A Sparse Attention-based Atmospheric Carbon Concentration Prediction Method and System Based on Missing Data Imputation

Abstract

The present invention provides a sparse attention-based atmospheric carbon concentration prediction method and system for missing data imputation, comprising the following steps: collecting atmospheric carbon data; wherein the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4; using a multiple interpolation method based on historical backfill auxiliary data to perform missing data imputation on the atmospheric carbon data to obtain a complete data sequence; performing standardization and outlier processing on the complete data sequence to obtain an atmospheric carbon data set; based on the atmospheric carbon data set and a sparse attention mechanism-based atmospheric carbon concentration prediction model, performing dual-mode multi-scale prediction of the atmospheric concentration to obtain a concentration prediction result. The technical solution of the present invention can effectively improve the accuracy of time series prediction of multivariate dual-carbon data.

Description

Technical Field

[0001] The present invention belongs to the technical fields of machine learning, deep learning, and time series prediction, and particularly relates to a sparse attention-based atmospheric carbon concentration prediction method and system for missing data imputation. Background Art

[0002] With the intensification of the global carbon emission problem, under the goal of carbon neutrality, the research on the concentration change of carbon products in the atmosphere has become increasingly urgent. Carbon dioxide (CO2) and methane (CH4) are important greenhouse gases and have a significant impact on global climate change. XCO2 and XCH4 are parameters describing the concentrations of XCO2 and XCH4 in the atmosphere, usually expressed in volume fractions, and are used to monitor and evaluate the impacts of the global carbon cycle, climate change, and greenhouse gas emissions. TCCON (Total Carbon Column Observing Network) is the most widely used and recognized monitoring network at present, providing researchers with an important reference for understanding the spatio-temporal changes of XCO2 and XCH4 in the atmosphere, and is used to monitor greenhouse gas emissions and evaluate the global carbon cycle.

[0003] Transformer has been widely used in time series prediction, but its ability to capture underlying sequence representations and describe multivariate correlations is poor, which limits its ability and generalization ability for different time series data. To solve this problem, Tsinghua University recently proposed a new Transformer-like model called iTransformer, which combines a more variable-centered global time series tagging representation and a multivariate correlation attention mechanism, thus significantly improving the performance. iTransformer solves the problem of insufficient ability of Transformer to predict long time series.

[0004] The present invention improves the multivariate attention component of the iTransformer model to obtain a sparse attention mechanism-based atmospheric carbon concentration prediction model with higher performance, and obtains more accurate atmospheric carbon concentration prediction results. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a sparse attention-based atmospheric carbon concentration prediction method and system for missing data imputation, which effectively improves the accuracy of multivariate carbon neutrality data time series prediction.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A sparse attention-based atmospheric carbon concentration prediction method for missing data imputation, comprising the following steps:

[0008] Collect atmospheric carbon data; among them, the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4;

[0009] Adopt a multiple interpolation method based on historical backfill auxiliary data to perform missing data imputation on the atmospheric carbon data and obtain a complete data sequence;

[0010] Perform standardization and outlier processing on the complete data sequence to obtain an atmospheric carbon data set;

[0011] Based on the atmospheric carbon data set and the sparse attention mechanism atmospheric carbon concentration prediction model, perform dual-mode multi-scale prediction of the atmospheric concentration to obtain a concentration prediction result.

[0012] Preferably, the method for performing missing data imputation on the atmospheric carbon data to obtain a complete data sequence is as follows:

[0013] Identify the missing data segments and adjacent complete data segments of the atmospheric carbon data;

[0014] Based on the time tags of the identified data segments, obtain the corresponding historical complete data sequences;

[0015] Based on dynamic time warping, calculate the similarity between the missing data segments and the corresponding historical complete data sequences;

[0016] Based on the multiple interpolation method, calculate the interpolation results of the missing data segments;

[0017] Based on the similarity and the interpolation results, obtain the final interpolation results of the missing data segments;

[0018] Based on the final interpolation results, obtain the complete data sequence.

