Method and device for predicting carbon emission of coal-fired power plant based on virtual sample generation

The virtual samples are generated through the Pearson correlation coefficient and t-SNE algorithm, and the carbon emission data set of coal-fired power plants is expanded, which solves the problem of low prediction accuracy caused by sparse samples, and significantly improves the accuracy of the carbon emission prediction model.

CN118967163APending Publication Date: 2024-11-15SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411055691.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The carbon emission data samples of coal-fired power plants are relatively sparse, resulting in low accuracy of the carbon emission prediction model and the inability to effectively discover information in high-dimensional space.

Method used

Pearson's correlation coefficient is used for feature selection, combining t-distributed random neighborhood embedding (t-SNE) and interpolation algorithm to generate virtual samples, expand the training data set, and integrate training through a random forest model to improve model accuracy.

Benefits of technology

By expanding the sample set, the accuracy of the carbon emission prediction model is improved, and the effectiveness of the virtual sample generation method in carbon emission data of coal-fired power plants is verified.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967163A_ABST
    Figure CN118967163A_ABST
Patent Text Reader

Abstract

The invention discloses a method and a device for predicting carbon emission of a coal-fired power plant based on virtual sample generation. The prediction method comprises the following steps: performing feature selection on production data of a coal-fired unit by using a Pearson's correlation coefficient to obtain an original data set; using the obtained original data set to initialize a random forest model; constructing a virtual sample set by using the obtained original data set and the initialized random forest model; the virtual sample set comprises input of virtual samples and output of the virtual samples; using a t distribution random neighborhood embedding and interpolation algorithm to generate the input of a virtual sample from the obtained original data set; using the obtained original data set and the generated virtual sample input to obtain the output of the virtual sample; and integrating the constructed virtual sample set and the obtained original data set, and reconstructing the random forest model. According to the method, the random forest model is trained by expanding the sample set, and the precision of the coal-fired power plant carbon emission prediction model is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a study on carbon dioxide emissions from coal-fired power plants, and in particular to a method for predicting carbon emissions from coal-fired power plants based on virtual sample generation. Background Art

[0002] Global climate change caused by greenhouse gas emissions such as carbon dioxide has become an environmental issue of common concern to countries and regions around the world. Coal-fired power generation is highly dependent on fossil energy such as coal, and its carbon emissions account for a large proportion of the carbon emissions in the current power industry. Accurate carbon emission prediction for coal-fired power plants is a prerequisite for achieving carbon asset management and carbon emission reduction in power plants.

[0003] Existing research on carbon emission prediction models can be divided into two categories: traditional prediction models and machine learning models. Traditional carbon emission prediction models include gray prediction models, time series models, regression models, STIRPTA models, etc. Traditional prediction models are mostly used for macro-level carbon emission influencing factor analysis, and cannot achieve stable and high-precision carbon emission predictions, while machine learning has significant advantages in carbon emission prediction accuracy. The prediction process is to substitute the historical carbon emission data stored by the enterprise and the historical energy consumption data associated with the historical carbon emission data into the machine learning model for iterative training, find the relationship between the historical energy consumption data and the historical carbon emission data, and then obtain the carbon emission prediction model. However, most of the current carbon emission data of coal-fired power plants are obtained through carbon accounting, in which the data such as fossil fuel quantity and elemental carbon content are derived from the annual or monthly production data of the enterprise, and the production data of the units in some months are similar, resulting in relatively scarce representative sample data, and it is often difficult to find valuable information in limited samples. These problems are called small sample problems, which will inevitably reduce the accuracy of the prediction model.

