Power distribution network carbon emission analysis method

Abstract

The invention provides a power distribution network carbon emission analysis method, which comprises the steps of constructing a user power consumption behavior portrait based on user electric quantity data, learning an influence weight of a user power consumption habit on carbon emission by utilizing an attention mechanism, mining carbon emission modes under different power consumption behaviors through a clustering algorithm, and forming a user-oriented carbon emission interpretation model; an Apriori association rule mining algorithm is adopted, a frequent pattern in carbon emission data is extracted from three dimensions of a time period, equipment and a user, a key factor combination which obviously influences the carbon emission is screened out, and a rule is embedded into a block chain through an intelligent contract to execute carbon emission traceability; the diversity of carbon emission data is expanded by utilizing the generative adversarial network, carbon emission conditions in different extreme scenes are simulated, visual tracing of carbon emission whole-process data is realized through a block chain browser, and a decision basis is provided for accurately checking carbon emission and formulating an emission reduction policy.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for analyzing carbon emissions in a distribution network. Background Art

[0002] In the process of carbon emission data collection and analysis, we are faced with the challenges of multi-dimensional data integration and precise modeling. First, how to effectively store heterogeneous carbon emission data collected by IoT sensors and smart meters in the blockchain network and ensure the authenticity and reliability of the data? Secondly, when processing high-dimensional carbon emission data, how to achieve effective dimensionality reduction while retaining key information? Furthermore, when establishing a carbon emission prediction model, how to simultaneously consider time series characteristics, equipment power consumption characteristics, and user behavior patterns to build a comprehensive multi-dimensional analysis model? Finally, in the process of model application, how to balance the accuracy and generalization ability of the model so that it can adapt to the carbon emission prediction needs of different regions and scenarios? These interrelated technical issues constitute a complex system engineering, which requires us to conduct in-depth research and innovation in multiple links such as data processing, model construction, and application promotion to achieve accurate carbon emission analysis and effective emission reduction strategy formulation. Summary of the invention

[0003] The present invention provides a distribution network carbon emission analysis method, which mainly includes:

[0004] According to the topology of the distribution network and the type of carbon emission sources, IoT sensors and smart meters are used to collect carbon emission values, timestamps, geographic locations and other multi-dimensional carbon emission raw data at different time and space granularities. Distributed ledger technology is used to store heterogeneous carbon emission data in a unified format in the blockchain network, and a multi-node consensus mechanism is used to verify the authenticity and reliability of carbon emission data.

[0005] The raw carbon emission data in the blockchain is cleaned, denoised and repaired, and a structured carbon emission data set is constructed by combining the carbon emission source type, intensity level and other attribute annotations. In view of the high-dimensional characteristics of carbon emission data, an autoencoder neural network is used to extract the low-dimensional semantic features of the data set to achieve dimensionality reduction and compression of carbon emission data;

[0006] According to the timestamp of carbon emission data, the sliding time window method is used to segment the continuous time series carbon emission data, and the long short-term memory network is used to model the changes in carbon emission patterns in different time periods, explore the nonlinear relationship between carbon emissions and time, and obtain the carbon emission rules in fine-grained time periods;

[0007] Based on the power load data collected by smart meters, a convolutional neural network is used to extract the power curve features of different types of power equipment. The softmax classifier is used to identify various types of power equipment. The carbon emission contribution of each device is calculated based on the rated power and usage time of the equipment, and a carbon emission decomposition model for the equipment is established.

[0008] Build a user electricity consumption behavior profile based on user electricity data, use the attention mechanism to learn the impact weight of user electricity consumption habits on carbon emissions, and use clustering algorithms to mine carbon emission patterns under different electricity consumption behaviors to form a user-oriented carbon emission explanation model;

[0009] The Apriori association rule mining algorithm is used to extract frequent patterns in carbon emission data from three dimensions: time period, equipment, and user, and to screen out key factor combinations that have a significant impact on carbon emissions. The rules are then embedded in the blockchain through smart contracts to perform carbon emission traceability.

[0010] Taking carbon emission influencing factors such as time period, equipment, user, and rules as input features, and carbon emissions as the target variable, the LightGBM regression algorithm is used to establish a carbon emissions prediction model. The pre-trained carbon emission model is applied to different regions and scenarios using the transfer learning method to improve the generalization performance of the model.

[0011] Adopting the model integration strategy, the time period model, equipment decomposition model, user interpretation model and rule mining model are weighted and integrated to form a fine-grained and multi-dimensional comprehensive carbon emission analysis model. Combined with the reinforcement learning algorithm, the carbon emission control strategy is optimized to achieve coordinated emission reduction in the source, grid, load and storage links.

