Atmospheric carbon data prediction method and system

Abstract

The invention discloses an atmospheric carbon data prediction method and system. The method comprises the following steps: preprocessing TCCON and GONGGA data set site multivariable time sequence carbon data; dividing a period trend term, performing time-frequency conversion by adopting parallel multi-scale wavelet transform after trend decomposition, and then performing adaptive threshold denoising processing and feature extraction to obtain data time-frequency features; and carrying out weighted fusion on the time-frequency features of the data through a feature fusion method, adjusting adaptive weight and parameter updating based on gradient descent, training an improved long-short-term memory network model based on the fused features, and carrying out atmospheric carbon concentration prediction to obtain a prediction result. According to the method, the accuracy and efficiency of atmospheric carbon data prediction are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of carbon data prediction, and particularly relates to a method and system for predicting atmospheric carbon data. Background Art

[0002] Under the background of carbon peaking and carbon neutrality, controlling and reducing carbon sources and increasing carbon sinks have become urgent tasks for maintaining ecological balance and social sustainability. However, due to the particularity of the geographical environment and the lack of monitoring facilities in China, there are a large number of missing carbon flux data, which poses challenges to accurately assessing the carbon cycle and carbon storage in this region. Carbon flux is one of the most basic concepts in carbon cycle research, which refers to the total amount of carbon elements passing through a certain ecological section by an ecosystem. The carbon flux of a forest ecosystem is the total amount of carbon cycle per unit time of this ecosystem. Among them, carbon sink is regarded as "negative carbon flux", which absorbs and stores carbon from the atmosphere, resulting in a decrease in the carbon concentration in the atmosphere. Carbon source is regarded as "positive carbon flux", which releases carbon into the atmosphere, resulting in an increase in the carbon concentration in the atmosphere. Carbon flux is a major driving factor of global climate change, because the increase in positive carbon flux (carbon source) causes air pollution and a decline in air quality, which has a negative impact on human health.

[0003] Under this background, the existing technology estimates the carbon flux data in a certain area in the next few years by using the GONGGA global carbon flux dataset as the main prediction dataset. TCCON is the most widely used and recognized monitoring network in the world, providing important data support. TCCON measures the total carbon column (XCO2) in the atmosphere by using infrared spectrometers through ground observation stations around the world. Its data is used to verify the prediction accuracy. These measurement data provide important references for researchers to understand the spatio-temporal changes of carbon gases in the atmosphere, help monitor greenhouse gas emissions and evaluate the global carbon cycle.

[0004] Existing time series prediction models often ignore the accurate denoising problem in the prediction of multi-variable time series. Especially when dealing with non-stationary time series with many mutations and frequent local changes, the prediction accuracy of traditional time series prediction models is insufficient. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method and system for predicting atmospheric carbon data based on multi-scale parallel wavelet denoising and feature extraction, so as to effectively reduce the accuracy of predicting noise time series in carbon peaking and carbon neutrality data under the channel-independent method.

[0006] Specifically, the present invention provides a method and system for predicting atmospheric carbon data. Among them, a method for predicting atmospheric carbon data includes:

[0007] Preprocessing the multi-variable time series carbon data of TCCON and GONGGA dataset sites;

[0008] Divide the periodic trend term. After trend decomposition, perform time-frequency conversion using parallel multi-scale wavelet transform, then perform adaptive threshold denoising processing and feature extraction to obtain the time-frequency features of the data;

[0009] Perform weighted fusion on the time-frequency features of the data through a feature fusion method, adjust the adaptive weights and parameter updates based on gradient descent, train an improved long short-term memory network model based on the fused features, and perform atmospheric carbon concentration prediction to obtain the prediction results.

[0010] Preferably, preprocess the multi-variable time series carbon data of the TCCON and GONGGA dataset sites, including regional rasterization and longitude and latitude alignment, missing value imputation, and self-supervised pre-training;

[0011] The regional rasterization and longitude and latitude alignment include:

[0012] Collect global carbon flux data through the GONGGA inversion system as regional carbon flux data, and then perform rasterization processing on the carbon flux data; further process the processed carbon flux dataset, and strip out the part where the net carbon flux is negative as a carbon sink for separate analysis;

[0013] Align the longitude and latitude with the drawn city or province boundary map to obtain the local carbon flux data, and at the same time impute and supplement the missing data of the local carbon flux data to obtain a complete dataset.

[0014] Preferably, the missing value imputation includes: regional rasterization and longitude and latitude calibration, output value imputation, sampling at a fixed time interval, outlier removal, and regularization;

[0015] Among them, the regional rasterization and longitude and latitude calibration is to rasterize the original data according to the region where it is located and calibrate the longitude and latitude of different data sources;

[0016] The output value imputation is to impute the output values in the data using the linear interpolation method;

[0017] The sampling at a fixed time interval is to sample the data at a fixed time interval;

[0018] The outlier removal is to identify the outliers in the data using the box plot method and remove the outliers;

[0019] The regularization is to perform regularization processing on the data, standardize the features and set the same scale range.

