CN115907178A - Clean ecosystem CO 2 Method for predicting exchange amount - Google Patents

Abstract

The invention discloses a clean ecosystem CO ₂ The invention discloses a prediction method of exchange capacity, which aims at the problem that the current NEE has numerous environmental influence factors which are mostly in a highly nonlinear relation and cannot be accurately predicted.

Description

A method for predicting net ecosystem CO2 exchange

Technical Field

The present invention relates to the field of environmental monitoring, and in particular to a method for predicting net ecosystem _CO2 exchange.

Background Art

_CO2 net ecosystem exchange (NEE) has received extensive attention, and accurate prediction of NEE is extremely important. However, there are many factors affecting NEE, and the relationship is highly nonlinear. Existing methods cannot accurately predict it. The present invention takes a variety of environmental driving factors into consideration, and uses the maximum mutual information coefficient MIC to screen the main controlling factors affecting NEE, constructs a database with the main controlling factors as independent variables and NEE as the target variable, and uses support vector regression machine theory and artificial bee colony optimization algorithm to establish an optimal prediction model for NEE, which effectively improves the prediction accuracy of NEE.

Summary of the invention

In order to solve the problem that the existing NEE environmental influencing factors are numerous and highly nonlinear and cannot be accurately predicted, the present invention provides a method for predicting the net ecosystem _CO2 exchange amount, which includes the following steps:

S1. Acquire continuous observation data of the eddy covariance system; the continuous observation data of the eddy covariance system includes: net ecosystem _CO2 exchange and environmental driving variable data;

S2. Preprocess the continuous observation data of the eddy covariance system to obtain effective net ecosystem _CO2 exchange and environmental driving variable data;

S3. Use the maximum mutual information coefficient method MIC to analyze the correlation between the effective net ecosystem _CO2 exchange and the environmental driving variable data, and select the top N main controlling factors with greater correlation with the net ecosystem _CO2 exchange from the environmental driving variable data; use the N main controlling factors as input variables and the corresponding net ecosystem _CO2 exchange as output variables to build an effective database;

S4, normalizing each item of data in the valid database to obtain normalized data; dividing the normalized data into a training set and a test set;

S5, establishing a support vector machine prediction model for net ecosystem _CO2 exchange, and training the support vector machine model using a training set to obtain a trained model;

S6. Optimizing the trained model using an artificial bee colony optimization algorithm to obtain an optimal prediction model;

S7, using the test set to test the optimal prediction model, and performing index evaluation on the optimal prediction model to obtain the final model;

S8. Use the final model to predict net ecosystem _CO2 exchange.

The beneficial effects provided by the present invention are: specifically considering the relationship between NEE and its main controlled environmental driving variables, using the selected main controlled factors as input variables of the NEE model, and effectively improving the prediction performance of the model.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the method of the present invention;

Figure 2 is the prediction result of the training database;

Figure 3 is the prediction result of the verification database.

DETAILED DESCRIPTION

In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

Please refer to Figure 1, which is a flow chart of the method of the present invention. The present invention provides a method for predicting the net ecosystem _CO2 exchange amount, comprising the following steps:

S1. Acquire continuous observation data of the eddy covariance system; the continuous observation data of the eddy covariance system includes: net ecosystem _CO2 exchange and environmental driving variable data;

Specifically, the present invention obtains net ecosystem CO ₂ exchange (NEE) data and environmental driving variable data (meteorological factor data) through the open-path eddy covariance system of the flux observation tower of Dajiu Lake in Shennongjia subalpine peat wetland. The meteorological factors mainly include 12 variables, including air temperature (Ta), relative humidity (RH), photon flux density (PPFD), net radiation (RN), downward shortwave radiation (SWIN), three-layer soil temperature (Ts1, Ts2, Ts3), three-layer soil moisture content (SWC ₁ , SWC ₂ , SWC ₃ ), and saturated vapor pressure difference (VPD);

S2. Preprocess the continuous observation data of the eddy covariance system to obtain effective net ecosystem _CO2 exchange and environmental driving variable data;

