Carbon emission factor correction method and device based on dynamic energy component analysis

Abstract

The invention relates to the technical field of intelligent monitoring, in particular to a carbon emission factor correction method and device based on dynamic energy component analysis, and the method comprises the steps: obtaining fire coal calorific value data and power grid green power proportion data, and constructing a multi-dimensional feature tensor in combination with time, space and energy dimension information; a physical constraint change model is constructed according to the multi-dimensional feature tensor, the physical constraint change model comprises an encoder and a decoder, a theoretical carbon emission factor and a dual-drive loss value are calculated according to the encoder and the decoder, and the deviation degree between prediction and the theoretical carbon emission factor is obtained; the method comprises the steps of determining the number of sampling times of Monte Carlo Dropout according to the deviation degree, obtaining a prediction result, calculating a carbon emission factor uncertainty interval, determining an attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval, tracing a key influence factor, correcting a model, and achieving accurate carbon emission factor prediction. Therefore, the problems that in the prior art, carbon emission factors are static, hysteresis and low in interpretability are solved.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology, and particularly to a carbon emission factor correction method and device based on dynamic energy component analysis. Background Art

[0002] Currently, the mainstream practice of carbon emission accounting generally relies on the static emission factor method, which mainly estimates based on default values of international institutions or specific regions, such as the IPCC standard. However, due to the dynamic fluctuation characteristics of energy components, this traditional method faces significant error challenges: the calorific value of coal can fluctuate by up to ±20% due to different origins and batches, and the penetration rate of green electricity in the power grid can fluctuate by more than 30% at the minute level with the real-time changes in wind and solar power generation. However, the traditional accounting method still uses annual average data, resulting in an accounting error generally exceeding 15%.

[0003] To address this issue, existing improvement schemes mainly focus on regression analysis of historical data or the use of simple time series models. However, these schemes still have the following three obvious deficiencies: First, these models fail to effectively couple with the physical combustion equation, resulting in a large deviation between the prediction results and the actual situation in the case of small samples; Second, the timeliness of the data is significantly insufficient, and the current widely used day-ahead prediction mode far fails to meet the urgent need for minute-level data updates in the carbon trading market; Finally, the problem of data islands across regions is very prominent, and the generalization performance of models trained in a single region is greatly reduced when facing heterogeneous energy structures. For example, the actual measurement data of a thermal power plant shows that the result of carbon emission accounting using the static factor method deviates from the measured value by 18.3%, and even for a dynamic model based on LSTM, due to ignoring the real-time changes in boiler efficiency, there is still an error of 9.6% in its accounting result. In addition, the existing technology also lacks an interpretable analysis of the changes in emission factors, making it difficult to meet the strict requirements of carbon verification audits. Summary of the Invention

[0004] This application provides a carbon emission factor correction method and device based on dynamic energy component analysis to solve the problems of static carbon emission factors, lag, and low interpretability in the prior art.

[0005] The first aspect embodiment of this application provides a carbon emission factor correction method based on dynamic energy component analysis, including the following steps: obtaining coal calorific value data and grid green electricity proportion data; constructing a multi-dimensional feature tensor according to the coal calorific value data, the grid green electricity proportion data, and combining time, space, and energy dimension information, where the multi-dimensional feature tensor includes decaying sine time encoding, geographical hash space encoding, and hierarchical energy type encoding; constructing a physical constraint change model according to the multi-dimensional feature tensor, where the physical constraint change model includes an encoder and a decoder, calculating a theoretical carbon emission factor and a dual-drive loss value according to the encoder and the decoder, and obtaining the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor according to the theoretical carbon emission factor and the dual-drive loss value; determining the sampling times of Monte Carlo Dropout according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor, obtaining a prediction result according to the sampling times and calculating a carbon emission factor uncertainty interval, determining an attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval, tracing the key influencing factors according to the attention weight matrix, and correcting the physical constraint change model according to the key influencing factors to achieve accurate carbon emission factor prediction.

[0006] Optionally, the formula for the decaying sine time encoding is:

[0007]

[0008] where t is the time step number; i is the dimension index; d is the embedding dimension; λ is the decay coefficient;

[0009] The formula for the theoretical carbon emission factor is:

[0010]

[0011] where C ad is the carbon content on the as-received basis of the fuel; A ad is the ash content on the as-received basis of the fuel; Q net,ar is the lower calorific value on the as-received basis of the fuel;

[0012] The formula for the dual-drive loss value is:

[0013] L = β · MSE(EF pred , EF real ) + (1 - β) · KL(EF pred || EF phy )

[0014] where α is the balance coefficient; EF pred represents the emission factor finally predicted by the model; EF real represents the measured emission factor; EF phyrepresents the theoretical emission factor; KL is the divergence function.

[0015] Optionally, obtaining a prediction result according to the number of samplings and calculating the uncertainty interval of the carbon emission factor, including:

[0016] The formula for calculating the dynamic carbon emission factor is:

[0017]

[0018] where N represents the number of samplings; EF n represents the model output value of the nth sampling; represents the predicted value of the final emission factor.

[0019] Optionally, tracing the key influencing factors according to the attention weight matrix, including: obtaining the attention weight matrix of the encoder; calculating the feature contribution degree according to the attention weight matrix, where the formula for the feature contribution degree is:

[0020]

[0021] where H is the number of attention heads, representing the number of different feature interaction modes for parallel learning of the model; T is the time step, representing the time window length of the input sequence; Q t represents the query vector at the current time step t; represents the key vector of the i-th feature in the h-th attention head; represents the attention weight of the h-th head to the feature i at the time step t; obtaining the key influencing factors according to the feature contribution degree.

[0022] Optionally, before constructing a multi-dimensional feature tensor by combining coal calorific value data and grid green electricity proportion data with time, space, and energy dimension information, including: pre-training across regions, where the pre-training includes obtaining multi-source data, training under a multi-task learning framework based on the multi-source data, predicting the dynamic carbon emission factor of the next time period, and initializing the model by migrating the underlying parameters from a general time series prediction model; dynamically adjusting the model according to the model initialization parameters through the cosine annealing strategy and the elastic weight consolidation algorithm, and implementing an automatic rollback mechanism to reach the target model.