[0020] Preferably, the process of dividing the periodic trend term, performing time-frequency conversion using parallel multi-scale wavelet transform after trend decomposition, then performing adaptive threshold denoising processing and feature extraction to obtain the time-frequency features of the data includes:

[0021] When performing periodic trend division on data, the moving average method is used to smooth the data;

[0022] The Fourier transform and the discrete wavelet transform are used to perform time-frequency domain conversion on the data respectively, and multi-graphic features of the data are extracted;

[0023] The features extracted by using the wavelet transform in the frequency domain are denoised, and the adaptive threshold denoising method is used to remove the noise components in the frequency domain.

[0024] Preferably, the process of smoothing the data by using the moving average method includes:

[0025] First, the multivariate time series is divided into several overlapping small blocks to obtain small block sequences;

[0026] Next, the moving average decomposition block is used to represent the trend and seasonal components, the original data sequence is filled, then the average pooling is applied to extract the trend component, and the seasonal component is separated by subtracting the trend component from the overall data.

[0027] Preferably, the process of weighted fusion of the time-frequency features of the data by using the feature fusion method includes:

[0028] According to the time-frequency features extracted by the multi-dimensional wavelet transform, the dimension features of different dimensions are separated, different weights are assigned, and different key dimension features are allocated;

[0029] The topological layer is used for local feature extraction, the multi-scale features are aggregated by the linear layer, the local dependencies in the time series are extracted by the morphological layer, and the multi-size measurements are integrated by the linear layer;

[0030] The processed data is converted from the frequency domain to the time domain, and the Fourier inverse transform is performed on the visualized prediction data and the verification data respectively.

[0031] Preferably, the adaptive weights and parameter updates are adjusted based on gradient descent, including the updates of the adaptive weights, the optimal wavelet basis, and the threshold parameters.

[0032] Preferably, the adaptive weights include the global weights represented by the Fourier transform and the local weights represented by the wavelet transform;

[0033] The selection of the optimal wavelet basis includes selecting the wavelet basis combination through the analysis of the experimental results of each data set and dynamically adjusting the corresponding weights according to the contributions;

[0034] The threshold parameter update includes assigning different initial thresholds to the seasonal term and the trend term respectively, using the product of the magnitude of the energy value and the initial threshold and continuously updating the initial threshold to continuously update the threshold to select the optimal threshold; and dynamically adjusting according to the transformation of the data set and the time series change in different time periods.

[0035] Preferably, based on the fused features, an improved long short-term memory network model is trained for atmospheric carbon concentration prediction. The process of obtaining the prediction result includes:

[0036] Using the improved long short-term memory model to input new atmospheric carbon data to be predicted into the trained prediction model, predicting the future trend of the atmospheric carbon concentration, and using the mean square performance and the mean absolute performance as model evaluation indicators to evaluate the performance of the model.

[0037] The present invention also provides an atmospheric carbon data prediction system, including:

[0038] A data preprocessing module for preprocessing the multi-variable time series carbon data of the TCCON and GONGGA data set sites;

[0039] A multi-variable time series time-frequency feature extraction and denoising module for dividing the periodic trend term, performing time-frequency conversion using parallel multi-scale wavelet transform after trend decomposition, and then performing adaptive threshold denoising processing and feature extraction to obtain the data time-frequency features;

[0040] A multi-variable time series time-frequency feature fusion module for weighted fusion of the data time-frequency features by a feature fusion method;

[0041] A multi-scale convolutional layer feature fusion module for adjusting the adaptive weights and parameter updates based on gradient descent;

[0042] An atmospheric carbon data prediction model training and optimization module for training an improved long short-term memory network model based on the fused features;

[0043] An atmospheric carbon data prediction module for predicting the atmospheric carbon concentration to obtain a prediction result.

[0044] Compared with the prior art, the present invention has the following advantages and technical effects:

[0045] The carbon data prediction method based on multi-scale parallel wavelet denoising and feature extraction proposed by the present invention significantly improves the accuracy and efficiency of atmospheric carbon data prediction through a series of innovative steps. The parallel multi-scale wavelet basis with frequency domain enhancement uses an adaptive threshold (MSW-EFD) method to alleviate the problems of residual and noise processing in multivariate time series prediction. This method first divides the multivariate time series into several overlapping small blocks to obtain small block sequences, enhancing the connection between scales and providing conditions for subsequent wavelet decomposition. Next, the moving average decomposition block is used to represent the trend and seasonal components, and then the Fourier transform is applied to transform them into the frequency domain. Subsequently, the adaptive parallel dual wavelet transform and adaptive threshold are used to attenuate the prominent high-frequency noise in the frequency domain and emphasize the relevant spectral features. After processing, the inverse fast Fourier transform (IFFT) is applied to reconstruct the time domain features, thereby reducing noise and enhancing the representation. It is added to each baseline. Subsequently, three independent real-world carbon concentration datasets and publicly available time series prediction datasets are used to evaluate the long-term atmospheric carbon data prediction model and system based on multi-scale parallel wavelet denoising and feature extraction. The results show that the model using the MSW-EFD method significantly improves the state-of-the-art baseline. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0047] Figure 1 It is a schematic flowchart of the method of the embodiment of the present invention;