[0012] Generative adversarial networks are used to expand the diversity of carbon emission data, simulate carbon emission conditions under different extreme scenarios, and realize visual traceability of the entire carbon emission process data through blockchain browsers, providing a decision-making basis for accurately calculating carbon emissions and formulating emission reduction policies.

[0013] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0014] The present invention discloses a refined carbon emission analysis method based on blockchain and artificial intelligence. The method collects multi-dimensional carbon emission data through Internet of Things devices and uses blockchain technology to ensure reliable data storage. In view of the high-dimensional characteristics of carbon emission data, an autoencoder neural network is used for dimensionality reduction, and the long short-term memory network is combined to model the temporal changes. The present invention innovatively deconstructs the sources of carbon emissions from two dimensions: equipment and users. It identifies the equipment type and calculates the carbon emission contribution through a convolutional neural network, and uses an attention mechanism to analyze the impact of user behavior on carbon emissions. Furthermore, the present invention integrates multi-dimensional models to construct a comprehensive carbon emission analysis model, and combines reinforcement learning to optimize emission reduction strategies. Finally, the present invention realizes the visual traceability of the entire process of carbon emission data, provides strong support for accurate accounting and the formulation of emission reduction policies, and effectively promotes the coordinated emission reduction of energy systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flow chart of a method for analyzing carbon emissions in a distribution network.

[0016] Figure 2 It is a schematic diagram of a distribution network carbon emission analysis method of the present invention.

[0017] Figure 3 This is another schematic diagram of a distribution network carbon emission analysis method according to the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 -3. A distribution network carbon emission analysis method in this embodiment may specifically include:

[0020] S101. According to the topological structure of the distribution network and the type of carbon emission sources, IoT sensors and smart meters are used to collect multi-dimensional carbon emission raw data such as carbon emission values, timestamps, geographic locations, etc. at different time and spatial granularities. The heterogeneous carbon emission data are stored in a unified format in the blockchain network through distributed ledger technology, and the authenticity and reliability of the carbon emission data are verified using a multi-node consensus mechanism.

[0021] IoT sensors and smart meters are deployed in all links of the distribution network. The placement and collection frequency of sensors and smart meters are determined according to the topological structure of the distribution network and the type of carbon emission sources. The carbon emission values, timestamps, geographic locations and other information of different time and spatial granularities of the distribution network are collected. The raw data of carbon emissions of the distribution network are obtained from sensors and smart meters. The obtained raw data of carbon emissions of the distribution network are cleaned and compressed on the edge computing device. The anomaly detection algorithm is used to identify and remove abnormal data in the raw data of carbon emissions of the distribution network to obtain a high-quality carbon emission data set of the distribution network. The processed carbon emission data set of the distribution network is uploaded to the blockchain network. The carbon emission data set of the distribution network is stored in a unified format using distributed ledger technology. The storage structure and access rules of the carbon emission data set of the distribution network are defined using smart contract technology to ensure the security and privacy of the carbon emission data set of the distribution network. A multi-node consensus mechanism is deployed in the blockchain network to verify the carbon emission data set of the distribution network on the chain. If the carbon emission data set of the distribution network passes the consensus verification, it will be permanently recorded in the blockchain ledger. If the carbon emission data set of the distribution network fails to pass the verification, it will be refused to be written into the ledger. For the distribution network carbon emission dataset recorded in the blockchain ledger, the temporal and spatial characteristics of the distribution network carbon emission are analyzed using time series decomposition and spatial autocorrelation analysis methods. The distribution network carbon emission dataset is obtained from the blockchain ledger. The SPSS tool is used to perform time series decomposition on the distribution network carbon emission dataset to obtain the high-incidence period of the distribution network carbon emission. The R language is used to perform spatial autocorrelation analysis on the distribution network carbon emission dataset to obtain the high-incidence area of ​​the distribution network carbon emission. According to the results of time series decomposition and spatial autocorrelation analysis, the key influencing factors of the distribution network carbon emission are determined. The distribution network carbon emission prediction model is established using the long short-term memory neural network (LSTM) algorithm. According to the distribution network carbon emission dataset, the historical carbon emission data is used as the training sample, and the determined key influencing factors of the distribution network carbon emission are used as the model input to train the multi-step prediction model. According to the trained distribution network carbon emission prediction model, the distribution network carbon emission forecast value in the future period is obtained. The distribution network carbon emission forecast value is compared with the actual distribution network carbon emission data obtained from sensors and smart meters. According to the comparison results of the predicted value of carbon emissions of the distribution network and the actual carbon emissions data of the distribution network, the accuracy of the distribution network carbon emission prediction model is evaluated. If the prediction error of the distribution network carbon emission prediction model exceeds the preset threshold, the newly collected distribution network carbon emission data is used to fine-tune and optimize the distribution network carbon emission prediction model. According to the predicted value of carbon emissions of the distribution network in the future, the genetic algorithm is used to formulate the energy-saving and emission reduction strategy of the distribution network. With the minimization of carbon emissions of the distribution network and the maximization of power supply reliability as the optimization goals, the optimal power grid dispatching scheme and equipment operating parameters are solved in the MATLAB optimization toolbox to obtain the optimal energy-saving and emission reduction strategy of the distribution network. The formulated optimal energy-saving and emission reduction strategy for the distribution network is sent to the distribution automation master station.The distribution network is optimized by controlling switches and adjusting equipment parameters through the distribution automation system. The implementation effect of the optimal energy-saving and emission reduction strategy of the distribution network is evaluated based on the real-time carbon emission data of the distribution network obtained from sensors and smart meters.