[0023] Optionally, the multi-source data includes the historical dispatching data of the power grid and the in-furnace coal industrial analysis data of coal-fired power plants. Among them, the in-furnace coal industrial analysis data of coal-fired power plants includes elemental carbon content data and calorific value data. The formula for the elastic weight consolidation algorithm is:

[0024]

[0025] where L newis the prediction error of the current fine-tuning data; λ is the weight controlling the importance of the old task; F i is the Fisher information matrix; θ i represents the current parameter value of the model; θ old,i is the parameter value after convergence in the pre-training stage.

[0026] Optionally, before constructing the multi-dimensional feature tensor by combining the coal calorific value data and the grid green electricity proportion data with time, space, and energy dimension information, it further includes: constructing a federated architecture; performing privacy protection on the data based on the federated architecture, where gradient encryption and differential privacy are used to ensure the security of gradient transmission, and the Intel SGX trusted execution environment is used to ensure the isolation and security of data storage; using the data sharing and privacy protection provided by the federated architecture, dynamically weighted averaging adjusts the weights according to data freshness and data volume. When anomalies are detected, inconsistent gradients are eliminated, and the global model is periodically aggregated and stored on the blockchain as evidence.

[0027] Optionally, the weights are adjusted according to data freshness and data volume, and the formula is:

[0028]

[0029] where n i is the data volume, and t i is the difference between the latest timestamp of the data and the current time.

[0030] The second aspect of the embodiments of the present application provides a carbon emission factor correction device based on dynamic energy component analysis, including: an acquisition module for acquiring coal calorific value data and grid green electricity proportion data; a construction module for constructing a multi-dimensional feature tensor according to the coal calorific value data and the grid green electricity proportion data by combining time, space, and energy dimension information, where the multi-dimensional feature tensor includes decaying sinusoidal time encoding, geohash space encoding, and hierarchical energy type encoding; a calculation module for constructing a physical constraint change model according to the multi-dimensional feature tensor, where the physical constraint change model includes an encoder and a decoder, calculating the theoretical carbon emission factor and the dual-drive loss value according to the encoder and the decoder, and obtaining the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor according to the theoretical carbon emission factor and the dual-drive loss value; a correction module for determining the sampling times of Monte Carlo Dropout according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor, obtaining the prediction result according to the sampling times and calculating the carbon emission factor uncertainty interval, determining the attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval, tracing the key influencing factors according to the attention weight matrix, and correcting the physical constraint change model according to the key influencing factors to finally achieve accurate carbon emission factor prediction.

[0031] A third aspect embodiment of the present application provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the program to perform the carbon emission factor correction method based on dynamic energy component analysis as described in the above embodiments.

[0032] Therefore, the present application has at least the following beneficial effects:

[0033] In the embodiments of the present application, multi-dimensional data such as coal calorific value and green electricity proportion are collected in real time through a high-frequency Internet of Things sensing network, a feature engine integrating spatio-temporal position encoding is constructed, and a Transformer model enhanced by physical constraints is designed to achieve second-level dynamic calibration of emission factors. The specific objectives include:

[0034] 1) Break through the linear assumption of the static factor method, establish a non-linear dynamic mapping relationship between calorific value and emission factors, and further reduce the accounting error;

[0035] 2) Integrate multi-region data through a federated learning framework to improve the generalization ability of the model while protecting data privacy;

[0036] 3) Use the self-attention mechanism to analyze key influencing variables and generate a traceability report that meets audit requirements;

[0037] 4) Support lightweight deployment at the edge to meet the real-time computing requirements of scenarios such as power plants and power grids;

[0038] The present application can be applied to carbon quota trading and enterprise process optimization. For example, dynamic carbon emission factors are used as the calculation benchmark for real-time carbon prices, or to guide coal-fired units to adjust the fuel blending ratio to reduce carbon costs, which is expected to promote the transformation of carbon accounting from "post-event statistics" to "process precise control". Thus, the problems of static, lagging, and low interpretability of carbon emission factors in the prior art are solved.

[0039] Additional aspects and advantages of the present application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The above and / or additional aspects and advantages of the present application will become obvious and easy to understand from the following description of the embodiments in conjunction with the drawings, where:

[0041] Figure 1 is a flowchart of a carbon emission factor correction method based on dynamic energy component analysis provided according to an embodiment of the present application;

[0042] Figure 2 is a flowchart of a carbon emission factor correction method based on dynamic energy component analysis provided according to an embodiment of the present application;

[0043] Figure 3 It is an example diagram of a carbon emission factor correction system based on dynamic energy component analysis provided according to an embodiment of the present application;

[0044] Figure 4 It is a block example diagram of a carbon emission factor correction device based on dynamic energy component analysis provided according to an embodiment of the present application;

[0045] Figure 5 It is a schematic structural diagram of an electronic device provided according to an embodiment of the present application. Detailed implementation manners

[0046] The embodiments of the present application will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present application, and should not be construed as a limitation to the present application.

[0047] The carbon emission factor correction method and device based on dynamic energy component analysis according to the embodiments of the present application will be described below with reference to the accompanying drawings. Aiming at the problem of static carbon emission factors mentioned in the above background art, the present application provides a carbon emission factor correction method based on dynamic energy component analysis. In this method, by integrating multi-dimensional data such as coal calorific value, the proportion of green electricity in the power grid, and time, space, and energy dimensions, a comprehensive feature tensor is constructed, and a specific coding technology is introduced to enhance the model's sensitivity to changes in time periodicity, geographical location, and energy type. Using the physical constraint change model established by the encoder and decoder, combined with the theoretical carbon emission factor and the dual-drive loss value, high-precision dynamic prediction of the carbon emission factor is achieved. In addition, the Monte Carlo Dropout technology is used to quantify the prediction uncertainty, and the key influencing factors are traced through the attention mechanism, further improving the accuracy and generalization ability of the model. Thus, the problems of static carbon emission factors, lag, and low interpretability in the prior art are solved.

[0048] The disaster warning control method, device, and equipment based on the change of meteorological element information according to the embodiments of the present application will be described below with reference to the accompanying drawings.

[0049] Specifically, Figure 1 It is a schematic flow diagram of a carbon emission factor correction method based on dynamic energy component analysis provided by an embodiment of the present application.

[0050] As Figure 1 shown, the carbon emission factor correction method based on dynamic energy component analysis includes the following steps:

[0051] In step S101, obtain the coal calorific value data and the proportion data of green electricity in the power grid.

