Carbon Dioxide Column Concentration Prediction Method, System and Electronic Device Based on the Fusion Model of Generative Diffusion and Atmospheric Diffusion

Abstract

The present application discloses a method, system and electronic device for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, belonging to the technical field of environmental prediction. It includes obtaining the moment dynamic characteristics of carbon dioxide column concentration data and meteorological data in the target area, as well as the spatio-temporal dynamic characteristics of regional grid meteorological data in the target area, and fusing the moment dynamic characteristics and spatio-temporal dynamic characteristics; performing forward diffusion operation and reverse denoising on the fused characteristics to obtain a preliminary prediction result; and constraining the smooth transition between adjacent moments in the prediction result through local smoothness loss, and constraining the consistency of the change trend between the preliminary prediction result and the existing data through trend consistency loss. The present application integrates spatio-temporal feature fusion, performs forward noise addition and reverse denoising reconstruction on prediction data based on the diffusion model, and ensures the smooth continuity of the time series and the overall physical rationality of the prediction result through local and trend consistency losses, thereby improving the prediction accuracy.

Description

Technical Field

[0001] This application belongs to the technical field of environmental prediction, and particularly relates to a method, system, and electronic device for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion. Background Art

[0002] The accurate prediction of carbon dioxide concentration is of great significance for climate research, environmental management, and policy making. Traditional methods mainly rely on atmospheric diffusion models (such as the WRF model), which simulate the diffusion process based on physical equations, but their accuracy is limited by the quality of input parameters, grid resolution, and cumulative errors in long-term prediction. In addition, physical models are difficult to capture complex spatio-temporal patterns and non-linear relationships in the data. At the same time, due to the limited spatial and temporal resolution of satellite XCO2 data, it is impossible to accurately reflect local complex terrain or short-term emission peaks, resulting in low prediction accuracy of carbon concentration in small-scale areas such as cities or industrial parks. Summary of the Invention

[0003] Object of the Invention: This application develops a method, system, and electronic device for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, aiming to solve the technical problem of low prediction accuracy of carbon concentration in the prior art.

[0004] Technical Solution: In a first aspect, an embodiment of this application provides a method for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, including:

[0005] Identifying a target area, and obtaining first data and second data at multiple existing moments in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes regional grid meteorological data of the target area;

[0006] Fusing the temporal dynamic characteristics of the first data and the spatio-temporal dynamic characteristics of the second data to obtain fused characteristics;

[0007] Performing forward diffusion and reverse denoising on the fused characteristics to obtain a preliminary prediction result;

[0008] Constraining the smooth transition between adjacent moments in the preliminary prediction result based on local smoothness loss, and constraining the consistency between the preliminary prediction result and the change trend of the existing carbon dioxide column concentration data based on trend consistency loss to obtain a final prediction result.

[0009] In some embodiments, the step of obtaining the temporal dynamic characteristics of the first data includes:

[0010] Converting the first data at each moment into a first vector; the representation formula of the first vector includes:

[0011] ;

[0012] in, For the corresponding time The first vector of , is the number of moments; The dimension is The matrix, is the number of parameter types in the first data;

[0013] Based on the sequence of the moments, the hidden state and the memory state of the multi-layer long short-term memory network are updated by using the plurality of the first vectors; the representation formula thereof includes:

[0014] ;

[0015] in, For the corresponding time The hidden state of the multi-layer long short-term memory network; For the corresponding time The memory unit state of the multi-layer long short-term memory network; It is a multi-layer long short-term memory network; For the corresponding time The first vector of For the corresponding time The hidden state of the multi-layer long short-term memory network; For the corresponding time The memory unit state of the multi-layer long short-term memory network;

[0016] The sequence of the hidden states after outputting the update is the moment dynamic feature of the first data.

[0017] In some embodiments, the step of obtaining the spatiotemporal dynamic characteristics of the second data includes:

[0018] Obtain grid information of the second data, and convert each second data into a second vector based on the grid information; the representation formula of the second vector includes:

[0019] ;

[0020] in, For the corresponding time The second vector of , is the number of moments; The dimension is The matrix, is the height of the grid, is the width of the grid, is the number of data channels;

[0021] Input the second data into a convolutional neural network to obtain the spatial features of the second data;

[0022] Input the spatial features into a multi-layer long short-term memory network, and based on the order of the moments, extract the spatio-temporal dynamic features of the second data through multiple of the spatial features; The representation formula includes:

[0023] ;

[0024] where, is the spatio-temporal dynamic feature of the second data; is the multi-layer long short-term memory network; is the spatial feature of the second data corresponding to the moment ; is the number of moments;[[ID=2,4]] is a matrix with dimension ; is the dimension of the moment dynamic feature of the output second data.

[0025] In some embodiments, the step of fusing the moment dynamic feature of the first data and the spatio-temporal dynamic feature of the second data to obtain a fusion feature includes:

[0026] Based on the dimension of the spatio-temporal dynamic feature, splice the moment dynamic feature and the spatio-temporal dynamic feature to obtain a spliced feature;

[0027] Based on the self-attention mechanism, perform weighted processing on the spliced feature to obtain a weighted feature;

[0028] Based on the self-attention mechanism, obtain the similarity between features in the weighted feature, and based on the similarity, increase the weight of key features to obtain the fusion feature.