[0022] Specifically, carbon emission sensors and smart meters are deployed at key nodes such as substations, distribution rooms, and distribution boxes in the distribution network. The sensor sampling frequency is set to 15 minutes / time, and the smart meter data collection frequency is set to 1 hour / time. The collected data is cleaned through the edge computing gateway, and abnormal values ​​exceeding 20% ​​of the range are eliminated. The data is encrypted using the AES-256 algorithm, and the compression rate is set to 50%. The processed data is uploaded to the blockchain network, and the Ethereum platform is used to build a private chain. The smart contract is designed to define the data structure, including fields such as data ID, timestamp, carbon emissions, location, etc., and data access rights are set. The blockchain network adopts the PBFT consensus algorithm, sets the number of consensus nodes to 7, and the consensus time to 10 seconds. The on-chain data is verified, and the data that passes the verification is packaged into blocks. The data that fails to pass the verification triggers an early warning and notifies the data providing node. Using the FP-growth association rule mining algorithm, with the threshold of minimum support 2 and minimum confidence 7, the frequent patterns of carbon emission data were mined. It was found that the time of high carbon emission was 10:00-15:00, and the high-incidence area was the industrial park. The influencing factors included load level and equipment utilization. The LSTM algorithm was used to build a carbon emission prediction model. The historical carbon emission data was divided into training set and test set with a ratio of 8:2. The load, temperature, etc. were used as model inputs. The training model predicted the carbon emissions in the next 24 hours, and the test set prediction error MAPE was 5%. Based on the prediction results, the NSGA-II multi-objective optimization algorithm was used to minimize carbon emissions and maximize power supply reliability. The optimal dispatching strategy of the distribution network was solved, including DG output, capacitor switching, load control, etc. After simulation verification, the carbon emissions were reduced by 15% and the power supply reliability was increased by 5% after optimization. Finally, the optimization strategy was sent to the distribution automation master station, and the strategy was executed through the intelligent distribution terminal, and the carbon emission data was monitored in real time. If the execution effect did not meet expectations, the strategy adjustment was triggered to form a closed-loop optimization.

[0023] S102. Clean, denoise and repair the original carbon emission data in the blockchain, and construct a structured carbon emission data set by combining the carbon emission source type, intensity level and other attribute annotations. In view of the high-dimensional characteristics of carbon emission data, an autoencoder neural network is used to extract the low-dimensional semantic features of the data set to achieve dimensionality reduction and compression of carbon emission data.