The data preprocessing in step S2 is specifically as follows: according to the preset threshold interval, the data outliers outside the threshold interval are eliminated. As an embodiment: the 30-min NEE of the present invention is obtained by processing the 10 Hz turbulence original data by the eddy-related flux calculation software EddyPro6.1, and its main steps include: wild point elimination, time delay correction, coordinate rotation, ultrasonic virtual temperature correction, air density effect correction and frequency response correction. For the calculated 30-min flux data, low-quality data are further eliminated according to the quality grade mark, and unreliable data caused by the contamination of the infrared gas analyzer mirror are eliminated. The NEE data is further corrected for the night friction wind speed, that is, the NEE data with a night friction wind speed less than 0.2 m/s is eliminated. The resulting non-missing data will be used as a valid data set for modeling;

Generally speaking, when removing abnormal values in step S2 of the present invention, three aspects of processing are performed: (1) removing data with a critical friction wind speed of 0.2 m/s at night; (2) removing data with precipitation greater than 0; (3) removing data other than -22.72 μmol·s ^-1 ·m ^-2 ≤ NEE ≤ 22.72 μmol·s ^-1 ·m ^-2 ; (4) removing data with a difference of more than 4 standard deviations between NEE and the average value. The effective data rate obtained after removal is 36.32%, which is higher than the average value of 35% of the international flux network (FLUXNET).

S3. Use the maximum mutual information coefficient method to analyze the correlation between the effective net ecosystem _CO2 exchange and the environmental driving variable data, and select the top N main controlling factors with the largest correlation with the net ecosystem _CO2 exchange from the environmental driving variable data; use the N main controlling factors as input and the corresponding net ecosystem _CO2 exchange as output to build an effective database;

6. Step S3 is specifically:

S31. Construct a scatter plot of the net ecosystem _CO2 exchange X and any environmental driving variable data Y in the form of data pairs D = {( _X1 , _Y1 ), ( _X2 , _Y2 ), ..., ( _Xn , _Yn ). Then divide the grids into a and b grids in the horizontal and vertical directions respectively. A set of a and b values corresponds to a division scale, and each division scale corresponds to multiple division schemes. The mutual information value of variables X and Y under each division scheme is:

S32, calculate the mutual information values under all the partitioning schemes through step S31, and take the maximum value as the maximum mutual information coefficient of the partitioning scale:

I ^* (X, Y) = max(I(X, Y));

S33: The formula in step S32 is normalized using log ₂ min(a,b) to obtain the normalized information values of X and Y on the division scale:

S33. Different a and b correspond to different M(D). The M(D) of all partitioning scales is calculated to form the information value characteristic matrix M(D) _a,b . Then the value of the maximum mutual information MIC of X and Y can be obtained by calculating the maximum value of the characteristic matrix of the information value:

Among them, B*(n)=n ^0.6 is the upper limit of the number of grids, and n is the size of the data;

S34. Calculate the MIC value between all environmental driving variable data and the net ecosystem _CO2 exchange, and give priority to environmental driving variable data with larger MIC values between them and the net ecosystem _CO2 exchange; the physical meaning of MIC is equivalent to the determination coefficient ( ^R2 ) in the traditional regression analysis method. The value range of MIC(X,Y) is [0,1]. When the MIC(X,Y) value is 1, it means that variable Y is completely correlated with variable X. When the MIC(X,Y) value is 0, it means that variable Y is completely independent of variable X.

S35. Calculate the MIC values among all environmental driving variable data, and finally determine the main controlling factors [xi _| i＝1,…,N] of NEEs.

S4, normalizing each item of data in the valid database to obtain normalized data; dividing the normalized data into a training set and a test set;

In the present invention, 75% of the normalized data set is used as the training database, and the remaining 25% is used as the validation database;

It should be noted that the following formula is used when performing normalization processing in step S4:

x'=(xx _min )/(x _max -x _min )

Wherein: x' is the normalized sample value, x is the true value of the sample, x _max is the maximum value of the sample, and x _min is the minimum value of the sample; the true value of the sample includes the data brought in from the database.