[0052] Among them, the coal calorific value refers to the heat released when a unit mass of coal is completely burned; the proportion data of green electricity in the power grid refers to the proportion of electricity from renewable energy in the power grid.

[0053] It can be understood that by obtaining and analyzing the coal calorific value data and the proportion data of green electricity in the power grid in the embodiments of the present application, problems and potential risks in the energy system can be discovered in a timely manner, the development trend of renewable energy in the power grid can be understood, and data support can be provided for energy transformation.

[0054] It should be noted that the method for obtaining the coal calorific value data is to use the fusion detection of a laser-induced breakdown spectroscopy sensor and a near-infrared spectroscopy sensor, which is deployed at the coal conveyor belt or the inlet of the coal mill, with a sampling frequency ≥ 1 time / minute and a calorific value detection error ≤ ±0.5 MJ / kg; the proportion data of green electricity in the power grid is obtained by collecting distribution node data through an intelligent electricity meter and accessing the API of the power dispatching system to obtain the real-time clean energy output ratio of cross-regional power transmission, with a data update cycle ≤ 15 seconds; sliding window normalization (window length 30 minutes, step size 5 minutes) is performed at the edge computing node, and the isolation forest algorithm is used to remove outliers in real time (confidence level > 99%).

[0055] In step S102, according to the coal calorific value data and the proportion data of green electricity in the power grid, a multi-dimensional feature tensor is constructed by combining time, space, and energy dimension information. Among them, the multi-dimensional feature tensor includes a decaying sine time encoding, a geohash space encoding, and a hierarchical energy type encoding.

[0056] Among them, the multi-dimensional feature tensor can be a data structure for storing feature data with multiple dimensions; the decaying sine time encoding can capture the periodicity and non-linearity characteristics of time series data by converting time information into a sine waveform and performing decay processing; the geohash space encoding can be a method of converting geographical coordinates into a fixed-length string representation for indexing and retrieving spatial data; the hierarchical energy type encoding can be a method of hierarchically encoding different types of energy to distinguish the differences and hierarchical relationships between different energy types.

[0057] It is understandable that in the embodiments of the present application, by integrating the coal combustion calorific value data with the green electricity proportion data of the power grid and introducing information in three dimensions of time, space, and energy, not only the accuracy and comprehensiveness of data analysis are improved, but also a rich perspective is provided for subsequent analysis and prediction. Through technologies such as decaying sine time encoding, geohash space encoding, and hierarchical energy type encoding, the multi-dimensional feature tensor can capture the periodic and non-linear characteristics of time series, achieve fast matching and query of spatial data, and fine-grained differentiation of energy types. The application of these encoding technologies further enhances the readability and analysis efficiency of the data. In addition, the multi-dimensional feature tensor also provides strong support for accurate prediction and in-depth analysis. By constructing a prediction model based on this tensor, accurate prediction of key indicators such as energy use and carbon emissions can be achieved.

[0058] In the embodiments of the present application, the formula for decaying sine time encoding is:

[0059]

[0060] where t is the time step number; i is the dimension index; d is the embedding dimension; and λ is the decay coefficient.

[0061] Specifically, for the time dimension: decaying sine time encoding is adopted, and the formula is:

[0062]

[0063] where t is a scalar, the time step number, representing the position of the current data point in the time series; i is an integer, the dimension index, controlling the frequency distribution of the sine function, cycling from 0 to d / 2 - 1; d is a fixed constant, the embedding dimension, representing the total length of the feature vector (which needs to be an even number); λ is a hyperparameter, the decay coefficient, controlling the exponential decay rate of the historical data weight, and the larger the value, the more important the recent data. It strengthens the weight of recent data and suppresses the interference of historical noise.

[0064] For the space dimension: a unique geohash space encoding (Geohash precision level 8) is assigned to the generator set / grid node and embedded into the feature vector.

[0065] For energy type encoding: a hierarchical label system (primary classification: fossil energy / renewable energy; secondary classification:

[0066] coal / gas / wind power / solar power) is used, and two-channel embedding vector fusion is adopted.

[0067] In the embodiments of the present application, before constructing a multi-dimensional feature tensor based on coal combustion calorific value data and the proportion of green electricity in the power grid, combined with time, space, and energy dimension information, it includes: performing pre-training across regions, where the pre-training includes obtaining multi-source data, training under a multi-task learning framework based on the multi-source data, predicting the dynamic carbon emission factor in the next time period, and initializing the model by migrating underlying parameters from a general time series prediction model; dynamically adjusting the model according to the model initialization parameters through a cosine annealing strategy and an elastic weight consolidation algorithm, and implementing an automatic rollback mechanism to reach the target model.

[0068] Among them, the multi-source data may include the historical dispatching data of the power grid and the industrial analysis data of the coal entering the furnace of coal-fired power plants. Among them, the industrial analysis data of the coal entering the furnace of coal-fired power plants may include elemental carbon content data and calorific value data.

[0069] It can be understood that the embodiments of the present application integrate diverse data from different regions covering multiple aspects such as energy use, carbon emissions, and economic activities through a cross-regional pre-training strategy. On this basis, a multi-task learning framework is adopted, enabling the model to simultaneously process multiple related tasks, such as predicting carbon emission factors and energy demands, thereby learning deeper features. At the same time, transfer learning is used to migrate underlying parameters from a general time series prediction model to initialize the new model, accelerating the training process and improving accuracy. During the training process, a cosine annealing strategy and an elastic weight consolidation algorithm are implemented for dynamic adjustment, and an automatic rollback mechanism is introduced to ensure the stability and reliability of model training, significantly improving the prediction accuracy of the model for the dynamic carbon emission factor in the next time period, enhancing the generalization ability of the model, and improving the training efficiency.

[0070] Specifically, a provincial power grid is simulated, covering 3 thermal power clusters, 5 wind farms, and 2 photovoltaic bases, with a total installed capacity of 18 GW.

[0071] In the embodiments of the present application, the formula of the elastic weight consolidation algorithm is:

[0072]

[0073] Among them, Lnew is the prediction error of the current fine-tuning data; λ is the weight controlling the importance of the old task; Fi is the Fisher information matrix; θ i represents the current parameter value of the model; θ old,i is the parameter value after convergence in the pre-training stage.