[0029] In some embodiments, the step of forward diffusion includes:

[0030] Obtain the weight coefficients corresponding to each moment;

[0031] Based on the weight coefficients, add Gaussian noise to the fusion feature to obtain the noise sequence at each moment; The representation formula includes:

[0032] ;

[0033] where, is the noise sequence corresponding to the th moment; is the fusion feature; is the weight coefficient at the corresponding time t; is the Gaussian noise.

[0034] In some embodiments, the step of reverse denoising includes:

[0035] Based on the weight coefficient, perform reverse denoising operation on the noise sequence through a noise prediction network to obtain a preliminary prediction result; its characterization formula includes:

[0036] ;

[0037] where is the result after reverse denoising at time is the noise intensity parameter in the t-th diffusion process, used to control the degree of Gaussian perturbation, ; is the noise prediction network; is the existing meteorological data, time encoding, and carbon dioxide column concentration distribution; is a random noise vector sampled from the standard normal distribution, used to simulate the noise in the diffusion process.

[0038] In some embodiments, the characterization formula of the local smoothness loss includes:

[0039] ;

[0040] ;

[0041] where is the local smoothness loss; is the time span of future prediction, used to balance the growth amplitude of trends under different time windows; is the serial number of the time; is in the reverse diffusion process the generation result at time is in the reverse diffusion process the generation result at time (·) is the dynamic weight function; is the control parameter for adjusting the deviation degree, used to control the decline rate of the weight; is the mean value of is the standard deviation of

[0042] In some embodiments, the characterization formula of the trend consistency loss includes:

[0043] ;

[0044] where is the trend consistency loss; is the average change trend of the predicted future data sequence; is the average change trend of the existing data sequence; is the explicit slope adjustment term, is the preset slope of the existing upward trend, is the future prediction time span, which is used to balance the growth amplitude of the trend under different time windows.

[0045] In a second aspect, an embodiment of the present application further provides a carbon dioxide column concentration prediction system based on a generative diffusion and atmospheric diffusion fusion model, including:

[0046] A data acquisition module, which is used to confirm the target area and acquire the first data and the second data at multiple existing moments in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes the regional grid meteorological data of the target area;

[0047] A spatio-temporal feature fusion module, which is used to fuse the moment dynamic features of the first data and the spatio-temporal dynamic features of the second data to obtain fusion features;

[0048] A concentration prediction module, which is used to perform forward diffusion and reverse denoising on the fusion features to obtain a preliminary prediction result;

[0049] A moment consistency constraint module, which is used to constrain the smooth transition between adjacent moments in the preliminary prediction result based on the local smoothness loss, and constrain the consistency between the preliminary prediction result and the change trend of the existing carbon dioxide column concentration data based on the trend consistency loss to obtain a final prediction result.

[0050] In a third aspect, an embodiment of the present application further provides an electronic device, including a memory and a processor, where the memory stores a computer program executable by the processor, and when the processor executes the computer program, it implements the steps of the carbon dioxide column concentration prediction method based on the generative diffusion and atmospheric diffusion fusion model as described in any item of the first aspect.

[0051] Beneficial effects: Compared with the prior art, a carbon dioxide column concentration prediction method based on a fusion model of generative diffusion and atmospheric diffusion provided by an embodiment of the present application includes obtaining the moment dynamic characteristics of carbon dioxide column concentration data and meteorological data in a target area, as well as the spatio-temporal dynamic characteristics of regional grid meteorological data in the target area, and fusing the moment dynamic characteristics and the spatio-temporal dynamic characteristics to obtain fusion characteristics; performing a forward diffusion operation and a reverse denoising operation on the fusion characteristics to obtain a preliminary prediction result; and ensuring the smooth transition between adjacent moments in the prediction result through local smoothness loss constraints, ensuring the smooth continuity of the time series, and ensuring the consistency of the change trend between the preliminary prediction result and the existing data through trend consistency loss constraints, ensuring the overall physical rationality. The present application integrates spatio-temporal features to provide sufficient conditional information for the diffusion model, then performs forward noise addition and reverse denoising on the fusion features based on the diffusion model to reconstruct future prediction data, and finally ensures the smooth continuity of the time series and the overall physical rationality of the prediction result through local and trend consistency losses, improving the prediction accuracy of carbon concentration. Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 It is a flowchart of the steps of the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application;

[0054] Figure 2 It is a flowchart of the steps of obtaining the moment dynamic characteristics of the first data in the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application;

[0055] Figure 3 It is a flowchart of the steps of obtaining the spatio-temporal dynamic characteristics of the second data in the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application;

[0056] Figure 4 It is a flowchart of the steps of fusing the moment dynamic characteristics of the first data and the spatio-temporal dynamic characteristics of the second data in the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application;

[0057] Figure 5 It is a module connection diagram of the carbon dioxide column concentration prediction system based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application;

[0058] Figure 6 Structural diagram of the electronic device provided by the embodiment of the present application;

[0059] Reference numerals: 10, data acquisition module; 20, spatio-temporal feature fusion module; 30, concentration prediction module; 40, moment consistency constraint module; 100, memory; 200, processor. Specific embodiments

[0060] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0061] The accurate prediction of carbon dioxide concentration is of great significance for climate research, environmental management, and policy making. Traditional methods mainly rely on atmospheric diffusion models (such as the WRF model), which simulate the diffusion process based on physical equations, but their accuracy is limited by the quality of input parameters, grid resolution, and cumulative errors in long-term predictions. In addition, physical models are difficult to capture complex spatio-temporal patterns and non-linear relationships in the data. It should be noted that there is currently a lack of relevant research using deep learning algorithms, especially deep learning algorithms based on the XCO2 dataset, to accurately predict future XCO2. The main advantage of deep learning methods lies in their powerful ability to automatically learn high-level features from a wide range of datasets, which is a key step in bridging the gap between data patterns at different feature levels. Given the excellent feature extraction performance of deep learning neural networks, they have great potential in fusing multi-source data to extract key information.