S5, establishing a support vector machine prediction model for net ecosystem _CO2 exchange, and training the support vector machine model using a training set to obtain a trained model;

It should be noted that, by combining the support vector machine theory and the training database [( _xi , j, _yj ) | i = 1, ..., N; j = 1, ..., _l1 ] obtained in step S4 ( _xi is the main control factor affecting NEE, _yi is the corresponding NEE value, _l1 is the total number of samples in the training database), a prediction model for NEE can be established, and its expression is as follows:

是非线性映射函数，w和b是待调整系数，可通过求解如下被约束的二次最优化问题得到待调整系数(w和b)：Where f(x) is the predicted value,

is a nonlinear mapping function, w and b are coefficients to be adjusted, and the coefficients to be adjusted (w and b) can be obtained by solving the following constrained quadratic optimization problem:

The constraints are:

Where c is the penalty factor, p is the sensitive loss coefficient describing the sparsity of the model solution, and ξ and ξ* are relaxation variables. The introduction of Lagrange multipliers (α, α*) can transform the optimization problem into a dual problem:

The constraints are:

Where K( _xi , _xj ) is the kernel function.

The support vector machine kernel function selects the Gaussian radial basis (RBF) kernel function that is most suitable for nonlinear prediction:

g is the kernel function parameter.

The parameter w to be adjusted can be calculated by Lagrange multipliers and kernel functions, and its expression is:

Finally, the SVR prediction model of NEE is:

S6. Optimizing the trained model using an artificial bee colony optimization algorithm to obtain an optimal prediction model;

The above three parameters (p, c, g) have an important influence on the prediction accuracy of the SVR model. It is necessary to optimize the three parameters simultaneously to obtain the optimal parameter combination. The optimal NEE prediction model is established through the optimal parameter combination. The present invention adopts the artificial bee colony optimization algorithm to optimize the parameters. The specific steps are as follows:

S61. Using the ten-fold cross validation method, the optimization objective function (f _u ) of the artificial bee colony optimization algorithm is constructed. The calculation formula is as follows:

u is the value of the three parameters (p, c, g), _fu is the cross-validation mean square error (CVMSE) of the training database, M is the number of training databases, S is the score by which the training database is divided by cross-validation, G _S is the Sth data used for validation, _yi is the actual value measured by the flux tower, and f( _xi )|u is the predicted value of NEE calculated by the SVR model when (p, c, g) is equal to u.

S62: Initialize the algorithm, set the maximum number of iterations maxCycle, the total number of bees NP, the number of food sources SN (NP/2), the dimension D, the maximum number of individual updates limit, and the value ranges of the parameters to be optimized c, g, and p, and use the artificial bee colony optimization algorithm to obtain the optimal solution of the parameter combination (p, c, g);

公式选择的解中生成NP个备选解，并通过步骤S61得到的f_u对备选解进行评估，然后通过轮盘赌的方式更新观察蜂的解；在侦查蜂阶段，如果存在某个解在迭代limit次后仍没有被更新，则采用公式u_ij＝u_imin+rand(0,1)(u_jmax-u_jmin)对其进行更新。S63: The artificial bee colony optimization algorithm includes three stages, namely the employed bee stage, the observation bee stage and the scout bee stage. In the employed bee stage, NP candidate solutions are generated according to the formula v _ij = u _ij + θ _ij (u _ij - _{u kj} ), and the candidate solutions are evaluated by f _u obtained in step S61, and then the employed bee solutions are updated by roulette. In the observation bee stage, the solution is updated according to the formula v ij = u ij + θ ij (u ij - u kj ).

NP candidate solutions are generated from the solution selected by the formula, and the candidate solutions are evaluated by f _u obtained in step S61, and then the solution of the observer bee is updated by roulette. In the scout bee stage, if a solution has not been updated after iteration limit times, it is updated by the formula u _ij =u _imin +rand(0,1)(u _jmax -u _jmin ).

S64: Record the parameter combination (p, c, g) corresponding to the minimum CVMSE in each cycle;

S65: Repeat steps S63 and S64 until the termination condition is reached, and output the parameter combination (p, c, g) corresponding to the smallest fu in all cycles, which is the optimal parameter combination.

S7, using the test set to test the optimal prediction model, and performing index evaluation on the optimal prediction model to obtain the final model;

S8. Use the final model to predict net ecosystem _CO2 exchange.

It should be noted that, when the prediction is finally performed, the normalized prediction result needs to be denormalized, and the following formula is used for the denormalization process: x=x'(x _max -x _min )+x _min .