[0019] Preferably, the method for obtaining the atmospheric carbon data set is as follows:

[0020] Convert the complete data sequence to a unified daily granularity scale and perform mean processing on the daily atmospheric carbon data to obtain an average value sequence;

[0021] Perform outlier detection and removal on the average value sequence;

[0022] Based on the average value sequence after removing outliers, obtain the atmospheric carbon data set and divide the atmospheric carbon data set into a training set, a validation set, and a test set.

[0023] Preferably, use the probabilistic sparse self-attention method to improve the self-attention module of the iTransformer model to obtain the sparse attention mechanism atmospheric carbon concentration prediction model; the method for constructing the sparse attention mechanism atmospheric carbon concentration prediction model is as follows:

[0024] Independently embed the original time series of different meteorological variables in the training set into tokens;

[0025] Based on probabilistic sparse self-attention, uniformly sample the embedded original time series and select local context;

[0026] Based on the selection of local context, compress the original time series into a short time series of a preset length;

[0027] Based on the short time series, calculate a sparse attention weight matrix;

[0028] Based on the sparse attention weight matrix, sparsely attend to the local neighborhood in the short time series to obtain local neighborhood-related information;

[0029] Based on the local neighborhood-related information, capture the periodicity, trend information of the original time series of different meteorological variables, as well as the dependence relationship between multivariate variables, and complete the construction of a sparse attention mechanism atmospheric carbon concentration prediction model.

[0030] Preferably, when performing dual-mode multi-scale prediction of atmospheric concentration, the dual mode includes multivariate predicting univariate and multivariate predicting multivariate.

[0031] The present invention also provides a sparse attention atmospheric carbon concentration prediction system based on missing data imputation for implementing the method, including:

[0032] A data acquisition module for collecting atmospheric carbon data; wherein, the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4;

[0033] An interpolation module for imputing missing data of the atmospheric carbon data by using a multiple interpolation method based on historical backfill auxiliary data to obtain a complete data sequence;

[0034] A data set construction module for standardizing and outlier processing the complete data sequence to obtain an atmospheric carbon data set;

[0035] A multi-scale prediction module for performing dual-mode multi-scale prediction of atmospheric concentration based on the atmospheric carbon data set and the sparse attention mechanism atmospheric carbon concentration prediction model to obtain a concentration prediction result.

[0036] Preferably, the interpolation module includes:

[0037] A data segment identification unit for identifying the missing data segment and adjacent complete data segment of the atmospheric carbon data;

[0038] A historical data acquisition unit for obtaining the corresponding historical complete data sequence based on the time tag of the identified data segment;

[0039] A similarity calculation unit for calculating the similarity between the missing data segment and the corresponding historical complete data sequence based on dynamic time warping;

[0040] An interpolation calculation unit for calculating the interpolation result of the missing data segment based on a multiple interpolation method;

[0041] A final interpolation acquisition unit for obtaining the final interpolation result of the missing data segment based on the similarity and the interpolation result; and obtaining the complete data sequence based on the final interpolation result.

[0042] Preferably, the dataset construction module includes:

[0043] A data mean normalization unit for converting the complete data sequence into a unified daily granularity scale and performing mean normalization processing on the daily atmospheric carbon data to obtain an average value sequence;

[0044] An outlier detection unit for detecting and removing outliers from the average value sequence;

[0045] A dataset construction unit for obtaining an atmospheric carbon dataset based on the average value sequence with outliers removed, and dividing the atmospheric carbon dataset into a training set, a validation set, and a test set.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes a multiple interpolation method based on historical backfill auxiliary data to solve the inherent missing value characteristics of carbon concentration time series data. It can retain the uncertainty of the missing data, making the results more reliable. Compared with a single interpolation method, it can better reflect the randomness of the missing data and is applicable to various types of missing patterns and statistical analysis methods. The constructed sparse attention mechanism atmospheric carbon concentration prediction model for accurate and multi-window carbon concentration prediction can capture long-term data dependencies. The complexity of the model is relatively low, so that the carbon time series data can be more effectively predicted in real time. The present invention realizes high-precision prediction of atmospheric carbon data and effectively solves the problems of computational complexity and real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 It is a flowchart of a sparse attention atmospheric carbon concentration prediction method based on missing data imputation according to an embodiment of the present invention.