[0062] The following engineering problems still remain to be solved:

[0063] Problem 1: The spatial and temporal resolutions of satellite XCO2 data are limited and cannot accurately reflect local complex terrains or short-term emission peaks, affecting the prediction accuracy of carbon concentration in small-scale areas such as cities or industrial parks.

[0064] Problem 2: The lack or imbalance of observational data leads to data sparsity problems in carbon concentration prediction models based on satellite remote sensing, reducing the reliability of predictions.

[0065] In view of this, an embodiment of the present application provides a method for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, including obtaining the moment dynamic characteristics of carbon dioxide column concentration data and meteorological data in a target area, as well as the spatio-temporal dynamic characteristics of regional grid meteorological data in the target area, and fusing the moment dynamic characteristics and spatio-temporal dynamic characteristics to obtain fusion characteristics; performing a forward diffusion operation and a reverse denoising operation on the fusion characteristics to obtain a preliminary prediction result; and ensuring the smooth transition between adjacent moments in the prediction result through local smoothness loss constraints, ensuring the smooth continuity of the moment sequence, and ensuring the consistency of the change trend between the preliminary prediction result and the existing data through trend consistency loss constraints, ensuring the overall physical rationality. The present application integrates spatio-temporal features to provide sufficient conditional information for the diffusion model, then performs forward noise addition and reverse denoising on the fusion features based on the diffusion model to reconstruct future prediction data, and finally ensures the smooth continuity of the moment sequence and the overall physical rationality of the prediction result through local and trend consistency losses, improving the prediction accuracy of carbon concentration.

[0066] In some embodiments, please refer to Figure 1 , Figure 1 is the step flow chart of the method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application. The method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application is specifically implemented through steps 100 to 400:

[0067] Step 100: Confirm the target area and obtain the first data and the second data at multiple moments existing in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes regional grid meteorological data of the target area.

[0068] In some embodiments, the historical carbon dioxide column concentration data and meteorological data are normalized to obtain the first data, where the meteorological data includes parameters such as wind speed, temperature, and humidity; the regional grid meteorological data output by the WRF model is subjected to convolutional dimensionality reduction and time series modeling to obtain the second data.

[0069] Step 200: Fuse the moment dynamic characteristics of the first data and the spatio-temporal dynamic characteristics of the second data to obtain fusion characteristics.

[0070] In some embodiments, please refer to Figure 2 , Figure 2 is the step flow chart of the method for obtaining the moment dynamic characteristics of the first data in the method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiment of the present application. The steps for obtaining the moment dynamic characteristics of the first data are specifically implemented through steps 210 to 230:

[0071] Step 210: Convert the first data at each moment into a first vector.

[0072] In some embodiments, the first data includes existing carbon dioxide column concentration data and meteorological data at several moments, and the representation formula of the first data at each moment includes:

[0073] ;

[0074] where is the first vector corresponding to the moment , , is the number of moments; is a matrix with a dimension of , is the number of parameter types in the first data. For example, the parameter types include carbon dioxide column concentration, wind speed, temperature, humidity, etc.

[0075] Step 220: Based on the order of moments, update the hidden state and memory state of the multi-layer long short-term memory network through multiple first vectors.

[0076] In some embodiments, to fully capture the time dynamic characteristics, this application uses a multi-layer long short-term memory network (LSTM) for modeling, and the steps are as follows:

[0077] Obtain a first input sequence based on the first vectors at multiple moments and perform normalization processing; where the representation formula of the first input sequence includes:

[0078] ;

[0079] where is the first input sequence; is the first vector corresponding to the moment , , is the number of moments.

[0080] Based on the order of moments, the representation formula for updating the hidden state and memory state of the multi-layer long short-term memory network through multiple first vectors includes:

[0081] ;

[0082] where is the hidden state of the multi-layer long short-term memory network corresponding to the moment ; is the memory cell state of the multi-layer long short-term memory network corresponding to the moment ; is the multi-layer long short-term memory network; is the corresponding moment The first vector; is the corresponding time of the hidden state of the multi-layer long short-term memory network; is the corresponding time of the memory cell state of the multi-layer long short-term memory network.

[0083] Step 230: Output the updated sequence of hidden states as the temporal dynamic features of the first data.

[0084] Specifically, after iterating through all time steps, the temporal dynamic features of the first data are obtained , is a matrix with a dimension of , is the output feature dimension.

[0085] In some embodiments, please refer to Figure 3 , Figure 3 is the flowchart of the steps for obtaining the spatio-temporal dynamic features of the second data in the carbon dioxide column concentration prediction method based on the generative diffusion and atmospheric diffusion fusion model provided by the embodiments of the present application. The steps for obtaining the spatio-temporal dynamic features of the second data are specifically implemented through steps 2,40 to 2,60:

[0086] Step 240: Obtain the grid information of the second data, and convert each second data into a second vector based on the grid information.