[0074] Specifically, the pre-training task is carried out under a multi-task learning framework, which includes the following key components:

[0075] Main task: The main goal is to predict the dynamic carbon emission factor for the next time period. To achieve this goal, the model uses the mean absolute error (MAE) loss as the optimization objective to minimize the error between the predicted carbon emission factor and the actual value.

[0076] Auxiliary task: To enhance the robustness and generalization ability of the model, an auxiliary task is introduced, which is to reconstruct the missing sensor data. This task uses a generative adversarial network (GAN) to simulate the scenario of missing data and trains the model to accurately reconstruct the missing data. This not only helps the model maintain performance when dealing with incomplete data but also promotes the model to learn deeper data features.

[0077] Model initialization: To improve the training efficiency and model accuracy, the method of transfer learning is used to transfer the underlying parameters from a general time series prediction model to initialize the new model. This approach makes full use of the prior knowledge of the existing model and provides a solid foundation for the model to learn on the new task.

[0078] Online fine-tuning mechanism:

[0079] Dynamic learning rate adjustment: To balance the convergence speed and stability of the model, the cosine annealing strategy is used to adjust the learning rate. The initial value of the learning rate is set to 1e-4, and a period of 24 hours is set, that is, the learning rate will be adjusted according to the change law of the cosine function every 24 hours. This strategy helps the model converge quickly in the initial stage of training and fine-tune the parameters in the later stage to achieve better performance.

[0080] Catastrophic forgetting suppression: To avoid the model forgetting the learned knowledge during training, the elastic weight consolidation algorithm is introduced. The formula of the elastic weight consolidation algorithm is:

[0081]

[0082] where L new is the prediction error of the current fine-tuning data; λ is the weight controlling the importance of the old task; F i is the Fisher information matrix; θ i represents the current parameter value of the model; θ old,i is the parameter value after convergence in the pre-training stage;

[0083] Model version management: When the validation set error rises continuously for 3 times by >5%, it will automatically roll back to the historical optimal version.

[0084] In the embodiments of the present application, before constructing a multi-dimensional feature tensor based on coal combustion calorific value data and the proportion of green electricity in the power grid, combined with time, space, and energy dimension information, it further includes: constructing a federated architecture; performing privacy protection on the data based on the federated architecture, where gradient encryption and differential privacy are used to ensure the security of gradient transmission, and the Intel SGX trusted execution environment is used to ensure the isolation and security of data storage; using the data sharing and privacy protection provided by the federated architecture, dynamically weighted averaging adjusts weights according to data freshness and data volume, and when anomalies are detected, inconsistent gradients are eliminated, and the global model is periodically aggregated and stored on the blockchain.

[0085] It can be understood that in the embodiments of the present application, by constructing a federated architecture, data holders are allowed to jointly perform model training and data analysis without sharing the original data. At the same time, a number of privacy protection measures are taken, including using gradient encryption and differential privacy technologies to ensure the security of gradient transmission, and introducing the Intel SGX trusted execution environment to ensure the isolation and security of data storage. In addition, by using the dynamic weighted averaging method to adjust weights, optimizing model training according to data freshness and data volume, and automatically eliminating abnormal or inconsistent gradients, not only the accuracy and generalization ability of the model are improved, but also the security of the data is significantly enhanced. The adoption of the federated architecture breaks data silos, promotes the efficient circulation and sharing of data, and ensures the transparency and traceability of the model by storing the global model on the blockchain.

[0086] In the embodiments of the present application, the weights are adjusted according to data freshness and data volume, and the weight formula is:

[0087]

[0088] where n i is the data volume, and t i is the difference between the data's latest timestamp and the current time.

[0089] Specifically, the federated architecture design:

[0090] Participant roles: Data holders: power generation enterprises (coal quality data), provincial power grids (green electricity data), industrial parks (energy consumption data); coordination nodes: deployed on a blockchain platform (such as Hyperledger Fabric), responsible for gradient aggregation and model distribution; communication protocol: gRPC long connection + Protobuf data serialization, supporting tens of thousands of concurrent updates per second.

[0091] Privacy protection mechanism:

[0092] Gradient Encryption: The Paillier homomorphic encryption algorithm is adopted. After locally calculating the gradients, they are encrypted and transmitted, and the public key is distributed by the coordination node; Differential Privacy: Gaussian noise is added during gradient update (noise standard deviation σ = 0.01, privacy budget ε = 1.0); Data Isolation: The local data of each participating party is stored in an encrypted container (Intel SGX trusted execution environment);

[0093] Federated Aggregation Strategy:

[0094] Dynamic Weighted Average: Weights are assigned according to data freshness (time decay factor) and data volume. The formula is:

[0095]

[0096] where n i is the data volume, and t i is the difference between the latest timestamp of the data and the current time;

[0097] Anomaly Detection: Calculate the cosine similarity between the gradients of the participating parties and the global gradient, and remove the abnormal nodes with a similarity < 0.7;

[0098] Model Version Control: Aggregate the global model every 6 hours and generate an immutable version hash for storage on the blockchain.

[0099] In step S103, a physical constraint change model is constructed based on the multi-dimensional feature tensor. Among them, the physical constraint change model includes an encoder and a decoder. The theoretical carbon emission factor and the dual-drive loss value are calculated according to the encoder and the decoder, and the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor is obtained based on the theoretical carbon emission factor and the dual-drive loss value.

[0100] Among them, the theoretical carbon emission factor is an estimated value of the carbon emission calculated according to physical principles and known conditions; the dual-drive loss value may refer to a comprehensive loss value that simultaneously considers the prediction error of the carbon emission and the model complexity, etc.; the encoder is set to a 12-layer multi-head self-attention mechanism (number of heads 8), and the input dimension is 512; the decoder is set to a 4-layer cross-attention mechanism, integrating the physical equation constraint module of the fuel carbon oxidation rate.

[0101] It is understandable that in the embodiments of the present application, by introducing an encoder and a decoder structure, a physical constraint change model capable of processing multi-dimensional feature tensors is constructed. This model can efficiently extract key information from complex factors such as energy type, production process, and equipment efficiency, and calculate the theoretical carbon emission factor in the latent space. By designing a dual-driven loss function, the model not only focuses on prediction accuracy but also ensures that the predicted value conforms to physical constraints, thereby improving the accuracy and physical rationality of the model. By comparing the predicted carbon emission factor with the theoretical carbon emission factor, the deviation degree of the model prediction can be quantified, providing a basis for model improvement, not only improving the prediction accuracy of the carbon emission factor but also enhancing the interpretability and robustness of the model.