[0049] Figure 2Schematic diagram of the multiple interpolation method for historical backfill auxiliary data in the embodiments of the present invention;

[0050] Figure 3 Schematic diagram of the sparse attention mechanism atmospheric carbon concentration prediction model in the embodiments of the present invention;

[0051] Figure 4 Schematic diagram of the structure of the sparse attention atmospheric carbon concentration prediction system based on missing data imputation in the embodiments of the present invention. Detailed implementation manners

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0054] Embodiment 1

[0055] As Figure 1 shown, a sparse attention atmospheric carbon concentration prediction method based on missing data imputation includes the following steps:

[0056] S1: Collect atmospheric carbon data. In this embodiment, data from TCCON ground detection stations: Lamont, Edwards, and Parkfalls are collected. The atmospheric carbon data includes two meteorological variables, XCO2 and XCH4; in this embodiment, the data of XCO2 and XCH4 (the number of carbon dioxide and methane molecules per million air molecules) are sampled and recorded at the same time interval and frequency. However, the amount of data per day is different, and there are abnormal data exceeding the empirical threshold and missing data on a few dates. Therefore, it is necessary to preprocess the collected atmospheric carbon data.

[0057] S2: Use the multiple interpolation method based on historical backfill auxiliary data to perform missing data imputation on the atmospheric carbon data to obtain a complete data sequence. As Figure 2 shown.

[0058] During the data analysis process, it can be observed that the same time period in different years shows a similar distribution trend. Therefore, the present invention believes that this characteristic can be utilized to obtain reference values for missing data during the substitution process. Based on this observation result, the present invention proposes a historical review method assisted by multiple interpolation.

[0059] A further embodiment lies in the method of imputing missing data in atmospheric carbon data to obtain a complete data sequence:

[0060] Identify the missing data segments and adjacent complete data segments in the atmospheric carbon data;

[0061] Based on the time tags of the identified data segments, obtain the corresponding historical complete data sequences; specifically, find the corresponding historical data from the same time period in previous years. The sequence lacking data is called the missing data sequence, and the sequence with complete historical data is called the historical complete data sequence.

[0062] Based on Dynamic Time Warping (DTW), calculate the similarity C between the missing data segment i and the corresponding historical complete data sequence j ij , and obtain the historical corresponding value at the missing position (time t) of sequence j.

[0063] Calculate the interpolation result M of the missing data segment based on the multiple interpolation method i ; specifically, when using spline interpolation, a cubic spline function can be constructed to fit the data and interpolate the missing points.

[0064] Based on the similarity and the interpolation result, obtain the final interpolation result of the missing data segment;

[0065] Based on the final interpolation result X, obtain the complete data sequence. Specifically, using the weighted average method, combine the interpolation result of the current data with the interpolation result of the historical data, and the calculation equation is as follows:

[0066] X = W1(i * C ij ) + W2(M i ).

[0067] W1: represents the weight of the interpolation part based on similarity, where i * C ij is the similarity score calculated by Dynamic Time Warping (DTW) between the historical sequence His-Len i and His-Len j . W1 emphasizes the contribution of historical pattern similarity to the interpolation process.

[0068] W2: represents the weight of the interpolation value M i generated from the multiple interpolation method. It reflects the importance of the interpolation prediction method for the final result without directly relying on historical data. W2 controls the influence of the prediction-based interpolation technique on the result.

[0069] In this embodiment, to determine the optimal values of W1 and W2, the grid search method is used. The goal of this search is to balance the contributions of the similarity-based interpolation and the multiple interpolation method.

[0070] Grid search parameters: Define a range for W1 and W2. Both weight values should be between 0 and 1, and satisfy W1 + W2 = 1 to ensure the balance of the two weights. For example, set W1 = {0.1, 0.2, 0.3,... 0.9} and W2 = 1 - W1.