[0087] In some embodiments, the representation formula of the second vector includes:

[0088] ;

[0089] where, is the second vector at the corresponding time , , is the number of time steps; is a matrix with a dimension of , is the height of the grid, is the width of the grid, is the number of data channels.

[0090] Furthermore, the regional gridded meteorological data is the regional gridded data output by WRF , which has obvious spatial correlation.

[0091] Step 250: Input the second data into a convolutional neural network to obtain the spatial features of the second data.

[0092] In some embodiments, for the second data at each moment, a convolutional neural network (CNN) is used to extract spatial features, and the specific process is as follows:

[0093] For each Perform multi-layer convolutional operations, and use the following formula to extract local spatial features:

[0094] ;

[0095] Wherein, and Are local spatial features; Is a convolutional operation; Is a matrix with a dimension of ; Is the number of feature channels extracted by the convolutional network.

[0096] Step 260: Input the spatial features into a multi-layer long short-term memory network, and based on the order of moments, extract the spatio-temporal dynamic features of the second data through multiple spatial features.

[0097] In some embodiments, the representation formula for extracting the spatio-temporal dynamic features of the second data through multiple spatial features includes:

[0098] ;

[0099] Wherein, Is the spatio-temporal dynamic feature of the second data; Is a multi-layer long short-term memory network; Is the spatial feature of the second data corresponding to the moment ; Is the number of moments; Is a matrix with a dimension of ; Is The dimension of the moment dynamic feature of the output second data.

[0100] In some embodiments, please refer to Figure 4 , Figure 4 Is the flowchart of the steps for fusing the moment dynamic feature of the first data and the spatio-temporal dynamic feature of the second data in the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided by the embodiments of the present application. The steps for fusing the moment dynamic feature of the first data and the spatio-temporal dynamic feature of the second data are specifically implemented through steps 270 to 290:

[0101] Step 270: Based on the dimension of the spatio-temporal dynamic feature, splice the moment dynamic feature and the spatio-temporal dynamic feature to obtain a spliced feature.

[0102] In some embodiments, the representation formula for the spliced feature includes:

[0103] ;

[0104] Among them, is the splicing feature; is the moment dynamic feature of the first data; is the spatio-temporal dynamic feature of the second data; is a matrix with a dimension of ).

[0105] Step 280: Perform weighted processing on the splicing feature based on the self-attention mechanism to obtain a weighted feature.

[0106] In some embodiments, in order to model the interaction relationship between features from different sources, the self-attention mechanism is used to weight the splicing feature, and its representation formula includes:

[0107] ;

[0108] Among them, is the self-attention mechanism; is the query vector, representing the target position where the attention needs to be calculated currently, = , is the key vector, which interacts with Q and is used to calculate the similarity, = , is the value vector, storing the information that actually needs to be extracted, = , is a learnable linear transformation matrix, is the splicing feature; is the normalized exponential function; is the transpose of K, is the dimension, is the transpose symbol.

[0109] Step 290: Obtain the similarity between features in the weighted feature based on the self-attention mechanism, and increase the weight of key features based on the similarity to obtain a fused feature.

[0110] It can be understood that the self-attention mechanism assigns higher weights to key features by calculating the similarity between features, thereby generating a fused spatio-temporal feature representation:

[0111] ;

[0112] Among them, is the fused feature; is the self-attention mechanism; is the splicing feature; is a matrix with a dimension of .

[0113] Step 300: Perform forward diffusion and reverse denoising on the fused features based on the diffusion model to obtain a preliminary prediction result.

[0114] In some embodiments, when performing forward diffusion, obtain the weight coefficients corresponding to each moment, and in the diffusion model, add Gaussian noise to the fused features based on the weight coefficients to obtain the noise sequences at each moment; its representation formula includes:

[0115] ;

[0116] where is the noise sequence corresponding to the th moment; is the fused feature; is the weight coefficient corresponding to the moment t; is the Gaussian noise.

[0117] After steps, basically degenerates into pure noise, laying the foundation for the reverse denoising diffusion process.

[0118] In some embodiments, when performing reverse denoising, based on the weight coefficients, perform reverse denoising operations on the noise sequences through the noise prediction network KAN to obtain a preliminary prediction result; its representation formula includes:

[0119] ;

[0120] where is the result after reverse denoising at the moment; is the noise intensity parameter in the t-th step of the diffusion process, used to control the degree of Gaussian perturbation, ; is the noise prediction network; is the existing meteorological data, time encoding, and carbon dioxide column concentration distribution; is a random noise vector sampled from the standard normal distribution, used to simulate the noise in the diffusion process.

[0121] After steps of iteration, the future XCO2 concentration distribution map predicted by the system is output. This generation process makes full use of the spatio-temporal conditional information extracted by the fusion module and the WRF regional meteorological data, significantly improving the prediction results in terms of spatial resolution and physical rationality.

[0122] In some embodiments, to train the noise prediction network KAN, the present application uses the standard mean squared error (MSE) loss function to fit the predicted noise and the actual sampled noise, and at the same time introduces the reconstruction error as an auxiliary loss. During the training process, the historical data and its forward diffusion obtained are correspondingly used for supervised learning to gradually optimize the network parameters θ, so that KAN can accurately predict the noise components added at each time step, thereby ensuring the stability and effectiveness of the inverse diffusion process.