[0102] In the embodiments of the present application, the formula for the theoretical carbon emission factor is:

[0103]

[0104] Where C ad is the carbon content on the as-received basis of the fuel; A ad is the ash content on the as-received basis of the fuel; Q net,ar is the lower calorific value on the as-received basis of the fuel;

[0105] The formula for the dual-driven loss value is:

[0106] L = β · MSE(EF pred , EF real ) + (1 - β) · KL(EF pred || EF phy )

[0107] Where α is the balance coefficient; EF pred represents the emission factor finally predicted by the model; EF real represents the measured emission factor; EF phy represents the theoretical emission factor; KL is the divergence function.

[0108] In step S104, the sampling times of Monte Carlo Dropout are determined according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor. The prediction result is obtained according to the sampling times, and the uncertainty interval of the carbon emission factor is calculated. The attention weight matrix is determined according to the region or characteristics of the carbon emission factor uncertainty interval. The key influencing factors are traced according to the attention weight matrix, and the physical constraint change model is corrected according to the key influencing factors to achieve accurate prediction of the carbon emission factor.

[0109] Among them, Monte Carlo Dropout is a method for estimating the prediction uncertainty of a model by randomly discarding some neurons in a neural network. The attention weight matrix can represent the degree of importance of different input features during the prediction process. The key influencing factors can be key factors or features that have a significant impact on the prediction results.

[0110] It can be understood that the embodiments of this application adopt the Monte Carlo Dropout sampling technique, flexibly adjust the sampling times according to the deviation degree between the predicted carbon emission factor and the theoretical value, effectively reduce the prediction random error, and improve the prediction accuracy. At the same time, by calculating the uncertainty interval of the carbon emission factor, a quantitative evaluation is provided for the prediction results, enhancing the scientificity of decision-making. Further, an attention weight matrix is constructed using the characteristics of the uncertainty interval, successfully tracing the key factors affecting carbon emission prediction, providing a strong basis for formulating emission reduction strategies. Finally, the physical constraint change model is optimized based on these key influencing factors, significantly improving the prediction ability and adaptability of the model, enabling it to more accurately handle various complex scenarios.

[0111] In the embodiments of this application, obtaining the prediction result according to the sampling times and calculating the uncertainty interval of the carbon emission factor includes: The formula for calculating the dynamic carbon emission factor is:

[0112]

[0113] where N represents the sampling times; EF n represents the model output value of the nth sampling; represents the final emission factor prediction value.

[0114] In the embodiments of this application, tracing the key influencing factors according to the attention weight matrix includes: Obtaining the attention weight matrix of the encoder; Calculating the feature contribution degree according to the attention weight matrix, where the formula for the feature contribution degree is:

[0115]

[0116] where H is the number of attention heads, representing the number of different feature interaction patterns for parallel learning of the model; T is the time step, representing the time window length of the input sequence; Q t represents the query vector at the current time step t; represents the key vector of the i-th feature in the h-th attention head; represents the attention weight of the h-th head to the i-th feature at the time step t;

[0117] Obtaining the key influencing factors according to the feature contribution degree.

[0118] It is understandable that the embodiments of the present application intuitively display the contribution degree of features in prediction through the attention weight matrix, quickly identify key influencing factors, significantly improve the prediction accuracy and reliability, which is particularly crucial in scenarios with high-precision requirements such as carbon emission prediction. At the same time, it promotes the optimization of the model structure, improves the operation efficiency and calculation speed. In addition, it enhances the interpretability of the prediction results and enables a clear understanding of the key factors affecting the prediction.

[0119] According to the carbon emission factor correction method based on dynamic energy component analysis proposed by the embodiments of the present application, by integrating multi-dimensional data such as coal combustion calorific value, the proportion of green electricity in the power grid, and time, space, and energy dimensions, a comprehensive feature tensor is constructed, and a specific coding technology is introduced to enhance the model's sensitivity to changes in time periodicity, geographical location, and energy type. Using the physical constraint change model established by the encoder and decoder, combined with the theoretical carbon emission factor and the dual-drive loss value, a high-precision dynamic prediction of the carbon emission factor is achieved. In addition, the Monte Carlo Dropout technology is used to quantify the prediction uncertainty, and the attention mechanism is used to trace the key influencing factors, further improving the accuracy and generalization ability of the model. Thus, the problems of static carbon emission factors, lag, and low interpretability in the prior art are solved.

[0120] The following will elaborate on the carbon emission factor correction method based on dynamic energy component analysis through a specific embodiment, as Figure 2 shown, including the following steps:

[0121] 1. Real-time data collection by Internet of Things sensors:

[0122] 1.1 Detection of coal combustion calorific value: Fusion detection using a laser-induced breakdown spectroscopy (LIBS) sensor and a near-infrared spectroscopy (NIR) sensor, deployed at the coal conveyor belt or the inlet of the coal mill, with a sampling frequency ≥ 1 time / minute and a calorific value detection error

[0123] ≤ ±0.5 MJ / kg;

[0124] 1.2 Tracking of the proportion of green electricity in the power grid: Collect distribution node data through smart electricity meters and access the power grid dispatching system API to obtain the real-time clean energy output proportion of cross-regional power transmission, with a data update cycle ≤ 15 seconds;

[0125] 1.3 Data preprocessing: Perform sliding window normalization (window length 30 minutes, step size 5 minutes) at the edge computing node, and use the isolated forest algorithm to remove outliers in real time (confidence level > 99%).

[0126] 2. Construction of multi-dimensional feature tensor:

[0127] 2.1 Spatiotemporal position encoding:

[0128] Time dimension: Use decaying sine positional encoding, with the formula:

[0129]

[0130] where t is a scalar, the time step number, representing the position of the current data point in the time series; i is an integer, the dimension index, controlling the frequency distribution of the sine function, cycling from 0 to d / 2 - 1; d is a fixed constant, the embedding dimension, representing the total length of the feature vector (which needs to be even); λ is a hyperparameter, the decay coefficient, controlling the exponential decay rate of the historical data weight. The larger the value, the more important the recent data. Strengthen the weight of recent data and suppress historical noise interference.