[0071] Grid search execution: Iterate over the preset W1 and W2 weight combinations. For each combination of W1 and W2, calculate the final interpolation result using the following formula: X = W1(i * C ij ) + W2(M i )

[0072] In this embodiment, in order to ensure the accuracy and reliability of the interpolation result, it is necessary to verify and optimize the interpolation result.

[0073] Cross-validation: Divide the data set into a training set and a validation set. Use the training set for interpolation calculation and evaluate the interpolation effect on the validation set. Through cross-validation, the generalization ability of the interpolation method can be detected.

[0074] Error analysis: Calculate the error between the interpolation result and the real data, such as the mean squared error (MSE), mean absolute error (MAE), etc. According to the error analysis results, adjust the interpolation method or parameters to improve the interpolation accuracy.

[0075] Visualization of interpolation results: By plotting a comparison graph of the interpolation result and the real data, the interpolation effect can be visually evaluated. Judge by plotting the time series graph before and after interpolation to observe whether the interpolation result is reasonable and smooth.

[0076] S3: Standardize the complete data sequence and process outliers to obtain the atmospheric carbon data set; A further implementation manner lies in that the method for obtaining the atmospheric carbon data set is:

[0077] Convert the complete data sequence to a unified daily granularity scale and perform mean processing on the daily atmospheric carbon data to obtain an average value sequence; Specifically, the carbon data is collected multiple times a day. To better analyze the data, the daily data is averaged, and the daily average value is retained as the only data at the time point.

[0078] Detect and remove outliers from the average value sequence; Specifically, use statistical methods or machine learning algorithms (such as the IQR method, Z-score, DBSCAN, etc.) to detect outliers and avoid the negative impact of these outliers on model training.

[0079] Based on the average value sequence after removing outliers, an atmospheric carbon dataset is obtained, and the atmospheric carbon dataset is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1. The training set is used to train the model, the validation set is used to adjust the model parameters, and the test set is used to evaluate the generalization ability of the model.

[0080] S4: Based on the atmospheric carbon dataset and the sparse attention mechanism atmospheric carbon concentration prediction model, perform dual-mode multi-scale prediction of the atmospheric concentration to obtain the concentration prediction result.

[0081] The overall structure of the proposed C-TimeSeries (sparse attention mechanism atmospheric carbon concentration prediction model) in the present invention has the same module layout as the Transformer encoder. In this model, the original sequences of different variables are independently embedded into tokens. This module reveals the multivariate correlations between variables by applying the multi-head self-attention mechanism to the embedded variable tokens.

[0082] As Figure 2 shown, the constructed C-TimeSeries deep learning model (sparse attention mechanism atmospheric carbon concentration prediction model) can capture the periodicity, trend information in time series data, and the complex dependencies between multiple variables. The C-TimeSeries model adopts an inverted version of the feed-forward network and applies it to the channels of different tokens to extract the complex features in the time series. By stacking inverted blocks, the observed time series is encoded, and the features of the future sequence are decoded using dense non-linear connections. The sparse self-attention method is an improved self-attention mechanism aimed at reducing the computational complexity of long sequence processing and maintaining high performance.

[0083] For independent time series, the same linear operations, that is, the combination of a linear predictor and a channel-independent strategy, help to extract sequence features. The time features extracted by the multi-layer perceptron (MLP) should be shared across the entire time series. The neurons of the MLP are trained to capture the intrinsic characteristics (such as amplitude periodicity) of any time series. This prediction model has an advantage over the self-attention mechanism that only depends on time points.

[0084] A further implementation manner lies in using the probabilistic sparse self-attention method to improve the self-attention module of the iTransformer model to obtain a sparse attention mechanism atmospheric carbon concentration prediction model; the method for constructing the sparse attention mechanism atmospheric carbon concentration prediction model is:

[0085] Independently embed the original time series of different meteorological variables in the training set into tokens;

[0086] Based on probabilistic sparse self-attention, uniformly sample the embedded original time series and select local contexts;

[0087] Based on the selection of local context, the original time series is compressed into a short time series of a preset length. In this way, the long input sequence is compressed into a shorter fixed-length representation, while reducing the computational amount.