[0048] Figure 2 It is a time domain comparison diagram before and after denoising of the seasonal term and trend term of the embodiment of the present invention;

[0049] Figure 3 It is a spectral comparison diagram before and after denoising of the model seasonal term of the embodiment of the present invention;

[0050] Figure 4 It is a diagram showing the fitting effect of the contrast Fourier and different wavelet bases used on local information and the extracted high-frequency information of the embodiment of the present invention;

[0051] Figure 5 It is a comparison result diagram of the prediction accuracy of each model at different noise levels of the embodiment of the present invention;

[0052] Figure 6 It is a schematic diagram of the system structure of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The following will describe the present application in detail with reference to the drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowchart of the drawings may be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than here.

[0055] Embodiment 1

[0056] As Figures 1-5 shown, in this embodiment, a method for predicting atmospheric carbon data is provided, including the following steps:

[0057] S1: Preprocess the multi-variable time series carbon data of the TCCON and GONGGA dataset sites, including regional rasterization and longitude and latitude alignment, missing value imputation, and self-supervised pre-training; the preprocessed data includes prediction data and verification data;

[0058] S2: Divide the periodic trend term, perform time-frequency conversion using parallel multi-scale wavelet transform after trend decomposition, and then perform adaptive threshold denoising processing and feature extraction to obtain the time-frequency features of the data;

[0059] S3: Weightedly fuse the extracted multi-scale features through a feature fusion method; specifically, perform feature fusion on the time-frequency features of the multi-variable time series, and introduce convolution and linear layers for fusion features after assigning different weights according to the different extracted features;

[0060] S4: Adjust the adaptive weights and parameter updates based on gradient descent, including adaptive weight, optimal wavelet basis, and threshold parameter updates;

[0061] S5: Train an improved long short-term memory network (LSTM) model based on the fused features to predict the atmospheric carbon concentration.

[0062] Furthermore, in step S1:

[0063] Data rasterization includes:

[0064] Collect global carbon flux data from the GONGGA inversion system as regional carbon flux prior data, then rasterize the carbon flux data, align the longitude and latitude with the drawn city or province boundary map, and then deduce the carbon flux in that area. At the same time, use the multiple imputation method to supplement a small amount of missing data to obtain a complete dataset. The processed carbon flux dataset is further processed to strip out the part where the net carbon flux is negative as a carbon sink for separate analysis.

[0065] Longitude and latitude alignment includes:

[0066] Drawing city boundary maps: Since the resolution of the dataset is 1*1°, draw 1*1° city boundary maps based on longitude and latitude for refined processing of urban carbon data.

[0067] Then, use the multiple imputation method to supplement a small amount of missing data, and further process the processed carbon flux dataset. Separate the carbon sink part for individual re-analysis.

[0068] Missing value imputation includes:

[0069] For incremental imputation of the original data, use the linear interpolation method. For fixed-interval hourly sampling, adopt a sampling strategy with a time interval of 24, and identify outliers through the boxplot method.

[0070] Specifically, it includes: regional rasterization and longitude and latitude calibration: rasterize the original data according to the region where it is located, and calibrate the longitude and latitude of different data sources to ensure the consistency of each data point.

[0071] Output value imputation: Use the linear interpolation method to impute the output values in the data to ensure the continuity of the time series.

[0072] Sampling interval sampling: Sample the data at a fixed time interval (such as 24 hours) to unify the time resolution of the data.

[0073] Outlier removal: Use the boxplot method to identify outliers in the data and remove these outliers to improve data quality.

[0074] Regularization: Regularize the data, standardize each feature, and set the same scale range to avoid some features playing a dominant role in model training.

[0075] The self-supervised pre-training stage includes:

[0076] In this embodiment, a self-supervised pre-training stage is introduced to enhance the ability of WE-LSTM, adopting the masked autoencoder paradigm of the time series data reported by MHCCL. It involves selectively masking input sequence blocks and then training WE-LSTM to accurately reconstruct these masked segments. The masked data serves as the training input, forcing the model to learn and infer the latent patterns and dependencies in the data. Use MSE as the loss function to optimize the model's reconstruction task, minimizing the difference between the predicted value and the actual masked value. Assume that xmasked represents the masked input segment and the output of the model is xmasked. The loss function is defined as:

[0077]

[0078] Where N is the number of mask segments, x(i)masked is the true value of the i-th mask segment, and masked is the corresponding value reconstructed by the model.