[0131] Spatial dimension: Assign a unique Geohash encoding (Geohash precision level 8) to the generator set / grid node and embed it into the feature vector.

[0132] 2.2 Energy type encoding: Use a hierarchical label system (primary classification: fossil energy / renewable energy; secondary classification: coal / gas / wind power / solar power), and adopt dual-channel embedding vector fusion.

[0133] 3. Physical constraint Transform modeling:

[0134] 3.1 Model structure:

[0135] Encoder: 12-layer multi-head self-attention mechanism (number of heads 8), input dimension 512;

[0136] Decoder: 4-layer cross-attention mechanism, integrating the physical equation constraint module of the fuel carbon oxidation rate. The theoretical carbon emission factor calculation formula is:

[0137]

[0138] where C ad is the carbon content on the as-received basis of the fuel (dry basis), detected by an elemental analyzer or a LIBS sensor; A ad is the ash content on the as-received basis of the fuel, reflecting the proportion of the non-combustible part in combustion; Q net,ar is the lower calorific value on the as-received basis of the fuel, representing the actual available energy value; 44 / 12 represents the molar mass ratio of CO2 to carbon, converting the carbon content to the CO2 emission coefficient.

[0139] Loss function: Define the physical-data dual-driven loss:

[0140] L = β·MSE(EF pred , EF real ) + (1 - β)·KL(EF pred ||EF phy )

[0141] Among them, α is a hyperparameter, a balance coefficient, which adjusts the weights of the data-driven loss and the physical consistency loss; EF pred is a scalar, representing the emission factor finally predicted by the model; EF real is a scalar, representing the measured emission factor, which is obtained through the continuous emission monitoring system (CEMS); EF phy is a scalar, representing the theoretical emission factor; KL is a divergence function, the Kullback-Leibler divergence, which measures the deviation between the predicted value distribution and the theoretical value.

[0142] 3.2 Online inference: The model calculation is triggered every 5 minutes, accelerated by the TensorRT engine, and the inference latency < 50ms.

[0143] 4. Trusted output and traceability:

[0144] 4.1 Uncertainty assessment: The Monte Carlo Dropout method is used to quantify the uncertainty of the carbon emission factor. The calculation formula for the dynamic carbon emission factor is:

[0145]

[0146] Among them, N represents the number of samplings, and the distribution of predicted values is calculated through multiple forward propagations; EF n represents the model output value of the nth sampling (keep Dropout activated during inference to simulate randomness); represents the average value of all sampling results, representing the final predicted value of the emission factor.

[0147] 4.2 Variable traceability analysis: Extract the attention weight matrix of the third layer of the encoder and calculate the feature contribution degree:

[0148]

[0149] Among them, H is the number of attention heads, representing the number of different feature interaction modes for parallel learning of the model; T is the time step, representing the time window length of the input sequence (such as data in the past 24 hours); Q t represents the query vector (Query) at the current time step t, which is obtained by linearly transforming the input features; represents the key vector (Key) of the i-th feature in the h-th attention head, which is obtained by linearly transforming the input features; represents the attention weight of the h-th head on the i-th feature at time step t. From this, the TOP3 influencing factors and weight proportions can be output (e.g., the wind power output increases by 25% → contribution degree 38.2%).

[0150] In addition, the carbon emission factor correction system based on dynamic energy component analysis is described according to the regional power grid, such as Figure 3As shown in the figure, it includes:

[0151] Implementation scenario: Simulate a provincial power grid covering 3 thermal power clusters, 5 wind farms and 2 photovoltaic bases, with a total installed capacity of 18 GW.

[0152] 1. Hardware Deployment and Data Integration

[0153] Dynamic data sources:

[0154] Thermal power plant side: 12 LIBS coal quality monitors (deployed in the coal conveying system of each power plant), and the calorific value of the coal fed into the furnace is transmitted in real time (average value 20.3 MJ / kg, fluctuation ±15%);

[0155] New energy side: The wind power / solar power SCADA system provides minute-level output data (sampling frequency 5 seconds);

[0156] Inter-provincial power transmission: Obtain the green electricity proportion of the exported power through the API of the power trading platform (real-time fluctuation range 10%-65%).

[0157] Federated nodes:

[0158] Participants: 3 municipal power grid companies (managing thermal power, wind power, and photovoltaic clusters respectively);

[0159] Data isolation: The data of each company is locally stored in an encrypted database (AES-256), and only the gradient parameters are uploaded to the coordination node.

[0160] 2. Model Operation and Dynamic Factor Calculation

[0161] Spatio-temporal feature construction:

[0162] Time dimension: Taking 15 minutes as the time window, embedding decay position encoding (λ = 0.02);

[0163] Space dimension: Divided by geographical grid (0.1°×0.1°), encoding the energy structure characteristics of each region.

[0164] Federated training:

[0165] Model structure: 12-layer Transformer encoder (hidden layer 512), and the physical constraint decoder integrates the theoretical formula of grid carbon emissions:

[0166]

[0167] Among them, EF i is the carbon emission factor of each power source, and P i is the real-time output;

[0168] Training parameters: gradient aggregation every 2 hours, Paillier encryption key length 2048 bits, differential privacy noise σ = 0.005.

[0169] Dynamic Output:

[0170] Update frequency: Release the dynamic emission factor of the entire network every 5 minutes;

[0171] Source analysis: Identify key influencing factors through attention weights (such as "the sudden drop in photovoltaic output at 14:00 caused the factor to increase by 0.12kg / kWh").

[0172] 3. Data comparison and effect verification are shown in Table 1 below.

[0173] Table 1 Data comparison

[0174]

[0175] Improved accuracy:

[0176] The mean absolute error (MAE) of the dynamic factor is 0.008 kg / kWh, which is 96% lower than that of the static factor method (0.199) and 86% lower than that of the day-ahead forecast model (0.057).

[0177] At 15:30, the photovoltaic output dropped by 70% due to sudden cloudy weather. The real-time response error of the dynamic factor was only 1.1%, while the error of the day-ahead prediction model reached 13.2%.

[0178] The actual application effects are as follows:

[0179] Carbon trading optimization: Based on dynamic factors, the power grid company sells quotas at a low carbon factor of 0.45kg / kWh during the green electricity peak period (09:00-11:00), earning 127,000 yuan more per day than the static factor method;

[0180] Dispatch decision support: When the dynamic factor exceeds the threshold (0.6kg / kWh), the energy storage system is automatically called first to reduce the output of high-carbon power sources, reducing carbon emissions by 142 tons per day.

