air interface flux distribution in S260, the optimal input grid combinations corresponding to each greenhouse gas are combined, and the automatic water-air interface greenhouse gas flux monitoring system for each target grid in the combined system is retained, while the automatic water-air interface greenhouse gas flux monitoring system for the remaining grids is removed. The greenhouse gas flux of each target grid is monitored in real time using the remaining systems, and the optimal models corresponding to each greenhouse gas are input separately to obtain the real-time greenhouse gas flux distribution of all grids. In this way, the flux distribution of the three greenhouse gases in an entire water area can be predicted using as few monitoring points as possible.

[0117] It should be noted that, in the above-mentioned method for predicting greenhouse gas flux distribution at the water-air interface, except for the steps of installing and dismantling the in-situ monitoring system for greenhouse gas flux at the water-air interface, the remaining steps can be automatically performed by electronic equipment.

[0118] In summary, this embodiment also provides a method for predicting greenhouse gas flux distribution at a water-air interface, which can achieve the following beneficial effects:

[0119] 1. By combining partial point monitoring with overall water area prediction, the greenhouse gas flux distribution of the entire water area can be obtained using as few automatic greenhouse gas flux monitoring devices at the water-air interface as possible, thus reducing the number of devices required for long-term monitoring, improving equipment utilization, and lowering the cost of greenhouse gas flux monitoring.

[0120] 2. Because neural network models need to learn complex flux distribution patterns based on a subset of grids, learning is challenging and the number of samples is limited, making it difficult to achieve ideal model accuracy all at once. To address this difficulty, this embodiment selectively selects representative grids from regions with large prediction errors as model input when model accuracy is poor. By increasing the number of input grids, the model is provided with more effective information, gradually reducing the learning difficulty.

[0121] 3. To address the limited availability of full-grid data samples, this embodiment uses multi-stage training to enable the model to first learn the patterns of the input data itself, and then conducts one-on-one supervised training using the monitored values ​​of each grid. Compared with directly using monitored data for supervised training, this method can accelerate model convergence and achieve the best possible training results using a limited sample set.

[0122] 4. For samples with missing data due to equipment or communication failures in some grids, this embodiment integrates uniform flux changes with monitoring values ​​and uses different loss value calculation methods for different grids. This fully utilizes the remaining data information in the sample and further avoids sample waste.

[0123] 5. After obtaining a full model with sufficient valid input grids, this embodiment uses a genetic algorithm to screen the input grids and remove redundant grids to prevent redundancy. This allows the automatic greenhouse gas flux monitoring system to predict flux distribution across the entire water surface with minimal model input. This can significantly reduce the number of devices needed and lower monitoring costs when the water surface is large and the flux distribution of each greenhouse gas needs to be predicted.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. An automatic monitoring system for greenhouse gas flux at the water-air interface, characterized in that: Applications to inland waters include: The wind speed sensor and air pressure and temperature sensor are used to collect the wind speed S of the air at the monitoring point. U1 , and atmospheric pressure S P and Kelvin temperature Sa k ; Infrared sensor, used to collect the content of CO2, CH4 and N2O in the air at the monitoring point Sa CO2 、Sa CH4 、Sa N2O ; Acidity sensor and temperature sensor are used to collect pH value of water at monitoring point S pH and water temperature in degrees Celsius Sw t ; Carbon dioxide sensor, used to collect bicarbonate concentration Sw in water at the monitoring point CO2 ; Methane sensor and nitrous oxide sensor are used to collect the content of CH4 and N2O dissolved in water at the monitoring point respectively. CH4 and Sw N2O ; The calculation module is used to calculate the flux of greenhouse gases at the water-air interface by: According to S U1 、Sw t and S pH , calculate the solubility K of CO2, CH4 and N2O in surface water CO2 , K CH4 and K N2O , water-air exchange rate k CO2 、k CH4 and k N2O , and Henry's equilibrium constant k for bicarbonate concentration in water h ; According to the following formula, calculate the concentration of dissolved CO2, CH4 and N2O in the water at the monitoring point Cw CO2 、Cw CH4 and Cw N2O : Q CO2 =Zoo CO2 ×K CO2 ×k h Q CH4 =Zoo CH4 ×K CH4 Q N2O =Zoo N2O ×K N2O According to the following formula, the concentrations of CO2, CH4 and N2O in the air on the surface of the water at the monitoring point are calculated as Ca CO2 , Ca CH4 and Ca N2O : in, is the correction coefficient of the ideal gas state equation, which is used to correct the ideal gas state equation to the real gas equation; According to the following formula, calculate the CO2, CH4 and N2O flux F at the water-air interface per unit area and per unit time in the water area to be monitored CO2 、F CH4 and F N2O : F CO2 =0.44×k CO2 ×(Cw CO2 -Ca CO2 ) F CH4 =0.16×k CH4 ×(Cw CH4 -Ca CH4 ) F N2O =0.44×k N2O ×(Cw N2O -Ca N2O )。 2. A method for predicting greenhouse gas flux distribution at a water-air interface, characterized in that: include: S210, dividing the water area to be monitored into a plurality of grids; S220, using each grid as a monitoring point and installing the automatic monitoring system for greenhouse gas flux at the water-air interface as claimed in claim 1, to monitor the flux of any greenhouse gas at the water-air interface of each grid in real time; and using the flux of each grid at the same time as a sample to construct a finite sample set; S230, averaging the flux distribution monitored over a period of time, and taking multiple extreme value grids and large gradient grids in the averaged flux distribution as multiple initial grids; and constructing a neural network model that takes the fluxes of the multiple initial grids at the same moment as input and takes the flux distribution of all grids at the same moment as output; S240, training and testing the neural network model using the finite sample set; if the accuracy of the neural network model does not meet the requirements, clustering the grids whose flux prediction values ​​differ from the monitored values ​​by more than a set threshold in the test results by geographical location, and incorporating the centers of the clusters into the neural network model input; S250, returning to S240 according to the new neural network model, and repeating this cycle until the accuracy of the neural network model meets the requirements; S260. Using a genetic algorithm, select the optimal input grid combination from the input grid set of the final neural network model that can make the neural network model meet the accuracy requirements and has the least number of grids, and use the optimal neural network model corresponding to the optimal input grid combination to predict the water-air interface flux distribution.