[0079] For a further optimization solution, the WE-LSTM model proposed in this embodiment integrates two components: an enhanced spectrum denoising module and an enhanced long short-term memory module. These components form a single layer that can be extended into multiple layers, which divides the input time series into multiple patches and adds positional encoding. Next, a moving average decomposition module is used to represent the trend and seasonal components; then, the signal is input into the ESD module for denoising and feature extraction, and finally input into the ELSTM network to capture complex time patterns.

[0080] Furthermore, in step S2: Periodic noise is enhanced by the moving average method, and the wavelet transform uses the discrete wavelet transform (DWT) and combines multiple wavelet bases to extract the time-frequency features of the data.

[0081] Specifically, for the moving average method: When dividing the data into periodic trends, the moving average method is used to smooth the data to eliminate short-term fluctuations and noise.

[0082] Fourier transform and wavelet transform: The Fourier transform and the discrete wavelet transform (DWT) are used to perform time-frequency domain conversion on the data respectively. The time-frequency domain conversion of the data is performed to extract the multi-graph features of the data.

[0083] Multi-wavelet threshold denoising: The features extracted by the wavelet transform in the frequency domain are used for denoising, and the adaptive threshold denoising or other appropriate methods are used to remove the noise components in the frequency domain.

[0084] More specifically, the time-frequency feature extraction and denoising method for multivariate time series includes:

[0085] The adaptive attenuation of noise in the trend and seasonal components is a key issue for time series data. High-frequency components usually represent rapid fluctuations that deviate from the basic trend or the signal of interest, showing strong randomness and thus being difficult to interpret. On the contrary, the low-frequency components reflect the main trend changes of the signal.

[0086] Furthermore, it includes the following steps: This embodiment uses the moving average technique to achieve the trend and seasonal decomposition of the time series.

[0087] This method first divides the multivariate time series into several overlapping small blocks to obtain small block sequences:

[0088]

[0089] Effectively solves the problems of information loss and boundary effects, enhances the connection between scales and provides

[0090] Next, the moving average decomposition block is used to represent the trend and seasonal components. First, the original data sequence is padded, and then average pooling is applied to extract the trend component.

[0091] xtrend = AveragePool(Padding(x))

[0092] The seasonal component is separated by subtracting the trend component from the overall data.

[0093] x seasonal = x - x trend

[0094] Then, the FFT is applied to transform them into the FD (frequency domain), and then denoising is performed. Specifically, the dual wavelet denoising method first decomposes the signal into separate approximation coefficients and detail coefficients (high frequency) through wavelet decomposition. This method uses dual wavelet bases suitable for global and local features, in parallel with the trend and seasonal modules. Then, the standard deviations of the two wavelet basis coefficients are calculated, and a factor (0.01) is used as the threshold to perform soft threshold denoising on each decomposition layer. The adaptive threshold is set in the trend and seasonal modules and continuously updated through backpropagation to further reduce the residual noise in the trend and seasonal components. Then, each denoised component is connected through a convolutional layer. In this embodiment, wavelet transform and adaptive threshold are used to attenuate the prominent high-frequency noise in the frequency domain and emphasize the relevant spectral features. After adaptive filtering of the frequency domain data, the model uses two sets of learnable filters applied to the trend and seasonal components respectively, allowing the model to dynamically adjust the importance of each filter based on data features: a global filter for learning from the original frequency domain data F, and a local filter for learning from the data Ffiltered obtained through adaptive filtering. Let WG and WL represent the learnable global and local filters respectively.

[0095] After processing, the inverse fast Fourier transform (IFFT) is applied to reconstruct the time-domain features, and the results from different wavelet bases in these two parts are concatenated. Then, these concatenated features are processed through a convolutional layer to capture the denoised features of the trend and seasonality. An activation function is introduced to further enhance the model's ability to process non-linear features. The processed trend component and seasonal component are combined along with the features captured from each component through addition.

[0096] In the final stage, it is fed into the ELSTM module for training. ELSTM is an advanced variant of the stable LSTM model, which includes exponential gating, memory mixing, and stable mechanisms.

[0097] Finally, the trained model is used to predict and analyze the carbon data.

[0098] The following are the detailed steps of multi-wavelet parallel denoising:

[0099] The decomposed trend and seasonal components are respectively input into the trend and seasonal modules of ESD for processing. First, the signal is decomposed into separate approximation coefficients and detail coefficients (high frequency) through discrete wavelet decomposition.