[0123] Step 400: Based on the local smoothness loss constraint, make the transition between adjacent moments in the preliminary prediction result smooth, and based on the trend consistency loss constraint, ensure the consistency between the preliminary prediction result and the changing trend of the existing carbon dioxide column concentration data, and obtain the final prediction result.

[0124] In some embodiments, when the preliminary prediction result is within the confidence range, MSE is used to evaluate the consistency. When the preliminary prediction result exceeds the confidence interval, it is considered not to directly penalize sudden events, but to weaken the penalty by reducing the weight, reflecting the inclusiveness of sudden situations. The representation formula of the local smoothness loss includes:

[0125] ;

[0126] ;

[0127] where is the local smoothness loss; is the time span of future prediction, used to balance the growth amplitude of trends under different time windows; is the serial number of the moment; is the generation result at the moment in the inverse diffusion process; is the generation result at the moment in the inverse diffusion process; (·) is the dynamic weight function; is the control parameter for adjusting the deviation degree, used to control the descending rate of the weight; is the mean value of; is the standard deviation of.

[0128] It can be understood that this local smoothness loss formula ensures the smoothness and continuity of the time series generated by the model by minimizing the difference between the prediction results of adjacent time steps. If there are sudden events (such as fires, large-scale emissions, etc.) in the data, this loss function will give greater tolerance to the prediction results of these events. It conforms to the physical characteristics of the carbon concentration prediction task, because the change of XCO2 is usually smooth and gradual, but there are still occasional sudden situations.

[0129] In some embodiments, in addition to local smoothness, the overall future prediction sequence should maintain a long-term change trend similar to that of historical data. To this end, this embodiment proposes a long-term trend consistency loss , and the representation formula of the trend consistency loss includes:

[0130] ;

[0131] where, is the trend consistency loss; is the average change trend of the predicted future data sequence; is the average change trend of the existing data sequence; is an explicit slope adjustment term, is the slope of the preset existing upward trend, is the time span of future prediction, which is used to balance the growth amplitude of the trend under different time windows.

[0132] It can be understood that this formula keeps the trend change amount of future prediction consistent with that of historical trend, and at the same time does not ignore a certain growth deviation of XCO2.

[0133] In some embodiments, to take into account both local smoothness and overall trend consistency, during model training, the two parts of the loss are weighted and summed to obtain the final time consistency loss :

[0134] ;

[0135] where, and are hyperparameters used to balance local and global constraints, and their values can be determined by cross-validation; is the model loss.

[0136] During the training process, and the original loss of the diffusion model are used as the total loss for backpropagation, so as to prompt the generated XCO2 prediction sequence to be both continuously smooth in time and reflect the historical long-term trend change.

[0137] Understandably, an embodiment of the present application provides a method for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, including obtaining the moment dynamic characteristics of carbon dioxide column concentration data and meteorological data in a target area, as well as the spatio-temporal dynamic characteristics of regional grid meteorological data in the target area, and fusing the moment dynamic characteristics and the spatio-temporal dynamic characteristics to obtain fusion characteristics; performing a forward diffusion operation and a reverse denoising operation on the fusion characteristics to obtain a preliminary prediction result; and ensuring the smooth transition between adjacent moments in the prediction result through a local smoothness loss constraint, and ensuring the consistency of the change trend between the preliminary prediction result and the existing data through a trend consistency loss constraint, so as to ensure the smooth continuity of the time series. The present application integrates spatio-temporal features to provide sufficient conditional information for the diffusion model, then performs forward noise addition and reverse denoising on the fusion features based on the diffusion model to reconstruct future prediction data, and finally ensures the smooth continuity of the time series and the overall physical rationality of the prediction result through local and trend consistency losses, improving the prediction accuracy of carbon concentration.

[0138] Through the detailed description of the present application, it can be seen that the solution provided by the present application realizes high-precision prediction of the future XCO2 concentration distribution by integrating spatio-temporal feature fusion, a prediction network based on a diffusion model and WRF data, and time consistency constraints. The solution of the present application uses LSTM and CNN to extract the features of historical and regional meteorological data respectively, and performs deep fusion through a self-attention mechanism to provide sufficient conditional information for the diffusion model; during the generation process, a forward noise addition and reverse denoising strategy is adopted to reconstruct future data, and the smooth continuity of the time series and the overall physical rationality are ensured through local and trend consistency losses. Experimental results show that the method of the present application is superior to traditional prediction methods in terms of prediction accuracy, spatio-temporal consistency, and physical constraint satisfaction, and has high engineering application value.

[0139] The implementation details of the above steps can be adjusted and expanded according to actual needs to ensure good results under different regions and different data conditions. The solution of the present application fully embodies the innovative idea of combining a deep generative model with physical meteorological data to realize complex spatio-temporal dynamic modeling, providing a new technical path for carbon concentration prediction. The following problems can be solved:

[0140] Problem 1: The spatial and temporal resolutions of satellite XCO2 data are limited and cannot accurately reflect local complex terrains or short-term emission peaks, affecting the prediction accuracy of carbon concentration in small-scale areas such as cities or industrial parks.

[0141] Problem 2: The lack or imbalance of observational data leads to a data sparsity problem in the carbon concentration prediction model based on satellite remote sensing, reducing the reliability of the prediction.

[0142] Therefore, more accurate prediction results can be obtained in the carbon concentration prediction task.