[0088] Based on the short time series, a sparse attention weight matrix is calculated. Specifically, most elements of the matrix are zero, and this sparse matrix can effectively reduce the computational cost and memory, thereby improving the efficiency of the model in long sequence tasks.

[0089] Based on the sparse attention weight matrix, sparse attention is paid to the local neighborhood in the short time series to obtain local neighborhood-related information and reduce the computational complexity.

[0090] Based on the local neighborhood-related information, the periodicity, trend information of the original time series of different meteorological variables and the dependence relationship between multiple variables are captured to complete the construction of the sparse attention mechanism atmospheric carbon concentration prediction model. The present invention realizes high-precision prediction of atmospheric carbon data and effectively solves the problems of computational complexity and real-time requirement.

[0091] To meet the characteristics of carbon data, we also have the following requirements:

[0092] Focus on local information: The model should focus on the local information of carbon time series data, such as the irregular changes in carbon data records caused by emergencies.

[0093] Capture long-term dependence relationships: The model should capture long-term data dependence relationships, such as seasonal characteristics.

[0094] Reduce complexity: The complexity of the model should be low so that carbon time series data can be predicted in real time more effectively.

[0095] A further implementation manner is that when performing dual-mode multi-scale prediction of atmospheric concentration, the dual mode includes predicting a single variable with multiple variables and predicting multiple variables with multiple variables.

[0096] Specifically,

[0097] Predicting a single variable with multiple variables: The prediction model uses multiple input variables to predict a single output variable.

[0098] Predicting a single variable with a single variable: The prediction model uses a single input variable to predict a single output variable.

[0099] In both the multi-variable and single-variable modes, the input prediction window is set to a step size of 96. After modeling XCO2 and XCH4, the output prediction windows are 12, 24, 36, 48, 96, 192, a total of six prediction scales.

[0100] The mean squared error (MSE) and the mean absolute error (MAE) were selected as the evaluation metrics of the model to quantitatively analyze and evaluate the experimental performance of the model. MSE and MAE are two commonly used prediction error evaluation metrics, which calculate the differences between the predicted values and the true values respectively as the final prediction evaluation metrics of the model.

[0101] To better process the input and output data, three different types of sliding window strategies were adopted in this embodiment, namely M, MS, and S modes:

[0102] M mode (multiple variables in, multiple variables out): The prediction model predicts multiple output variables based on multiple input variables. For example, for the multi-variables XCO2 and XCH4, the prediction model predicts the future multi-variables XCO2 and XCH4.

[0103] MS mode (multiple variables in, single variable out): The prediction model uses multiple input variables to predict a single output variable. For example, predicting the future values of XCO2 and XCH4 is based on the past values of multiple variables (such as XCO2 and XCH4).

[0104] S mode (single variable in, single variable out): The prediction model uses a single input variable to predict a single output variable. For example, predicting XCO2 and XCH4 in the next few days is only based on the previous XCO2 and XCH4 data respectively.

[0105] Most of the time series prediction studies in recent years have adopted the M or MS mode. In this study, the TCCON XCO2 and XCH4 data of three sites were used. There are three types of data in the dataset: date, XCO2, and XCH4 data. Therefore, the present invention also selects the M and MS processing modes to compare and predict XCO2 and XCH4 at three sites respectively.

[0106] To conduct ablation experiments and quantitatively test the effectiveness of the improved model, experiments were conducted on two models respectively. The specific experimental steps are as follows:

[0107] Experiment of M mode:

[0108] Input variables: XCO2 and XCH4.

[0109] Output result: The common prediction result of XCO2 and XCH4.

[0110] Experimental set: Data from three sites.

[0111] Experiment of MS mode:

[0112] Input variables: XCO2 and XCH4.

[0113] Output results: Predict XCO2 and XCH4 respectively.

[0114] Experimental set: Data from three sites.