[0004] In fact, the small sample problem exists in many fields, such as biomedical engineering, manufacturing and process industries. Virtual sample generation (VSG) is an effective method to deal with the small sample problem. It mainly generates virtual samples for auxiliary model learning based on the prior knowledge of the training set data, so as to effectively expand the number of samples in the training set and enhance the sample information of the training set, thereby improving the quality of the model. The VSG method based on feature extraction has made great progress in both theoretical research and practical application. It mainly relies on feature extraction algorithms or deep networks to find feature spaces suitable for processing high-dimensional samples. When predicting carbon emissions based on multiple factors such as power generation and actual boiler evaporation, the prediction model data set contains multidimensional data. In the multidimensional space, the distribution structure of the data is very loose, and it is difficult to extract effective knowledge information. Summary of the invention

[0005] The technical problem to be solved by the present invention is to propose a method for predicting carbon emissions from coal-fired power plants that improves the accuracy of carbon emission prediction models.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] Carbon emission prediction methods for coal-fired power plants, including:

[0008] Step 1: Use the Pearson correlation coefficient to perform feature selection on the production data of coal-fired units to obtain the original data set;

[0009] Step 2: Initialize the random forest model using the original data set;

[0010] Step 3: Generate virtual sample input using t-SNE and interpolation algorithm;

[0011] Step 4: Use the original data set and the generated virtual sample input to obtain the output of the virtual sample;

[0012] Step 5: Integrate the virtual sample set and the original data set to reconstruct the random forest model.

[0013] Furthermore, in step 1, the Pearson correlation coefficient is used to perform feature selection on the production data of the coal-fired unit to obtain the original data set. Specifically, the Pearson correlation coefficient between each input feature and the monthly carbon emissions of the coal-fired unit is calculated based on the sample data set, and the feature variables with an absolute value of the correlation coefficient greater than 0.75 are selected as input features. Finally, the data set is min-max normalized to obtain the original data set.

[0014] Furthermore, in step 2, the random forest model is initialized using the original data set, specifically:

[0015] Through grid search and 5-fold cross validation, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as performance evaluation indicators. After determining the appropriate hyperparameters, the original training set D is used to divide the tain Initialize the random forest model α.

[0016] Furthermore, in step 3, the input of the virtual sample is generated using t-distributed stochastic neighbor embedding (t-SNE) and interpolation algorithm, specifically:

[0017] 1) The t-SNE algorithm is used to extract the input features of the original data set of carbon emissions from coal-fired power plants into three-dimensional space. The input features of the original data set of carbon emissions from coal-fired power plants are X train =[x1, x2, …, x8], the extracted feature is X′ train =[x′1, x′2, x′3], the main extraction process is as follows:

[0018] First, similarity calculation in high-dimensional space. For the input feature X in high-dimensional space train Each pair of data points z i and z j , calculate the similar conditional probability p j|i , the calculation formula is as follows:

[0019]

[0020] Among them, σ i The data point z i The difference of Gaussian distribution centered at .

[0021] Then calculate the joint distribution probability p of two sample points in high-dimensional space ij , the calculation formula is as follows:

[0022]

[0023] Secondly, similarity calculation in low-dimensional space. For each pair of data points z′ in low-dimensional space i and z′ j , calculate the joint probability distribution q ij , the calculation formula is as follows:

[0024]

[0025] Finally, minimize the Kullback-Leibler divergence. KL divergence is used to measure the similarity between two distributions, and the high similarity between the low-dimensional distribution and the high-dimensional distribution is achieved by minimizing the KL divergence. The calculation formula is as follows:

[0026]

[0027] Optimize by gradient descent method, iteratively update the data point position in low-dimensional space until convergence, and obtain the extracted feature X′ train ;

[0028] 2) In the three-dimensional space, a new virtual sample input is generated using the nearest neighbor interpolation algorithm, that is, the mean of the mapping point z′ and the k points closest to it is calculated, and the mean is inserted into the three-dimensional space as a new virtual feature, where the two mapping points z′ in the three-dimensional space are i , z′ j The distance between them is calculated as follows:

[0029]