[0100]

[0101] Then, it extracts additional high frequencies from the approximation coefficients of the further decomposition. The expression of the decomposition level is given by the following formula:

[0102]

[0103] This method adopts a multi-scale wavelet basis suitable for global and local features, in parallel with the trend and seasonal modules.

[0104] Then, the standard deviations of the two wavelet basis coefficients are calculated, and soft threshold denoising is performed on each decomposition layer using the initial factor (0.01) as the threshold. The adaptive threshold is set in the trend and seasonal modules and continuously updated through backpropagation to further reduce the residual noise in the trend and seasonal components. Then, each denoised component is connected through a convolutional layer. In the model of this embodiment, the dynamic threshold λ is adaptively calculated based on the signal feature E(x) and the learnable parameter θ, as shown below:

[0105] E(x) = standard deviation(x)

[0106] T = E(x)·θ

[0107]

[0108] Here, λ represents the dynamic threshold, E(x) represents the signal characteristic (such as the standard deviation), and λ is a trainable parameter initialized as λ0 and updated via gradient descent during training. In this work, the signal feature E(x) is explicitly defined as the standard deviation of the signal:

[0109] FL = WL ⊙ Ffiltered

[0110] FG = WG ⊙ F

[0111] Fintegrated = FG + FL

[0112] The dynamic threshold parameter λ is updated through backpropagation. This adaptive mechanism ensures that the threshold parameter is optimized for the overall learning process of the model. After adaptive filtering of the frequency-domain data, the model employs two sets of learnable filters applied to the trend and seasonal components respectively, allowing the model to dynamically adjust the importance of each filter based on data characteristics: a global filter for learning from the original frequency-domain data F, and a local filter for learning from the data Ffiltered obtained through adaptive filtering. Let WG and WL denote the learnable global and local filters respectively.

[0113] Next, these filtered features are integrated to capture comprehensive spectral details. Finally, the IFFT is applied to the processed trend and seasonal components respectively and enters the feature fusion module. After applying the IFFT to the trend component and the seasonal component, the results from different wavelet bases in these two parts are concatenated. Then, these concatenated features are processed through a convolutional layer to capture the denoised features of the trend and seasonality. An activation function is introduced to further enhance the model's ability to process non-linear features. The processed trend component and seasonal component are combined along with the features captured from each component through addition. In the last stage, the complete output is refined through a linear layer, enabling the model to capture the overall trend of the time series data.

[0114] Generally speaking, this feature fusion method improves the model's recognition ability, enabling it to capture deeper features from time series data.

[0115] Furthermore, in step S3, feature fusion is performed by the weighted average method, and the process includes:

[0116] Two weights are respectively set for the trend and seasonal terms, and the weights are continuously updated during superposition according to their different contributions to obtain the optimal weights. The Fourier transform and wavelet transform are distinguished in a parallel manner, and weights are respectively assigned to the global features captured by the Fourier transform and the local feature distributions captured by the wavelet transform, serving as global and local filters for feature extraction.

[0117] More specifically, it includes the following steps:

[0118] Feature extraction: According to the time-frequency features extracted by multi-dimensional wavelet transform, the dimensional features of different dimensions are separated and different weights are assigned. The assignment of weights is based on feature importance or correlation analysis, and different weights are assigned to different key dimensional features.

[0119] Feature fusion: Use a topological layer for local feature extraction and aggregate multi-scale features through a linear layer to capture more complex spatio-temporal patterns at high depths. The morphological layer is used to extract local dependencies in the time series, while the linear layer helps to integrate multi-size measurements.

[0120] Frequency domain to time domain conversion: Convert the processed data from the frequency domain to the time domain, and perform inverse Fourier transforms on the visualized prediction data and the verification data respectively.

[0121] Further optimization scheme, the multivariate time series feature fusion and long time series prediction method in this implementation includes:

[0122] In this embodiment, in view of the poor performance of the common multivariate time series based on channel-independent methods in the face of longer time series prediction, it is noted that there is a large amount of noise inherently in the multivariate carbon data, and because the data is updated slowly and the annual data has a lag, long time series prediction is required. This work selects LSTM as the backbone network for further noise control. This embodiment proposes an enhanced LSTM (ELSTM) module, aiming to improve the ability to capture complex patterns in the long term and short term. In ELSTM, a smoother softplus activation function and an exponential gating unit are added to the input gate to enhance the model stability. It also integrates gradient clipping and Xavier initialization to reduce the risk of gradient explosion and vanishing. To further improve the memory ability, ELSTM uses a candidate memory module. In addition, ELSTM combines a convolutional neural network (CNN), combining their advantages in capturing local patterns to better learn and represent complex global and local patterns.