[0181] Cross-regional collaboration: After integrating data from neighboring provinces through federated learning, the model’s prediction error in border areas dropped from 7.3% to 2.8%.

[0182] In summary, through the Transformer model enhanced by IoT dynamic perception and physical constraints, minute-level high-precision correction of carbon emission factors is achieved, reducing the average error of traditional static methods from 15%-40% to within 3%. At the same time, it supports cross-regional secure collaborative computing and traceability of influencing factors. Compared with existing technologies, this method has made breakthroughs in data timeliness (second-level update), model interpretability (attention weight attribution), and multi-source data fusion (federated learning), providing real-time and reliable data support for carbon market trading and enterprise emission reduction process optimization;

[0183] With the help of high-precision IoT sensors, second-level acquisition of key parameters such as coal calorific value and green electricity proportion is achieved. Combining spatio-temporal attenuation coding technology, a multi-dimensional feature tensor is constructed, thus overcoming the data lag problem existing in traditional static factor methods; A Transformer architecture with enhanced physical constraints is carefully designed, and the fuel carbon oxidation rate equation is cleverly integrated in the decoding layer, perfectly integrating theoretical emission calculation and data-driven prediction, effectively solving the "black box" problem commonly existing in pure AI models; Using the Monte Carlo Dropout method, carbon emission factors with confidence intervals are generated, and through self-attention weight analysis technology, key influencing variables are accurately identified. This mechanism meets the strict requirements of carbon verification audits for interpretability; Based on the federated learning framework, joint training of multi-region models is achieved. By adopting Paillier homomorphic encryption and differential privacy protection technologies, data security is ensured, breaking the island phenomenon of energy data while meeting privacy compliance requirements.

[0184] Secondly, a carbon emission factor correction device based on dynamic energy component analysis proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0185] Figure 4 It is a block diagram of a carbon emission factor correction device based on dynamic energy component analysis according to an embodiment of the present application.

[0186] As Figure 4 shown, the carbon emission factor correction device 10 based on dynamic energy component analysis includes: an acquisition module 100, a construction module 200, a calculation module 300, and a correction module 400.

[0187] Among them, the acquisition module 100 is used to acquire coal calorific value data and grid green electricity proportion data; the construction module 200 is used to construct a multi-dimensional feature tensor according to the coal calorific value data and the grid green electricity proportion data, combined with time, space and energy dimension information, where the multi-dimensional feature tensor includes decaying sine time encoding, geographical hash space encoding and hierarchical energy type encoding; the calculation module 300 is used to construct a physical constraint change model according to the multi-dimensional feature tensor, where the physical constraint change model includes an encoder and a decoder, calculate the theoretical carbon emission factor and the dual-drive loss value according to the encoder and the decoder, and obtain the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor according to the theoretical carbon emission factor and the dual-drive loss value; the correction module 400 is used to determine the sampling times of Monte Carlo Dropout according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor, obtain the prediction result according to the sampling times and calculate the carbon emission factor uncertainty interval, determine the attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval, trace the key influencing factors according to the attention weight matrix, and correct the physical constraint change model according to the key influencing factors, so as to finally achieve accurate carbon emission factor prediction.

[0188] It should be noted that the foregoing explanation of the embodiment of the carbon emission factor correction method based on dynamic energy component analysis also applies to the carbon emission factor correction device based on dynamic energy component analysis of this embodiment, and will not be repeated here.

[0189] According to the carbon emission factor correction device based on dynamic energy component analysis proposed in the embodiment of the present application, by integrating multi-dimensional data of coal calorific value, grid green electricity proportion, and time, space and energy dimensions, a comprehensive feature tensor is constructed, and specific encoding techniques are introduced to enhance the model's sensitivity to time periodicity, geographical location and energy type changes. The physical constraint change model established by the encoder and the decoder, combined with the theoretical carbon emission factor and the dual-drive loss value, realizes high-precision dynamic prediction of the carbon emission factor. In addition, the Monte Carlo Dropout technology is used to quantify the prediction uncertainty, and the key influencing factors are traced through the attention mechanism, further improving the accuracy and generalization ability of the model. Thus, the problems of static carbon emission factors, lag and low interpretability in the prior art are solved.

[0190] Figure 5 The structural schematic diagram of the electronic device provided by the embodiment of the present application. The electronic device may include:

[0191] A memory 501, a processor 502, and a computer program stored on the memory 501 and executable on the processor 502.

[0192] When the processor 502 executes the program, it implements the carbon emission factor correction method based on dynamic energy component analysis provided in the foregoing embodiment.

[0193] Furthermore, the electronic device further includes:

[0194] A communication interface 503 for communication between the memory 501 and the processor 502.

[0195] A memory 501 for storing a computer program that can run on the processor 502.

[0196] The memory 501 may include a high-speed RAM memory and may also include a non-volatile memory, such as at least one disk memory.

[0197] If the memory 501, the processor 502, and the communication interface 503 are implemented independently, the communication interface 503, the memory 501, and the processor 502 can be interconnected via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 only a thick line is shown in the figure, but it does not mean that there is only one bus or one type of bus.

[0198] Optionally, in a specific implementation, if the memory 501, the processor 502, and the communication interface 503 are integrated on a single chip, the memory 501, the processor 502, and the communication interface 503 can communicate with each other through an internal interface.

[0199] The processor 502 may be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0200] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or N embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0201] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of this application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0202] Any process or method description shown in a flowchart or described in other ways herein can be understood to represent a module, segment, or part of code including one or N executable instructions for implementing a customized logical function or process, and the scope of the preferred embodiments of this application includes additional implementations, where the functions can be executed in a substantially simultaneous manner or in a reverse order according to the involved functions, rather than in the order shown or discussed, which should be understood by those skilled in the art to which the embodiments of this application belong.