In the method for predicting the net ecosystem _CO2 exchange amount described in the present invention, the environmental driving variables and NEE data are monitored by collecting the open-path eddy covariance system of the flux observation tower, and the MIC-SVR prediction method is adopted to effectively screen the characteristic variables of the prediction model, construct the SVR model between the environmental driving variables and NEE, and improve the prediction accuracy of NEE.

Finally, as an embodiment, the specific experimental process of the present invention is as follows:

L1. Collect data from the flux observation tower built near Lake No. 3 of Dajiu Lake in Shennongjia from January to December 2018. The data mainly include the original data of _CO2 net ecosystem exchange (NEE) and 12 environmental variables (Ta, RH, PPFD, RN, SWIN, VPD, Ts ₁ , Ts ₂ , Ts ₃ , SWC ₁ , SWC ₂ , SWC ₃ ).

L2. Perform quality control on the collected raw data, use MATLAB to eliminate the raw data according to the above method, and finally obtain the valid data set for 2018.

Table 1 Missing half-hourly NEE flux in the Dajiuhu subalpine peat wetland in Shennongjia in 2018

L3. Based on the valid data set obtained above, the correlation between environmental variables and NEE is analyzed by the maximum mutual information coefficient MIC method. The main controlling factors of NEE are selected according to the calculated MIC value and the principle described in claim 3. Finally, six main controlling factors are determined: PPFD, Ta, Ts ₃ , VPD, RH, and SWC ₃ .

Table 2 Maximum Mutual Information Coefficient (MIC) calculation results

L4, using the main control factors determined in L3 as input variables and the corresponding NEE as output variables, construct an effective database for SVR modeling [(PPFD _j , Ta _j , Ts _3j , VPD _j , RH _j , SWC _3j , y _j )|j＝1,…,l], where l is the number of samples in the database, a total of 6364 samples;

L5. Normalize each item of data in the database, use 75% of the valid data set as the training database (4773 samples), and the remaining 25% as the verification database (1591 samples);

L6, using the support vector machine machine learning algorithm, the data normalized in step L5 is used as input variables, and is introduced into the support vector regression formula to establish a support vector regression model; the vector regression formula is:

L7. Use the artificial bee colony optimization algorithm to optimize the support vector regression model established in step L6 and establish the optimal NEE prediction model.

L8. Please refer to Figure 2, which is the prediction result of the training database; the training database of L5 is brought into the NEE prediction model established by L7, the determination coefficient ^R2 is 0.91, and the prediction root mean square error Rmse is 0.011.

L9, please refer to Figure 3, which is the prediction result of the verification database; bring the training database of L5 into the NEE model established by L7 for prediction and denormalization to obtain the predicted value of NEE. The determination coefficient ^R2 is 0.85 and the prediction root mean square error Rmse is 0.025.

In summary, the beneficial effects of the present invention are as follows: the relationship between NEE and its main controlled environmental driving variables is specifically considered, and the selected main controlled factors are used as input variables of the NEE model, which effectively improves the prediction performance of the model.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. A method for predicting net ecosystem _CO2 exchange, characterized in that it comprises the following steps: S1. Acquire continuous observation data of the eddy covariance system; the continuous observation data of the eddy covariance system includes: net ecosystem _CO2 exchange and environmental driving variable data; S2. Preprocess the continuous observation data of the eddy covariance system to obtain effective net ecosystem _CO2 exchange and environmental driving variable data; S3. Use the maximum mutual information coefficient method MIC to analyze the correlation between the effective net ecosystem _CO2 exchange and the environmental driving variable data, and select the top N main controlling factors with greater correlation with the net ecosystem _CO2 exchange from the environmental driving variable data; use the N main controlling factors as input variables and the corresponding net ecosystem _CO2 exchange as output variables to build an effective database; S4, normalizing each item of data in the valid database to obtain normalized data; dividing the normalized data into a training set and a test set; S5, establishing a support vector machine prediction model for net ecosystem _CO2 exchange, and training the support vector machine model using a training set to obtain a trained model; S6. Optimizing the trained model using an artificial bee colony optimization algorithm to obtain an optimal prediction model; S7, using the test set to test the optimal prediction model, and performing index evaluation on the optimal prediction model to obtain the final model; S8. Use the final model to predict net ecosystem _CO2 exchange. 2. A method for predicting net ecosystem _CO2 exchange as claimed in claim 1, characterized in that: the data preprocessing in step S2 specifically includes: according to a preset threshold interval, eliminating data outliers outside the threshold interval. 3. The method for predicting net ecosystem _CO2 exchange according to claim 1, wherein step S3 specifically comprises: S31. Construct a scatter plot of the net ecosystem _CO2 exchange X and any environmental driving variable data Y in the form of data pairs D = {( _X1 , _Y1 ), ( _X2 , _Y2 ), ..., ( _Xn , _Yn )}, and then divide a and b grids in the horizontal and vertical directions respectively. A set of a and b values corresponds to a division scale, and each division scale corresponds to multiple division schemes. The mutual information value of variables X and Y under each division scheme is:

S32, calculate the mutual information values under all the partitioning schemes through step S31, and take the maximum value as the maximum mutual information coefficient of the partitioning scale: I ^* (X, Y) = max(I(X, Y)); S33: The formula in step S32 is normalized using log ₂ min(a,b) to obtain the normalized information values of X and Y on the division scale:

S33. Different a and b correspond to different M(D). The M(D) of all partitioning scales is calculated to form the information value characteristic matrix M(D) _a,b . Then the value of the maximum mutual information MIC of X and Y can be obtained by calculating the maximum value of the characteristic matrix of the information value:

Among them, B*(n)=n ^0.6 is the upper limit of the number of grids, and n is the size of the data; S34, calculating the MIC values between all environmental driving variable data and the net ecosystem _CO2 exchange amount, and giving priority to the environmental driving variable data with a larger MIC value between the net ecosystem _CO2 exchange amount; S35. Calculate the MIC values between all environmental driving variable data, follow the principle that the selected environmental driving variable data are independent of each other, and finally determine the main controlling factors [xi _| i＝1,…,N] of NEEs. 4. A method for predicting net ecosystem _CO2 exchange capacity as claimed in claim 1, characterized in that: the construction process of the net ecosystem _CO2 exchange capacity support vector machine prediction model in step S5 is as follows: S51. Construct the initial net ecosystem _CO2 exchange support vector machine prediction model expression:

Where f(x) is the predicted value,

is a nonlinear mapping function, w and b are coefficients to be adjusted; S52, obtaining the coefficients w and b to be adjusted by solving the following constrained quadratic optimization problem;

The constraints are:

Where c is the penalty factor, p is the insensitive loss function, ξ and ξ* are slack variables; S53, introducing Lagrange multipliers (α, α*) to transform the quadratic optimization problem into a dual problem:

The constraints are:

Where K( _xi , _xj ) is the kernel function; the support vector machine kernel function selects the Gaussian radial basis function (RBF) kernel function which is most suitable for nonlinear prediction:

g is the kernel function parameter; S54, the parameter w to be adjusted is calculated by Lagrange multiplier and kernel function, and its expression is:

Finally, the prediction model is:

5. A method for predicting net ecosystem _CO2 exchange as claimed in claim 4, characterized in that: the specific process of obtaining the optimal prediction model in step S6 is: S61. Using the ten-fold cross validation method, the optimization objective function (f _u ) of the artificial bee colony optimization algorithm is constructed. The calculation formula is as follows:

u is the value of the three parameters (p, c, g), _fu is the cross-validation mean square error (CVMSE) of the training database, M is the number of training databases, S is the fraction into which the training database is divided by cross-validation, G _S is the Sth portion of data used for validation, _yi is the actual value measured by the flux tower, and f( _xi )|u is the predicted value of net ecosystem _CO2 exchange calculated by the prediction model when (p, c, g) is equal to u; S62. Initialize the algorithm, set the maximum number of iterations maxCycle, the total number of bees NP, the number of food sources SN (NP/2), the dimension D, the maximum number of individual updates limit, and the value range of the parameters p, c, and g to be optimized. The artificial bee colony optimization algorithm can be used to obtain the optimal solution for the parameter combination (p, c, g).

Images