3. The method according to claim 2, characterized in that The training and testing of the neural network model using the finite sample set includes: Dividing the finite sample set into a first training set, a second training set and a test set; Training the neural network model once using the first training set, and updating the neural network model parameters by minimizing the flux change of adjacent grids during the training; The neural network model after the first training is trained for a second time using the second training set, wherein the parameters of the neural network model are updated by minimizing the difference between the flux prediction value and the monitoring value of each grid; The test set is used to test the neural network model after secondary training.

4. The method according to claim 3, characterized in that The second training set is used to perform a second training on the neural network model after the first training, wherein the neural network model parameters are updated by minimizing the difference between the flux prediction value and the monitoring value of each grid in the second training, including: If there are some grids with missing monitoring data in the samples in the second training set, the neural network model parameters are updated in the secondary training by minimizing the flux changes between the grids and the adjacent grids, and minimizing the differences between the flux prediction values ​​and the monitoring values ​​of other grids.

5. The method according to claim 4, characterized in that The updating of the neural network model parameters by minimizing the flux change between the partial grid and the adjacent grids and minimizing the difference between the flux prediction value and the monitoring value of other grids includes: Construct the following loss function: Among them, L3 represents the loss function value, i1 and j1 represent the two-dimensional index of the data complete grid, i2 and j2 represent the two-dimensional index of the data missing grid, and F(,) represents the flux prediction value of the corresponding grid. represents the flux monitoring value of the corresponding grid, abs represents the absolute value, and α and β represent the weight coefficients respectively.

6. The method according to claim 2, characterized in that The method of using a genetic algorithm to select an optimal input grid combination from the input grid set of the final neural network model that can ensure the accuracy of the neural network model meets the requirements and has the least number of grids includes: According to the size of the input grid set of the final neural network model, a chromosome sequence structure of equal length is constructed, wherein each numerical bit in the chromosome sequence structure corresponds one-to-one to each grid in the set, and the 1 / 0 value of the numerical bit respectively represents whether the corresponding grid is used as the input grid of the neural network model; generating a current population of a genetic algorithm according to the chromosome sequence structure; Pruning the final neural network model according to each chromosome in the current population, and training the pruned final neural network model three times using the finite sample set; Determine whether the final neural network model after three trainings meets the optimal goal. If not, generate a new current population according to the genetic algorithm and return to the pruning operation of the final neural network model until the final neural network model after three trainings meets the optimal goal. Among them, the optimal goal is that the accuracy of the final neural network model after three trainings meets the requirements, and the final neural network model with the least number of input grids has been obtained in the three trainings conducted so far.

7. The method according to claim 6, characterized in that The generating of a new current population according to the genetic algorithm comprises: According to the change pattern of the flux prediction value between each input grid in the chromosome, a fitness function is constructed; According to the fitness function, excellent chromosomes are selected from the mutated current population to construct a new current population.

8. The method according to claim 7, characterized in that The fitness function is constructed according to the change form of the flux prediction value between each input grid in the chromosome, including: Construct the following fitness function: Among them, fit is the fitness value, num(G) represents the number of extreme grids in the flux distribution predicted by the model after three trainings in which the input grids with a value of 1 in the chromosome are trained; i3 and j3 are the two-dimensional indexes of the input grids with a value of 1 in the chromosome, F(,) represents the flux prediction value of the corresponding grid, and abs represents the absolute value; num(0) represents the number of grids with a value of 0 in the chromosome; γ, δ, and ε represent the weight coefficients respectively; the larger the fitness value, the better the chromosome.

9. The method according to claim 2, characterized in that The method of using the optimal neural network model corresponding to the optimal input grid combination to predict the water-air interface flux distribution includes: retaining the water-air interface greenhouse gas flux automatic monitoring system of each target grid in the optimal input grid combination, and removing the water-air interface greenhouse gas flux automatic monitoring system of the remaining grids; The water-air interface greenhouse gas flux automatic monitoring system of each target grid is used to monitor the gas flux of each target grid in real time, and the optimal neural network model corresponding to the optimal input grid combination is input to obtain the real-time flux distribution of all grids.

10. The method according to claim 2, characterized in that include: For each greenhouse gas, operations S220 to S260 are performed; wherein, using the optimal neural network model corresponding to the optimal input grid combination to predict the water-air interface flux distribution includes: Take the union of the optimal input grid combinations corresponding to various greenhouse gases, retaining the aforementioned and centralized water-air interface greenhouse gas flux automatic monitoring systems for each target grid, and dismantling the water-air interface greenhouse gas flux automatic monitoring systems for the remaining grids; The greenhouse gas flux automatic monitoring system of the water-air interface of each target grid is used to monitor the greenhouse gas flux of each target grid in real time, and the optimal neural network model corresponding to the optimal input grid combination of various greenhouse gases is input respectively to obtain the real-time greenhouse gas flux distribution of all grids.

Images