[0203] It should be understood that each part of this application can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by a combination of any one or more of the following technologies well known in the art: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0204] Those of ordinary skill in the art of this technology can understand that all or part of the steps carried by the method for implementing the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

Claims

1. A method for correcting carbon emission factors based on dynamic energy component analysis, characterized in that, It includes the following steps: Obtain coal calorific value data and grid green electricity proportion data; Construct a multi-dimensional feature tensor according to the coal calorific value data and the grid green electricity proportion data, combined with time, space, and energy dimension information. Among them, the multi-dimensional feature tensor includes decaying sine time encoding, geographical hash space encoding, and hierarchical energy type encoding; Construct a physical constraint change model according to the multi-dimensional feature tensor. Among them, the physical constraint change model includes an encoder and a decoder. Calculate the theoretical carbon emission factor and the dual-drive loss value according to the encoder and the decoder, and obtain the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor according to the theoretical carbon emission factor and the dual-drive loss value; Determine the sampling times of Monte Carlo Dropout according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor. Obtain the prediction result according to the sampling times and calculate the carbon emission factor uncertainty interval. Determine the attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval. Trace the key influencing factors according to the attention weight matrix. Modify the physical constraint change model according to the key influencing factors to achieve accurate carbon emission factor prediction.

2. The carbon emission factor correction method based on dynamic energy component analysis according to claim 1, characterized in that, The formula for the decaying sine time encoding is: where t is the time step number; i is the dimension index; d is the embedding dimension; λ is the decay coefficient; The formula for the theoretical carbon emission factor is: Among them, C ad is the carbon content on as-received basis of the fuel; A ad is the ash content on as-received basis of the fuel; Q net,ar is the net calorific value on as-received basis of the fuel; The formula for the dual-drive loss value is: L = β·MSE(EF pred , EF real ) + (1 - β)·KL(EF pred || EF phy ) where α is the balance coefficient; EF pred represents the emission factor finally predicted by the model; EF real represents the measured emission factor; EF phy represents the theoretical emission factor; KL is the divergence function.

3. The carbon emission factor correction method based on dynamic energy component analysis according to claim 1, characterized in that, Obtain the prediction result according to the sampling times and calculate the carbon emission factor uncertainty interval, including: The formula for calculating the dynamic carbon emission factor is: Among them, N represents the number of sampling times; EF n represents the model output value of the nth sampling; represents the predicted value of the final emission factor.

4. The carbon emission factor correction method based on dynamic energy component analysis according to claim 1, wherein Trace the key influencing factors according to the attention weight matrix, including: Obtain the attention weight matrix of the encoder; Calculate the feature contribution degree according to the attention weight matrix. Among them, the formula for the feature contribution degree is: Among them, H is the number of attention heads, representing the number of different feature interaction patterns for parallel learning of the model; T is the time step, representing the time window length of the input sequence; Q t represents the query vector at the current time step t; represents the key vector of the i-th feature in the h-th attention head; represents the attention weight of the h-th head on the i-th feature at time step t; Obtain the key influencing factors according to the feature contribution degree.

5. The carbon emission factor correction method based on dynamic energy component analysis according to claim 1, wherein Before constructing a multi-dimensional feature tensor according to the coal calorific value data and the grid green electricity proportion data, combined with time, space, and energy dimension information, it includes: Perform pre-training across regions. Among them, the pre-training includes obtaining multi-source data, training under a multi-task learning framework based on the multi-source data, predicting the dynamic carbon emission factor in the next time period, and initializing the model by migrating the underlying parameters from a general time series prediction model; Dynamically adjust the model according to the model initialization parameters through the cosine annealing strategy and the elastic weight consolidation algorithm, and implement an automatic rollback mechanism to reach the target model.

6. The carbon emission factor correction method based on dynamic energy component analysis according to claim 5, wherein, The multi-source data includes the historical dispatching data of the power grid and the in-furnace coal industrial analysis data of coal-fired power plants. Among them, the in-furnace coal industrial analysis data of coal-fired power plants includes elemental carbon content data and calorific value data. The formula for the elastic weight consolidation algorithm is: Among them, L new is the prediction error of the current fine-tuning data; λ is the weight controlling the importance of the old task; F i is the Fisher information matrix; θ i represents the current parameter value of the model; θ old,i is the parameter value after convergence in the pre-training stage.

7. The carbon emission factor correction method based on dynamic energy component analysis according to claim 1, characterized in that, Before constructing a multi-dimensional feature tensor according to the coal calorific value data and the grid green electricity proportion data, combined with time, space, and energy dimension information, it also includes: Construct a federated architecture; Perform privacy protection on the data based on the federated architecture. Among them, use gradient encryption and differential privacy to ensure the security of gradient transmission, and use the Intel SGX trusted execution environment to ensure the isolation and security of data storage; Using the data sharing and privacy protection provided by the federal architecture, the dynamic weighted average adjusts the weights according to data freshness and data volume. When anomalies are detected, inconsistent gradients are eliminated, and the global model is periodically aggregated and stored on the blockchain.

8. The carbon emission factor correction method based on dynamic energy component analysis according to claim 7, wherein Adjust the weights according to data freshness and data volume, and the weight formula is: where n i is the data volume, and t i is the difference between the latest timestamp of the data and the current time.

9. A carbon emission factor correction device based on dynamic energy component analysis, characterized in that, Include: An acquisition module for acquiring coal calorific value data and the proportion of green electricity in the power grid; A construction module for constructing a multi-dimensional feature tensor according to the coal calorific value data and the proportion of green electricity in the power grid, in combination with time, space, and energy dimension information, where the multi-dimensional feature tensor includes decaying sinusoidal time encoding, geographical hash space encoding, and hierarchical energy type encoding; A calculation module for constructing a physical constraint change model according to the multi-dimensional feature tensor, where the physical constraint change model includes an encoder and a decoder, calculating a theoretical carbon emission factor and a dual-drive loss value according to the encoder and the decoder, and obtaining the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor according to the theoretical carbon emission factor and the dual-drive loss value; A correction module for determining the sampling times of Monte Carlo Dropout according to the deviation degree between the predicted carbon emission factor and the theoretical carbon emission factor, obtaining a prediction result according to the sampling times and calculating the uncertainty interval of the carbon emission factor, determining an attention weight matrix according to the region or characteristics of the carbon emission factor uncertainty interval, tracing the key influencing factors according to the attention weight matrix, and correcting the physical constraint change model according to the key influencing factors, finally realizing accurate prediction of the carbon emission factor.

10. An electronic device, characterized in that, Include: A memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement the carbon emission factor correction method based on dynamic energy component analysis according to any one of claims 1-8.

Images