[0030] In the extracted feature X′ train In the example, k points adjacent to the mapping point z′ are selected one by one, and the average values ​​of each dimension are calculated. The average values ​​of these points are used as virtual features x′ virpInsert to ensure the rationality of the inserted data. The calculation formula is as follows:

[0031]

[0032] Repeat this step to obtain the three-dimensional virtual sample input set X′ vir =[x′ vir1 , x′ vir2 , x′ vir3 ];

[0033] 3) Assume that the input of the high-dimensional virtual sample is X vir =[x vir1 , x vir2 , …, x vir8 ], the input of the virtual sample is reconstructed by the random forest algorithm, and the input variable of each reconstruction model is the extracted low-dimensional feature X′ train , the output is a virtual sample X of the original high-dimensional space train Therefore, 8 models need to be established to reconstruct the input features of carbon emission data of coal-fired power plants. By inputting the reconstruction model β with the original space virtual sample, the extracted low-dimensional features can be mapped to the corresponding input feature data of carbon emission data of coal-fired power plants, and then the three-dimensional space virtual sample input X′ vir Get the corresponding original space virtual sample input X vir .

[0034] Furthermore, in step 4, the output of the virtual sample is obtained by using the original data set and the generated virtual sample input, specifically:

[0035] According to the random forest model α constructed in step 2, X vir As the input of the constructed model, the virtual sample output Y vir, Get the virtual sample set D vir .

[0036] Furthermore, in step 5, the virtual sample set and the original data set are integrated to reconstruct the random forest model, specifically:

[0037] The generated virtual sample D vir Compared with the original training sample D train Integration to form a mixed sample data set D mix Use D mix Train a random forest model α′.

[0038] The present invention also provides a coal-fired power plant carbon emission prediction device, comprising:

[0039] one or more processors;

[0040] A memory for storing one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the provided method for predicting carbon emissions from coal-fired power plants based on virtual sample generation.

[0042] The present invention also provides a storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the provided method for predicting carbon emissions from coal-fired power plants based on virtual sample generation are implemented.

[0043] Advantages of the invention:

[0044] The present invention provides a method for predicting carbon emissions from coal-fired power plants based on virtual sample generation, the method comprising using the Pearson correlation coefficient to perform feature selection on the production data of coal-fired units to obtain an original data set, using the original data set to initialize a random forest model, using t-SNE and an interpolation algorithm to generate virtual sample inputs, using the original data set and the generated virtual sample inputs to obtain virtual sample outputs, and integrating the virtual sample set and the original data set to reconstruct the random forest model. Feature extraction is performed on high-dimensional data using a t-distributed random neighbor embedding algorithm, and the characteristics between data are maintained while the data dimension is reduced, so that data points with similar distances in high-dimensional space remain similar in low-dimensional space, and then virtual samples that conform to the original sample distribution are generated by an interpolation algorithm. The present invention can improve the accuracy of the carbon emission prediction model for coal-fired power plants by training a random forest model with an expanded sample set, and the experimental results on the carbon emission data set of coal-fired power plants verify the effectiveness of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flowchart of the carbon emission prediction model for coal-fired power plants based on t-SNE-VSG;

[0046] Figure 2 Virtual feature interpolation map;

[0047] Figure 3 Original space virtual sample input reconstruction algorithm. DETAILED DESCRIPTION

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment provides a method for predicting carbon emissions from coal-fired power plants based on virtual sample generation, comprising the following steps:

[0050] Step 1: Obtain the original dataset of carbon emissions from coal-fired power plants

[0051] The production data used are derived from the monthly power generation report of a power generation company from 2020 to 2022, with a total of 118 valid monthly data samples and 46 dimensions of original variables. The monthly carbon emissions of each unit are calculated based on the "Guidelines for Enterprise Greenhouse Gas Emissions Accounting Methods and Reporting-Power Generation Facilities". In order to analyze the correlation between the input features and the carbon emissions of coal-fired power plants and screen out features with high correlation, the Pearson correlation coefficient between each input feature and the monthly carbon emissions of the coal-fired unit is calculated based on the sample data set, and the feature variables with an absolute value of the correlation coefficient greater than 0.75 are screened as input features. The results are shown in Table 1. After feature selection, the data set contains 8 dimensions of input features and 1 dimension of output features, namely the calculated value of the monthly carbon emissions of the coal-fired unit. Finally, the data set is min-max normalized to obtain the original data set.