[0143] Correspondingly, an embodiment of the present application further provides a carbon dioxide column concentration prediction system based on a generative diffusion and atmospheric diffusion fusion model. Please refer to Figure 5 , Figure 5 which is a module connection diagram of the carbon dioxide column concentration prediction system provided by the embodiment of the present application. The carbon dioxide column concentration prediction system includes:

[0144] A data acquisition module 10, which is used to confirm the target area and acquire the first data and the second data at multiple existing moments in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes regional grid meteorological data of the target area;

[0145] A spatio-temporal feature fusion module 20, which is used to fuse the moment dynamic features of the first data and the spatio-temporal dynamic features of the second data to obtain fused features;

[0146] A concentration prediction module 30, which is used to perform forward diffusion and reverse denoising on the fused features to obtain a preliminary prediction result;

[0147] A moment consistency constraint module 40, which is used to constrain the smooth transition between adjacent moments in the preliminary prediction result based on the local smoothness loss, and constrain the consistency between the preliminary prediction result and the change trend of the existing carbon dioxide column concentration data based on the trend consistency loss to obtain the final prediction result.

[0148] In some embodiments, the spatio-temporal feature fusion module 20 is specifically used for:

[0149] Convert the first data at each moment into a first vector; the representation formula of the first vector includes:

[0150] ;

[0151] where is the first vector corresponding to the moment , , is the number of moments; is a matrix with a dimension of , is the number of parameter types in the first data;

[0152] Based on the order of the moments, update the hidden state and memory state of the multi-layer long short-term memory network through multiple first vectors; its representation formula includes:

[0153] ;

[0154] Among them, is the hidden state of the multi-layer long short-term memory network at the corresponding moment; is the memory cell state of the multi-layer long short-term memory network at the corresponding moment; is the multi-layer long short-term memory network; is the first vector at the corresponding moment; is the hidden state of the multi-layer long short-term memory network at the corresponding moment; is the memory cell state of the multi-layer long short-term memory network at the corresponding moment; is the first vector at the corresponding moment; is the hidden state of the multi-layer long short-term memory network at the corresponding moment; is the memory cell state of the multi-layer long short-term memory network at the corresponding moment; is the memory cell state of the multi-layer long short-term memory network at the corresponding moment; is the memory cell state of the multi-layer long short-term memory network at the corresponding moment;

[0155] The sequence of the updated hidden state output is the moment dynamic feature of the first data.

[0156] In some embodiments, the spatio-temporal feature fusion module 20 is specifically configured to:

[0157] Obtain the grid information of the second data, and convert each second data into a second vector based on the grid information; the representation formula of the second vector includes:

[0158] ;

[0159] Among them, is the second vector at the corresponding moment, is the number of moments; is a matrix with a dimension of is the height of the grid, is the width of the grid, is the data channel number; is the height of the grid, is the width of the grid, is the data channel number;

[0160] Input the second data into the convolutional neural network to obtain the spatial feature of the second data;

[0161] Input the spatial feature into the multi-layer long short-term memory network, and based on the order of moments, extract the spatio-temporal dynamic feature of the second data through multiple spatial features; its representation formula includes: [[ID=6']]

[0162] ;

[0163] Among them, is the spatio-temporal dynamic feature of the second data; is the multi-layer long short-term memory network; is the spatial feature of the second data at the corresponding moment; is the number of moments; is the number of moments; is a matrix with a dimension of . is the dynamic feature dimension of the second output data at a moment.

[0164] In some embodiments, the spatio-temporal feature fusion module 20 is specifically configured to:

[0165] Based on the dimension of the spatio-temporal dynamic feature, splice the moment dynamic feature and the spatio-temporal dynamic feature to obtain a spliced feature;

[0166] Based on the self-attention mechanism, perform weighted processing on the spliced feature to obtain a weighted feature;

[0167] Based on the self-attention mechanism, obtain the similarity between features in the weighted feature, and based on the similarity, increase the weight of the key feature to obtain a fused feature.

[0168] In some embodiments, the concentration prediction module 30 is specifically configured to:

[0169] Obtain the weight coefficient corresponding to each moment;

[0170] In the diffusion model, based on the weight coefficient, add Gaussian noise to the fused feature to obtain the noise sequence at each moment; its representation formula includes:

[0171] ;

[0172] Where is the noise sequence corresponding to the th moment; is the fused feature; <b is the weight coefficient corresponding to the moment t; is Gaussian noise.

[0173] In some embodiments, the concentration prediction module 30 is specifically configured to:

[0174] Based on the weight coefficient, perform an inverse denoising operation on the noise sequence through a noise prediction network to obtain a preliminary prediction result; its representation formula includes:

[0175] ;

[0176] Where is the result after inverse denoising at the moment; is the noise intensity parameter in the t-th diffusion process, used to control the degree of Gaussian perturbation, ; is the noise prediction network; is the existing meteorological data, moment encoding, and carbon dioxide column concentration distribution; is a random noise vector sampled from the standard normal distribution and is used to simulate the noise in the diffusion process.

[0177] Accordingly, please refer to Figure 6 , Figure 6 is the structural diagram of the electronic device provided by the embodiment of the present application. The electronic device provided by the embodiment of the present application includes:

[0178] A memory 100 and a processor 200 are communicatively connected to each other. Computer instructions are stored in the memory 100. The processor 200 executes the computer instructions to perform the carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion provided in the embodiment of the present application. Specifically, the memory 100 and the processor 200 are connected through a communication bus.