[0115] Next, based on the XCO2 and XCH4 time series data, the proposed C-TimeSeries model was experimented 6 times in the same way. In the experiment, the input prediction window (SeqLen) was set to a length of 96, and the output prediction windows of 6 scales were 12, 24, 36, 48, 96, and 192 respectively. Corresponding training and testing were carried out, and the MSE and MAE values of each group of experiments were calculated to quantitatively evaluate the performance of the model.

[0116] C-TimeSeries helps to understand carbon trends and provides information for management strategies. As shown in Table 1 (Lamont experimental results), Table 2 (Parkfalls experimental results), and Table 3 (Edwards experimental results).

[0117] Table 1

[0118]

[0119] Table 2

[0120]

[0121] Table 3

[0122]

[0123] In practical applications, the present invention proposes a complete set of solutions. For those skilled in the art, they can recognize that the technical solution of the present invention can be implemented through a computer program and its corresponding general hardware platform. Based on this understanding, the technical solution of the present invention, or rather its contribution to the prior art, can be embodied as a computer program software product, which is stored in the above storage medium and contains several instructions for guiding a device including a data processing unit to execute the methods described in various embodiments or specific parts of the embodiments of the present invention.

[0124] Embodiment 2

[0125] As Figure 4 shown, the present invention also provides a sparse attention atmospheric carbon concentration prediction system based on missing data imputation for implementing the method, including:

[0126] A data acquisition module for acquiring atmospheric carbon data; among them, the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4;

[0127] An interpolation module for imputing missing data in the atmospheric carbon data using a multiple interpolation method based on historical backfill auxiliary data to obtain a complete data sequence;

[0128] A dataset construction module for standardizing and processing outliers in the complete data sequence to obtain an atmospheric carbon dataset;

[0129] A multi-scale prediction module for performing dual-mode multi-scale prediction of atmospheric concentration based on the atmospheric carbon dataset and a sparse attention mechanism atmospheric carbon concentration prediction model to obtain a concentration prediction result.

[0130] A further implementation manner is that the interpolation module includes:

[0131] A data segment identification unit for identifying missing data segments and adjacent complete data segments in the atmospheric carbon data;

[0132] A historical data acquisition unit for obtaining corresponding historical complete data sequences based on the time tags of the identified data segments;

[0133] A similarity calculation unit for calculating the similarity between the missing data segment and the corresponding historical complete data sequence based on dynamic time warping;

[0134] An interpolation calculation unit for calculating the interpolation result of the missing data segment based on the multiple interpolation method;

[0135] A final interpolation acquisition unit for obtaining the final interpolation result of the missing data segment based on the similarity and the interpolation result;

[0136] Based on the final interpolation result, a complete data sequence is obtained.

[0137] A further implementation manner is that the dataset construction module includes:

[0138] A data mean normalization unit for converting the complete data sequence to a unified daily granularity scale and performing mean normalization processing on the daily atmospheric carbon data to obtain an average value sequence;

[0139] An outlier detection unit for detecting and removing outliers from the average value sequence;

[0140] A dataset construction unit for obtaining an atmospheric carbon dataset based on the average value sequence with outliers removed and dividing the atmospheric carbon dataset into a training set, a validation set, and a test set.