[0052] Table 1 Pearson correlation coefficients between characteristic variables and carbon emissions of coal-fired units

[0053]

[0054] Step 2: Initialize the random forest model using the original dataset

[0055] The input features of the original data set D of coal-fired power plant carbon emissions, including the on-grid power, actual boiler evaporation, plant power supply, desalted water supplied to the boiler, average utilization hours of power generation equipment, power plant power consumption, power generation and fly ash production (excluding water), are recorded as X, and the monthly carbon emissions of coal-fired units are recorded as Y. Through grid search and 5-fold cross validation, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as performance evaluation indicators. After determining the appropriate hyperparameters, the original training set D is used to divide the data into two groups. tain Initialize the random forest model α.

[0056] Step 3: Generate virtual sample input using t-SNE and interpolation algorithms

[0057] 1) The t-SNE algorithm is used to extract the input features of the original data set of carbon emissions from coal-fired power plants into three-dimensional space. The input features of the original data set of carbon emissions from coal-fired power plants are X train =[x1, x2, …, x8], the extracted feature is X′ train =[x′1, x′2, x′3], the main extraction process is as follows:

[0058] First, similarity calculation in high-dimensional space. For the input feature X in high-dimensional space train Each pair of data points z i and z j , calculate the similar conditional probability p j|i , the calculation formula is as follows:

[0059]

[0060] Among them, σ i The data point z i The difference of Gaussian distribution centered at .

[0061] Then calculate the joint distribution probability pij of two sample points in the high-dimensional space. The calculation formula is as follows:

[0062]

[0063] Secondly, similarity calculation in low-dimensional space. For each pair of data points z′ in low-dimensional space i and z′ j , calculate the joint probability distribution q ij , the calculation formula is as follows:

[0064]

[0065] Finally, minimize the Kullback-Leibler divergence. KL divergence is used to measure the similarity between two distributions, and the high similarity between the low-dimensional distribution and the high-dimensional distribution is achieved by minimizing the KL divergence. The calculation formula is as follows:

[0066]

[0067] Optimize by gradient descent method, iteratively update the data point position in low-dimensional space until convergence, and obtain the extracted feature X′ train ;

[0068] 2) In the three-dimensional space, a new virtual sample input is generated using the nearest neighbor interpolation algorithm, that is, the mean of the mapping point z′ and the k points closest to it is calculated, and the mean is inserted into the three-dimensional space as a new virtual feature, where the two mapping points z′ in the three-dimensional space are i , z′ j The distance between them is calculated as follows:

[0069]

[0070] In the extracted feature X′ train In the above example, five points adjacent to the mapping point z′ are selected one by one, and the average values ​​of each dimension are calculated. The average values ​​of these points are used as the virtual feature x′. virp Insert to ensure the rationality of the inserted data. The calculation formula is as follows:

[0071]

[0072] Repeat this step to obtain the three-dimensional virtual sample input set X′ vir =[x′ vir1 , x′vir2 , x′ vir3 ]. Figure 2 The data origin and interpolation points in the three-dimensional space are shown. The red points are the original sample points of the coal-fired power plant production data mapped using t-SNE, and the blue points are the virtual sample points generated by interpolation based on the mapped data origin.