[0179] In some embodiments, the processor 200 may be a central processing unit (CPU). The processor 200 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. chips, or combinations of the above types of chips.

[0180] In some embodiments, the memory 100, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the method embodiments of the present application. The processor 200 runs the non-transitory software programs, instructions, and modules stored in the memory 100 to perform various functional applications and data processing of the processor 200, that is, to implement the methods in the above method embodiments.

[0181] In some embodiments, the memory 100 may include a program storage area and a data storage area. The program storage area may store an operating device and application programs required for at least one function. The data storage area may store data created by the processor 200 and the like. In addition, the memory 100 may include a high-speed random access memory 100, and may also include a non-transitory memory 100, such as at least one disk memory device 100, a flash memory device, or other non-transitory solid-state memory devices 100. In some embodiments, the memory 100 may optionally include a memory 100 that is remotely disposed relative to the processor 200, and these remote memories 100 may be connected to the processor 200 through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0182] In some embodiments, one or more modules are stored in the memory 100 and, when executed by the processor 200, perform the methods in the above method embodiments.

[0183] In some embodiments, the specific details of the above electronic device may be understood by referring to the corresponding related descriptions and effects in the above method embodiments, and will not be elaborated here.

[0184] The present application has introduced in detail a carbon dioxide column concentration prediction method, system, and electronic device provided by the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. At the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation on the present application.

Claims

1. A method for predicting carbon dioxide column concentration based on a fusion model of generative diffusion and atmospheric diffusion, characterized in that, Including: Confirm the target area, and obtain the first data and the second data at multiple existing moments in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes the regional grid meteorological data of the target area; Fuse the temporal dynamic characteristics of the first data and the spatio-temporal dynamic characteristics of the second data to obtain fused characteristics; wherein: The steps of obtaining the temporal dynamic characteristics of the first data include: Convert the first data at each moment into a first vector; the representation formula of the first vector includes: Among them, x t is the first vector corresponding to the moment t, where t = 1, 2,..., T input , and T input is the number of moments; is a matrix with dimension d main , and d main is the number of parameter types in the first data, and the parameter types include the carbon dioxide column concentration data parameter and the meteorological data parameter; Based on the order of the moments, update the hidden state and memory state of the multi-layer long short-term memory network through multiple first vectors; its representation formula includes: h t ,c t = LSTM(x t ,h t-1 ,c t-1 ); where h t is the hidden state of the multi-layer long short-term memory network at the corresponding time t; c t is the memory cell state of the multi-layer long short-term memory network at the corresponding time t; LSTM is the multi-layer long short-term memory network; x t is the first vector at the corresponding time t; h t-1 is the hidden state of the multi-layer long short-term memory network at the corresponding time t-1; c t-1 is the memory cell state of the multi-layer long short-term memory network at the corresponding time t-1; Output the sequence of the updated hidden state as the temporal dynamic characteristics of the first data; The steps of obtaining the spatio-temporal dynamic characteristics of the second data include: Obtain the grid information of the second data, and convert each second data into a second vector based on the grid information; the representation formula of the second vector includes: Among them, y t is the second vector corresponding to the moment t, where t = 1, 2,..., T input , T input is the number of moments; is a matrix with dimensions H×W×C, where H is the height of the grid after the meteorological data is gridded, W is the width of the grid after the meteorological data is gridded, and C is the number of data channels; Input the second vector into a convolutional neural network to obtain the spatial characteristics of the second data; Input the spatial characteristics into a multi-layer long short-term memory network, and extract the spatio-temporal dynamic characteristics of the second data through multiple spatial characteristics based on the order of the moments; its representation formula includes: Among them, F aux is the spatio-temporal dynamic feature of the second data; LSTM is the multi-layer long short-term memory network; F spatial,t is the spatial feature of the second data at the corresponding moment t; T input is the number of moments; is a matrix of dimension T inpnt ×d aux_features where d aux_features is the dimension of the moment dynamic feature of the second data output by LSTM; Perform forward diffusion and reverse denoising on the fused characteristics based on a diffusion model to obtain a preliminary prediction result; Constrain the smooth transition between adjacent moments in the preliminary prediction result based on the local smoothness loss, and constrain the consistency between the preliminary prediction result and the change trend of the existing carbon dioxide column concentration data based on the trend consistency loss to obtain the final prediction result.

2. The carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion according to claim 1, wherein The steps of fusing the temporal dynamic characteristics of the first data and the spatio-temporal dynamic characteristics of the second data to obtain fused characteristics include: Based on the dimension of the spatio-temporal dynamic characteristics, splice the temporal dynamic characteristics and the spatio-temporal dynamic characteristics to obtain a spliced characteristic; Perform weighted processing on the spliced characteristic based on a self-attention mechanism to obtain a weighted characteristic; Obtain the similarity between features in the weighted characteristic based on the self-attention mechanism, and increase the weight of key features based on the similarity to obtain the fused characteristic.

3. The method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion according to claim 1, wherein The steps of the forward diffusion include: Obtain the weight coefficients corresponding to each moment; In the diffusion model, add Gaussian noise to the fused characteristic based on the weight coefficients to obtain the noise sequence at each moment; its representation formula includes: Among them, X t is the noise sequence corresponding to the t-th moment; X fused is the fusion feature; α t is the weight coefficient corresponding to the moment t; ∈ is the Gaussian noise.