[0123] This embodiment solves the long time dependence problem of time series by aligning variables before the prediction process, aiming to improve the accuracy and robustness of the prediction. Its main steps include: ELSTM introduces Dropout in the output layer to reduce overfitting and enhance the generalization ability, and at the same time adopts an exponential gating mechanism to control the input gate and the forget gate to make the information flow more effective. The relevant formula expressions include:

[0124] i t =exp(i tilda ),f t =σ(f tilda )

[0125] m t =log(exp(log(f t )+m t-1 )+exp(log(i t )))

[0126]

[0127]

[0128] Furthermore, the process of adjusting the adaptive weights and parameter updates based on gradient descent includes:

[0129] Wavelet basis selection: Select the wavelet basis combination by analyzing the experimental results of each dataset, and dynamically adjust the corresponding weights according to their contributions.

[0130] Global and local weights: The global weights represented by the Fourier transform and the local weights represented by the wavelet transform should be evaluated and continuously updated by the gradient descent method.

[0131] Dynamic threshold adjustment: Different initial thresholds are assigned to the seasonal term and the trend term. The product of the energy value and the initial threshold is used and the initial threshold is continuously updated to select the optimal threshold; and it is dynamically adjusted according to the changes in the dataset and the time series changes in different time periods.

[0132] Furthermore, based on the fused features, an improved long short-term memory network (LSTM) model is trained. The process of predicting the atmospheric carbon concentration includes:

[0133] Improve the long short-term memory network so that it can capture long-term and short-term dependence patterns. The important features of the fused carbon data are used to train the long-term prediction training model through the enhanced long short-term memory network; finally, the trained atmospheric carbon data prediction model is used to process the obtained atmospheric carbon data to be predicted, and the prediction result of the atmospheric carbon data is obtained.

[0134] Even further, the improved long short-term memory model inputs the new atmospheric carbon data to be predicted into the trained prediction model for prediction. Its features include using the enhanced long short-term memory network (ELSTM) for data prediction, and inputting the atmospheric carbon concentration data to be predicted into the trained prediction model. Using the relationships between multiple variables learned by the model, the future trend of the atmospheric carbon concentration is predicted. The mean square performance (MSE) and the mean absolute performance (MAE) are used as model evaluation indicators to evaluate the performance of the model.

[0135] The improvement process of LSTM includes: modifying its structure to enhance the ability to capture long-term and short-term dependence patterns. The improved LSTM network includes ELSTM and a convolutional fusion layer;

[0136] Input window: Set the input prediction window to 512 steps, and model multiple variables (such as XCO2, carbon sink, etc.). Through the analysis of historical data, the model can predict the future carbon concentration based on the current multi-variable time series.

[0137] Prediction measurements: The output prediction results include multiple prediction measurements, specifically 96 steps, 192 steps, 336 steps, and 720 steps, corresponding to different prediction accuracies and different prediction requirements.

[0138] Model evaluation: Using the mean square deviation (MSE) and mean absolute error (MAE) evaluation metrics, quantitatively analyze the prediction performance of the model. MSE and MAE calculate the differences between the predicted values and the true values respectively, so as to evaluate the accuracy of the prediction results.

[0139] Furthermore, for the optimized solution, input the atmospheric carbon concentration data to be predicted into the trained prediction model. Utilize the relationships between multiple variables learned by the model to predict the future trend of atmospheric carbon concentration. Output the prediction results of atmospheric carbon concentration, including the change trend of carbon concentration within a certain future time. On the basis of considering multiple carbon data variables such as carbon sources and sinks, adopt the channel dependence mode to train and predict multiple variables such as future XCO2 and carbon sinks. For multi-variable prediction of multi-variables, the prediction model can use multiple input variables to predict multiple output variables. In the present invention, the selected input prediction window is set to a step size of 512. After modeling multiple variables such as XCO2 and carbon sinks, the prediction windows output are of 4 prediction scales with step sizes of 96, 192, 336, and 720. The experiment selects MSE and MAE as the evaluation metrics of the model, and quantitatively analyzes and evaluates the experimental performance of the model. Mean square error (MSE) and mean absolute error (MAE) are two commonly used prediction error evaluation metrics. They calculate the differences between the predicted values and the true values respectively. Use them as the final prediction evaluation metrics of the model.

[0140] Its main steps include:

[0141] Step1: Data preprocessing and model input

[0142] Step1-1: Data preprocessing and pre-training: First, perform preprocessing steps such as rasterization and alignment of urban longitude and latitude for the atmospheric carbon concentration and carbon sink data to be predicted. This step ensures the alignment of the data in terms of time and regional longitude and latitude. After that, the multi-variable data undergoes various preprocessing steps including instance normalization, patch segmentation, etc.

[0143] Step1-2: Model input: For the preprocessed multiple datasets (including private datasets and public datasets) including multiple variables (such as XCO2 and carbon sinks, etc.), after these variables are aligned, they are used as input data and fed into the pre-trained model for pre-training.