[0073] 3) Assume that the input of the high-dimensional virtual sample is X vir =[x vir1 , x vir2 , …, x vir8 ]. The input of the virtual sample is reconstructed by the random forest algorithm, and the input variable of each reconstruction model is the extracted low-dimensional feature X′ train , the output is a virtual sample X of the original high-dimensional space train Therefore, 8 models need to be established to reconstruct the input features of carbon emission data of coal-fired power plants. By inputting the reconstruction model β with the original space virtual sample, the extracted low-dimensional features can be mapped to the corresponding input feature data of carbon emission data of coal-fired power plants, and then the three-dimensional space virtual sample input X′ vir Get the corresponding original space virtual sample input X vir The structure of the original space virtual sample input reconstruction algorithm is as follows: Figure 3 shown.

[0074] Step 4: Use the original data set and the generated virtual sample input to obtain the output of the virtual sample

[0075] According to the random forest model α constructed in step 2, X vir As the input of the constructed model, the virtual sample output Y vir . Get the virtual sample set D vir .

[0076] Step 5: Integrate the virtual sample set and the original data set to reconstruct the random forest model

[0077] The generated virtual sample D vir Compared with the original training sample D train Integration to form a mixed sample data set D mix Use D mix Train a random forest model α′.

[0078] The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to evaluate the performance of the model. The original data set is divided by 5-fold cross validation, and different numbers of virtual samples are integrated into the training data set each time to obtain the average results of each model. The results of Ave(RMSE) and Ave(MAPE) are shown in Table 2:

[0079] Table 2 Model evaluation with different numbers of virtual samples

[0080]

[0081] It can be seen from Table 2 that the coal-fired power plant carbon emission prediction model trained after adding virtual samples has higher accuracy.

[0082] Example 2

[0083] This embodiment provides a device for predicting carbon emissions from a coal-fired power plant, comprising:

[0084] one or more processors;

[0085] A memory for storing one or more programs;

[0086] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the coal-fired power plant carbon emission prediction method based on virtual sample generation provided in Example 1.

[0087] Example 3

[0088] The present invention also provides a storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the method for predicting carbon emissions from coal-fired power plants based on virtual sample generation provided in Example 1 are implemented.

Claims

1. Carbon emission prediction method for coal-fired power plants based on virtual sample generation, It is characterized in that include: Step 1: Use the Pearson correlation coefficient to perform feature selection on the production data of coal-fired units to obtain the original data set; Step 2: Initialize the random forest model using the original data set obtained in step 1; Step 3: Using the original data set obtained in step 1 and the random forest model initialized in step 2, a virtual sample set is constructed; the virtual sample set includes the input of the virtual sample and the output of the virtual sample; Generate virtual sample input from the original data set obtained in step 1 using the t-distributed random neighborhood embedding and interpolation algorithm; Using the original data set obtained in step 1 and the generated virtual sample input, the output of the virtual sample is obtained; Step 4: Integrate the virtual sample set constructed in step 3 and the original data set obtained in step 1 to reconstruct the random forest model.

2. The method for predicting carbon emissions from coal-fired power plants based on virtual sample generation according to claim 1, It is characterized in that In step 1, the Pearson correlation coefficient is used to perform feature selection on the production data of coal-fired units to obtain the original data set, specifically: Based on the sample data set, the Pearson correlation coefficient between each input feature and the monthly carbon emissions of coal-fired units was calculated, and the characteristic variables with an absolute value of the correlation coefficient greater than 0.75 were selected as input features. Finally, the data set was min-max normalized to obtain the original data set.

3. The method for predicting carbon emissions from coal-fired power plants based on virtual sample generation according to claim 1, It is characterized in that In step 2, the random forest model is initialized using the original data set, specifically: Through grid search and 5-fold cross validation, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as performance evaluation indicators. After determining the appropriate hyperparameters, the original training set D is used to divide the tain Initialize the random forest model α.