4. The method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion according to claim 3, wherein The steps of the reverse denoising include: Based on the weight coefficients, perform reverse denoising operations on the noise sequence through a noise prediction network to obtain a preliminary prediction result; its representation formula includes: Among them, X t-1 is the result after reverse denoising at time t-1; β t is the noise intensity parameter in the t-th diffusion process, used to control the degree of Gaussian perturbation, β t ∈(0,1); ∈ θ is the noise prediction network; C is the existing meteorological data, time encoding, and carbon dioxide column concentration distribution; z is a random noise vector sampled from the standard normal distribution, used to simulate the noise in the diffusion process.

5. The method for predicting carbon dioxide column concentration based on the fusion model of generative diffusion and atmospheric diffusion according to claim 1, wherein The representation formula of the local smoothness loss includes: Among them, L local is the local smoothness loss; T future is the time span of future prediction, which is used to balance the growth rate of trends under different time windows; t is the sorting number of time; X t is the generation result at time t in the diffusion inverse process; X t-1 is the generation result at time t - 1 in the diffusion inverse process; ω(·) is the dynamic weight function; α t is the control parameter for adjusting the deviation degree, which is used to control the decline rate of the weight; μ t is the mean value of X t ; σ t is the standard deviation of X t .

6. The carbon dioxide column concentration prediction method based on the fusion model of generative diffusion and atmospheric diffusion according to claim 1, characterized in that The representation formula of the trend consistency loss includes: Among them, L trend is the trend consistency loss; is the average change trend of the predicted future data sequence; is the average change trend of the existing data sequence; k t ·T future is an explicit slope adjustment term, k t is the slope of the preset existing upward trend, T future is the time span of future prediction, which is used to balance the growth amplitude of trends under different time windows.

7. A carbon dioxide column concentration prediction system based on a fusion model of generative diffusion and atmospheric diffusion, characterized in that, Including: A data acquisition module (10), and the data acquisition module (10) is used to confirm the target area and obtain the first data and the second data at multiple existing moments in the target area; the first data includes carbon dioxide column concentration data and meteorological data, and the second data includes the regional grid meteorological data of the target area; A spatio-temporal feature fusion module (20), which is used to fuse the moment dynamic features of the first data and the spatio-temporal dynamic features of the second data to obtain fused features; where: The steps of obtaining the moment dynamic features of the first data include: Converting the first data at each moment into a first vector; the representation formula of the first vector includes: Among them, x t is the first vector corresponding to the moment t, where t = 1, 2,..., T input , and T input is the number of moments; is a matrix with dimension d main , and d main is the number of parameter types in the first data, and the parameter types include the carbon dioxide column concentration data parameter and the meteorological data parameter; Based on the order of the moments, updating the hidden state and memory state of a multi-layer long short-term memory network through multiple of the first vectors; its representation formula includes: h t ,c t = LSTM(x t ,h t-1 ,c t-1 ); where h t is the hidden state of the multi-layer long short-term memory network at the corresponding time t; c t is the memory cell state of the multi-layer long short-term memory network at the corresponding time t; LSTM is the multi-layer long short-term memory network; x t is the first vector at the corresponding time t; h t-1 is the hidden state of the multi-layer long short-term memory network at the corresponding time t - 1; c t-1 is the memory cell state of the multi-layer long short-term memory network at the corresponding time t - 1; Outputting the sequence of the updated hidden state as the moment dynamic features of the first data; The steps of obtaining the spatio-temporal dynamic features of the second data include: Obtaining the grid information of the second data, and converting each of the second data into a second vector based on the grid information; the representation formula of the second vector includes: where y t is the second vector corresponding to the moment t, where t = 1, 2,..., T input , and T tnput is the number of moments; is a matrix of dimension H×W×c, where H is the height of the grid after the meteorological data is gridded, W is the width of the grid after the meteorological data is gridded, and C is the number of data channels; Inputting the second vector into a convolutional neural network to obtain the spatial features of the second data; Inputting the spatial features into a multi-layer long short-term memory network, and extracting the spatio-temporal dynamic features of the second data through multiple of the spatial features based on the order of the moments; its representation formula includes: Among them, F aux is the spatio-temporal dynamic feature of the second data; LSTM is the multi-layer long short-term memory network; F spatial,t is the spatial feature of the second data at the corresponding moment t; T input is the number of moments; is a matrix of dimension T input ×d aux_features where d aux_features is the dimension of the moment dynamic feature of the second data output by the LSTM; A concentration prediction module (30), which is used to perform forward diffusion and reverse denoising on the fused features to obtain a preliminary prediction result; A moment consistency constraint module (40), which is used to constrain the smooth transition between adjacent moments in the preliminary prediction result based on the local smoothness loss, and to constrain the consistency between the preliminary prediction result and the change trend of the existing carbon dioxide column concentration data based on the trend consistency loss to obtain a final prediction result.

8. An electronic device, characterized in that, It includes a memory (100) and a processor (200), the memory (100) stores a computer program executable by the processor (200), and when the processor (200) executes the computer program, it implements the steps of the carbon dioxide column concentration prediction method based on the generative diffusion and atmospheric diffusion fusion model according to any one of claims 1-6.

Images