[0144] Step2: Model training and testing process

[0145] Step2-1: Determine the number of wavelet decomposition levels and the type of wavelet basis. The number of wavelet decomposition levels is determined according to the maximum decomposition level, and its formula is given above. The type of wavelet basis is determined by a large number of experiments to obtain the best wavelet basis combination.

[0146] Step2-3: Determine the size of the threshold and the parallel wavelet basis weights, and update using gradient descent.

[0147] Step2-4: Determine the local weights after wavelet denoising and the global weights after Fourier processing

[0148] Step2-5: Model learning: The prediction model uses the relationships between multiple variables learned through a series of processes such as denoising, feature extraction, and feature fusion during pre-training and training to predict the future trend of atmospheric carbon concentration. The model analyzes the temporal dependence and relationships between multiple variables in historical data to predict the future trend of carbon concentration changes.

[0149] Step2-6: Output prediction results: The model outputs the predicted results of carbon concentration for a future period, including the predicted trends of concentration changes of variables such as XCO2 and carbon sinks for a future period. These prediction results can be used to evaluate future environmental changes and formulate corresponding policy measures.

[0150] Step3: Multi-variable and multi-scale window prediction settings

[0151] Step3-1: Prediction window settings: To achieve predictions at different time scales, multiple prediction window lengths are selected to train and evaluate the model. The input prediction window is set to a step size of 512, and the output prediction windows include four step sizes: 96, 192, 336, and 720. These step sizes represent different prediction time ranges, covering the prediction requirements from long-term to ultra-long-term time series.

[0152] Step3-2: Prediction scale: By modeling variables such as XCO2 and carbon sinks, the model can make predictions at different time scales. These scales provide a multi-angle understanding of carbon concentration changes, helping to more comprehensively evaluate the future trend of carbon concentration.

[0153] Step4: Model performance evaluation and comparison

[0154] Step4-1: Evaluation metrics: Mean Squared Error (MSE) and Mean Absolute Error (MAE) are selected as the evaluation metrics for model performance. These two metrics are used to quantify the accuracy of the prediction results:

[0155] Step4-2: Mean Squared Error (MSE): The Mean Squared Error is used as the main evaluation metric to calculate the average of the squared differences between the predicted values and the true values, which is used to measure the overall prediction error of the model.

[0156] Step4-3: Quantitative analysis and evaluation: Quantitative analysis and comparative evaluation of the experimental performance of the model and other current SOTA models are carried out through MSE and MAE. These metrics help to determine the performance of the model under different prediction windows, guiding further model optimization and adjustment.

[0157] It should be noted that in this embodiment, the prediction results of each model for the TCCON site dataset by the model are shown in Table 1, the experimental results for the carbon source and carbon sink datasets in Xinjiang region are shown in Table 2, the comparative experimental results for the Xianghe site dataset are shown in Table 3, and the ablation experimental results for the Xianghe site dataset are shown in Table 4.

[0158] Table 1

[0159]

[0160] Table 2

[0161]

[0162] Table 3

[0163]

[0164] Table 4

[0165]

[0166] Embodiment 2

[0167] As Figure 6 shown, based on the same inventive concept, this embodiment also provides an atmospheric carbon data prediction system, including:

[0168] A data preprocessing module, configured to preprocess the multi-variable time series carbon data of the TCCON and GONGGA dataset sites;

[0169] A multi-variable time series time-frequency feature extraction and denoising module, configured to divide the periodic trend term, perform time-frequency conversion using parallel multi-scale wavelet transform after trend decomposition, and then perform adaptive threshold denoising processing and feature extraction to obtain the data time-frequency features;

[0170] A multi-variable time series time-frequency feature fusion module, configured to perform weighted fusion on the data time-frequency features through a feature fusion method;

[0171] A multi-scale convolutional layer feature fusion module, configured to adjust the adaptive weights and update the parameters based on gradient descent;

[0172] An atmospheric carbon data prediction model training and optimization module, configured to train an improved long short-term memory network model based on the fused features;

[0173] An atmospheric carbon data prediction module, configured to perform atmospheric carbon concentration prediction to obtain prediction results.

[0174] An atmospheric carbon data prediction system provided in this embodiment has all the advantages of the atmospheric carbon data prediction method provided in Embodiment 1.

[0175] The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting atmospheric carbon data, characterized in that: include: Preprocessing of multivariate time series carbon data of TCCON and GONGGA dataset sites; Divide the periodic trend items, use parallel multi-scale wavelet transform to perform time-frequency conversion after trend decomposition, and then perform adaptive threshold denoising and feature extraction to obtain the time-frequency characteristics of the data; The time-frequency features of the data are weightedly fused through a feature fusion method, the adaptive weights are adjusted and the parameters are updated based on gradient descent, and an improved long short-term memory network model is trained based on the fused features to predict the atmospheric carbon concentration and obtain the prediction results.