4. The method for predicting carbon emissions from coal-fired power plants based on virtual sample generation according to claim 1, It is characterized in that In step 3, the input of the virtual sample is generated using the t-distributed random neighborhood embedding and interpolation algorithm, specifically: 1) The t-SNE algorithm is used to extract the input features of the original data set of carbon emissions from coal-fired power plants into three-dimensional space. The input features of the original data set of carbon emissions from coal-fired power plants are X train =[x 1 ,x 2 ,…,x 8 ], the extracted feature is X t ′ rain =[x 1 ′ ,x ′ 2 ,x ′ 3 ], the main process of extraction is as follows: First, similarity calculation in high-dimensional space. For the input feature X in high-dimensional space train Each pair of data points z i and z j , calculate the similar conditional probability p j|i , the calculation formula is as follows: Among them, σ i The data point z i The difference of Gaussian distribution centered at . Then calculate the joint distribution probability p of two sample points in high-dimensional space ij , the calculation formula is as follows: Secondly, similarity calculation in low-dimensional space. For each pair of data points z in low-dimensional space i ′ and z j ′ , calculate the joint probability distribution q ij , the calculation formula is as follows: Finally, minimize the Kullback-Leibler divergence. KL divergence is used to measure the similarity between two distributions, and the high similarity between the low-dimensional distribution and the high-dimensional distribution is achieved by minimizing the KL divergence. The calculation formula is as follows: Optimize by gradient descent method, iteratively update the position of data points in low-dimensional space until convergence, and obtain the extracted feature X t ′ rain ; 2) In three-dimensional space, a new virtual sample input is generated using the nearest neighbor interpolation algorithm, that is, the mapping point z is calculated ′ and the mean of the k points closest to it, and insert the mean into the three-dimensional space as a new virtual feature, where the two mapping points z in the three-dimensional space i ′ , z j ′ The distance between them is calculated as follows: In the extracted feature X t ′ rain In the example, select and map points z one by one ′ For k adjacent points, calculate the average value of each dimension, and use the average value of these points as the virtual feature x ′ virp Insert, the calculation formula is as follows: Repeat this step to obtain the three-dimensional virtual sample input set X v ′ ir =[x ′ vir1 ,x ′ vir2 ,x ′ vir3 ]; 3) Input X through a three-dimensional virtual sample v ′ ir Get the corresponding original space virtual sample input X vir , X vir =[x vir1 ,x vir2 ,…,x vir8 ], the three-dimensional virtual sample input is reconstructed by the random forest algorithm, and the input variable of the training reconstruction model is the extracted low-dimensional feature X t ′ rain , the output is a virtual sample X of the original high-dimensional space train Therefore, 8 models need to be established to reconstruct the input features of carbon emission data of coal-fired power plants. The reconstruction model β is input through the original space virtual sample, and the extracted low-dimensional features are mapped to the corresponding input feature data of carbon emission data of coal-fired power plants, and then the three-dimensional space virtual sample input X v ′ ir Get the corresponding original space virtual sample input X vir。 5. The method for predicting carbon emissions from coal-fired power plants based on virtual sample generation according to claim 4, It is characterized in that In step 3, the output of the virtual sample is obtained by using the original data set and the generated virtual sample input, specifically: According to the random forest model α constructed in step 2, X vir As the input of the constructed model, the virtual sample output Y is obtained vir , get the virtual sample set D vii .

6. The method for predicting carbon emissions from coal-fired power plants based on virtual sample generation according to claim 5, It is characterized in that In step 4, the virtual sample set constructed in step 3 and the original data set obtained in step 1 are integrated to reconstruct the random forest model, specifically: The generated virtual sample D iii Compared with the original training sample D irain Integration to form a mixed sample data set D mix ; Using the mixed sample dataset D mix Training a random forest model α ′ .

7. A device for predicting carbon emissions from coal-fired power plants, include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method for predicting carbon emissions of coal-fired power plants based on virtual sample generation as described in any one of claims 1-6.

8. A storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the method for predicting carbon emissions of a coal-fired power plant based on virtual sample generation as described in any one of claims 1 to 6 are implemented.