[0141] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A sparse attention-based atmospheric carbon concentration prediction method using missing data imputation, characterized in that, It includes the following steps: Collect atmospheric carbon data; among them, the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4; Adopt a multiple interpolation method based on historical backfill auxiliary data to interpolate the missing data in the atmospheric carbon data to obtain a complete data sequence; Standardize and process outliers for the complete data sequence to obtain an atmospheric carbon data set; the method for obtaining the atmospheric carbon data set is: Convert the complete data sequence to a unified daily granularity scale and average the daily atmospheric carbon data to obtain an average value sequence; Detect and remove outliers from the average value sequence; Based on the average value sequence with outliers removed, obtain an atmospheric carbon data set, and divide the atmospheric carbon data set into a training set, a validation set, and a test set; based on the atmospheric carbon data set and the sparse attention mechanism atmospheric carbon concentration prediction model, perform dual-mode multi-scale prediction of atmospheric concentration to obtain a concentration prediction result; Use the probability-based sparse self-attention method to improve the self-attention module of the iTransformer model to obtain the sparse attention mechanism atmospheric carbon concentration prediction model; the method for constructing the sparse attention mechanism atmospheric carbon concentration prediction model is: Independently embed the original time series of different meteorological variables in the training set into tokens; Based on probability-based sparse self-attention, uniformly sample the embedded original time series and select local context; Based on the selection of local context, compress the original time series into a short time series of a preset length; Based on the short time series, calculate the sparse attention weight matrix; Based on the sparse attention weight matrix, sparsely focus on the local neighborhood in the short time series to obtain local neighborhood-related information; Based on the local neighborhood-related information, capture the periodicity, trend information of the original time series of different meteorological variables, and the dependencies between multiple variables to complete the construction of the sparse attention mechanism atmospheric carbon concentration prediction model.

2. The sparse attention-based atmospheric carbon concentration prediction method based on missing data imputation according to claim 1, wherein The method for interpolating the missing data in the atmospheric carbon data to obtain a complete data sequence is: Identify the missing data segments and adjacent complete data segments in the atmospheric carbon data; Based on the time tags of the identified data segments, obtain the corresponding historical complete data sequences; Based on dynamic time warping, calculate the similarity between the missing data segments and the corresponding historical complete data sequences; Calculate the interpolation results of the missing data segments based on the multiple interpolation method; Based on the similarity and the interpolation results, obtain the final interpolation results of the missing data segments; Based on the final interpolation results, obtain the complete data sequence.

3. The sparse attention-based atmospheric carbon concentration prediction method based on missing data imputation according to claim 1, characterized in that, When performing dual-mode multi-scale prediction of atmospheric concentration, the dual mode includes multi-variable predicting single variable and multi-variable predicting multi-variable.

4. A sparse attention-based atmospheric carbon concentration prediction system for imputing missing data, which is used to implement the method according to any one of claims 1-3, characterized in that It includes: A data acquisition module for collecting atmospheric carbon data; among them, the atmospheric carbon data includes two meteorological variables, XCO2 and XCH4; An interpolation module for interpolating the missing data in the atmospheric carbon data by using a multiple interpolation method based on historical backfill auxiliary data to obtain a complete data sequence; A data set construction module for standardizing and processing outliers for the complete data sequence to obtain an atmospheric carbon data set; A multi-scale prediction module for performing dual-mode multi-scale prediction of atmospheric concentration based on the atmospheric carbon dataset and the sparse attention mechanism atmospheric carbon concentration prediction model to obtain a concentration prediction result.

5. The sparse attention-based atmospheric carbon concentration prediction system for missing data imputation according to claim 4, wherein The interpolation module includes: A data segment identification unit for identifying missing data segments and adjacent complete data segments of the atmospheric carbon data; A historical data acquisition unit for obtaining corresponding historical complete data sequences based on the time tags of the identified data segments; A similarity calculation unit for calculating the similarity between the missing data segments and the corresponding historical complete data sequences based on dynamic time warping; An interpolation calculation unit for calculating the interpolation results of the missing data segments based on multiple interpolation methods; A final interpolation acquisition unit for obtaining the final interpolation results of the missing data segments based on the similarity and the interpolation results; and obtaining the complete data sequences based on the final interpolation results.

6. The sparse attention-based atmospheric carbon concentration prediction system based on missing data imputation according to claim 4, wherein The dataset construction module includes: A data mean normalization unit for converting the complete data sequences into a unified daily granularity scale and performing mean normalization processing on the daily atmospheric carbon data to obtain a mean value sequence; An outlier detection unit for detecting and removing outliers from the mean value sequence; A dataset construction unit for obtaining an atmospheric carbon dataset based on the mean value sequence with outliers removed and dividing the atmospheric carbon dataset into a training set, a validation set, and a test set.

Images