2. The method according to claim 1, characterized in that Preprocessing of multivariate time series carbon data from TCCON and GONGGA datasets, including regional rasterization and longitude and latitude alignment, missing value interpolation, and self-supervised pre-training; The regional rasterization and longitude and latitude alignment include: The global carbon flux data are collected as regional carbon flux data through the GONGGA inversion system, and then the carbon flux data are rasterized; the processed carbon flux data set is further processed to separate the negative net carbon flux as a carbon sink for separate analysis; The longitude and latitude of the drawn city or province boundary map are aligned to obtain the ground carbon flux data, and the missing data of the ground carbon flux data are interpolated and supplemented to obtain a complete data set.

3. The method according to claim 2, characterized in that The missing value interpolation includes: regional gridding and latitude and longitude calibration, output value interpolation, sampling interval sampling, outlier extraction, and regularization; The regional rasterization and longitude and latitude calibration are to rasterize the original data according to the region and calibrate the longitude and latitude of different data sources; The output value interpolation is to interpolate the output value in the data using a linear interpolation method; The sampling interval sampling is to sample the data at fixed time intervals; The outlier extraction is to use a box plot method to identify outliers in the data and extract the outliers; The regularization is to perform regularization processing on the data, standardize the features and set the same scale range.

4. The method according to claim 1, characterized in that: Divide the periodic trend items, use parallel multi-scale wavelet transform to perform time-frequency conversion after trend decomposition, and then perform adaptive threshold denoising and feature extraction. The process of obtaining the time-frequency characteristics of the data includes: When dividing the data into periodic trends, the sliding average method is used to smooth the data; Use Fourier transform and discrete wavelet transform to transform the data into time and frequency domains respectively, and extract the multi-graphic features of the data; The features extracted by wavelet transform are used for denoising in the frequency domain, and the adaptive threshold denoising method is used to remove the noise components in the frequency domain.

5. The method according to claim 4, characterized in that The process of smoothing data using the sliding average method includes: First, the multivariate time series is divided into several overlapping small blocks to obtain a small block sequence; Next, the original data series is padded using a moving average decomposition block to represent the trend and seasonal components, and then average pooling is applied to extract the trend component, which is then subtracted from the overall data to separate the seasonal component.

6. The method according to claim 1, characterized in that The process of weighted fusion of the time-frequency features of the data by the feature fusion method includes: According to the time-frequency features extracted by multi-dimensional wavelet transform, the dimensional features of different dimensions are separated and assigned different weights to different key dimensional features; The topological layer is used to extract local features, and the linear layer aggregates multi-scale features, extracts local dependencies in the time series through the morphological layer, and integrates multi-scale measurements through the linear layer; The processed data is converted from the frequency domain to the time domain, and the visualization prediction data and the verification data are subjected to inverse Fourier transform respectively.

7. The method according to claim 1, characterized in that Adaptive weights and parameter updates are adjusted based on gradient descent, including adaptive weights, optimal wavelet basis and threshold parameter updates.

8. The method according to claim 7, characterized in that The adaptive weights include global weights represented by Fourier transform and local weights represented by wavelet transform; The selection of the optimal wavelet basis includes selecting a wavelet basis combination by analyzing the quality of the experimental results of each data set and dynamically adjusting the corresponding weights according to the contribution; The threshold parameter update includes assigning different initial thresholds to the seasonal item and the trend item respectively, using the product of the energy value and the initial threshold and continuously updating the initial threshold, so that the threshold is continuously updated to select the optimal threshold; and dynamically adjusting according to the transformation of the data set and the changes in the time series of different time periods.

9. The method according to claim 7, characterized in that: Based on the fused features, the improved long short-term memory network model is trained to predict atmospheric carbon concentration. The process of obtaining the prediction results includes: The improved long short-term memory model is used to input new atmospheric carbon data to be predicted into the trained prediction model to predict the future trend of atmospheric carbon concentration. The mean square performance and average absolute performance are used as model evaluation indicators to evaluate the performance of the model.

10. An atmospheric carbon data prediction system, characterized in that: include: Data preprocessing module, used to preprocess multivariate time series carbon data of TCCON and GONGGA dataset sites; The multivariate time series time-frequency feature extraction and denoising module is used to divide the periodic trend items. After trend decomposition, parallel multi-scale wavelet transform is used for time-frequency conversion, and then adaptive threshold denoising and feature extraction are performed to obtain the data time-frequency features. A multivariate time series time-frequency feature fusion module, used to perform weighted fusion of the data time-frequency features through a feature fusion method; Multi-scale convolutional layer feature fusion module, used to adjust adaptive weights and parameter updates based on gradient descent; Atmospheric carbon data prediction model training and optimization module, used to train improved long short-term memory network models based on fused features; The atmospheric carbon data prediction module is used to predict the atmospheric carbon concentration and obtain the prediction results.

Images