[0024] The raw carbon emission data is obtained from the blockchain network. Regular expressions are used to clean the data, remove illegal characters and outliers, and the nearest neighbor interpolation method is used to repair missing values ​​to obtain the cleaned carbon emission data set. According to the type, industry, scale and other attributes of the carbon emission source, the cleaned carbon emission data set is annotated to construct a structured carbon emission data set. The structured carbon emission data set is divided into a training set and a test set with a ratio of 8:2. The categorical attributes in the training set are converted into numerical features using One-Hot encoding. In view of the high-dimensional characteristics of the carbon emission data set, the principal component analysis method is used for feature selection, and the principal components with a cumulative variance contribution rate greater than 80% are selected as new feature subsets to achieve dimensionality reduction of carbon emission data. Based on the carbon emission training data set after dimensionality reduction, the autoencoder neural network model is used to extract the low-dimensional semantic features of the data. The encoder is set to contain two fully connected layers, and the decoder is symmetrical with the encoder. The Adam optimizer and MSE loss function are used to train the model until convergence. The encoder part of the autoencoder neural network model is extracted as a feature extraction module for carbon emission data. The feature extraction module is used to extract features from the reduced-dimensional carbon emission training data to obtain the compressed carbon emission feature vector. The K-means clustering algorithm is used to cluster the compressed carbon emission feature vector, and the optimal cluster number K is determined by the elbow rule. According to the clustering results, the carbon emission data is divided into K carbon emission intensity levels, and the cluster labels are used as the annotations of carbon emission intensity. The decision tree algorithm is used to establish a carbon emission intensity classification model, with the compressed carbon emission feature vector as input and the carbon emission intensity level as output. The optimal features are selected by the information gain ratio to generate a carbon emission intensity classification decision tree. The accuracy, precision, recall and other indicators of the decision tree classification model are evaluated on the test set. The support vector regression algorithm is used to construct a carbon emission prediction model, and the Gaussian kernel function is selected. The compressed carbon emission feature vector and historical carbon emissions are used as input, and the future carbon emissions are used as output. The model hyperparameters are optimized by grid search, and the carbon emission prediction model is trained. The mean square error, mean absolute error and other indicators of the support vector regression model are evaluated on the test set. For new carbon emission data, we first use the feature extraction module to extract its low-dimensional semantic features, and then input the extracted features into the carbon emission intensity classification model to determine the carbon emission intensity level to which it belongs. Finally, we input the carbon emission feature vector into the carbon emission prediction model to obtain the predicted carbon emission value for a period of time in the future. The comprehensive carbon emission intensity level and carbon emission prediction results provide data support for the formulation of carbon emission reduction strategies.

[0025] Specifically, after obtaining the raw carbon emission data from the blockchain network, regular expressions are used to clean the data. For example, the regular expression "[^a-zA-Z0-9\u4e00-\u9fa5]" is used to match illegal characters and replace them with empty strings, and the regular expression "^\d+(\。\d+)?$" is used to match legal numeric data, and outliers that do not conform to the format are deleted. For missing values, the nearest neighbor interpolation method is used to repair them, that is, the nearest k non-missing values ​​before and after the missing value are selected, and their average value is calculated as the estimated value of the missing value. The commonly used k value is 3 or 5. According to the type, industry, scale and other attributes of the carbon emission source, the cleaned carbon emission data set is labeled to construct a structured carbon emission data set. For example, carbon emission sources are divided into types such as electricity, industry, and transportation, industries are divided into steel, cement, and chemicals, and scales are divided into large, medium, and small. One-Hot encoding is used to convert these categorical attributes into numerical features, that is, each category is mapped into a binary vector, the length of the vector is equal to the number of categories, and the value at each position is 0 or 1, indicating whether it belongs to the category. In view of the high-dimensional characteristics of the structured carbon emission data set, the principal component analysis method is used for feature selection. By calculating the covariance matrix and eigenvalues ​​of the features, the eigenvectors corresponding to the first k eigenvalues ​​with a cumulative variance contribution rate greater than 80% are selected as the principal components, and the original data set is mapped to the principal component space to achieve dimensionality reduction of carbon emission data. Based on the carbon emission data set after dimensionality reduction, the autoencoder neural network model is used to extract the low-dimensional semantic features of the data. The encoder is set to contain two fully connected layers. The first layer maps the input data to 128 dimensions and the second layer maps it to 64 dimensions. The decoder is symmetrical with the encoder and contains two fully connected layers and an output layer. The ReLU activation function and Adam optimizer are used. MSE is used as the loss function to train the model until convergence. The encoder part of the trained autoencoder is extracted as the feature extraction module of carbon emission data. The low-dimensional semantic features of the carbon emission data after dimensionality reduction are extracted to obtain a 64-dimensional compressed carbon emission feature vector. The K-means clustering algorithm is used to cluster the compressed carbon emission feature vector. By calculating the silhouette coefficient of the clustering results under different K values, the K value with the largest silhouette coefficient is selected as the optimal number of clusters, and the carbon emission data is divided into K carbon emission intensity levels. The decision tree algorithm is used to establish a carbon emission intensity classification model, with the compressed carbon emission feature vector as input and the carbon emission intensity level as output. By calculating the information gain ratio corresponding to each feature, the feature with the largest information gain ratio is selected as the root node, and the decision tree is recursively constructed until all samples belong to the same category or reach the preset depth threshold.The carbon emission prediction model was constructed using the support vector machine algorithm, and the Gaussian kernel function was selected. The compressed carbon emission feature vector and historical carbon emissions were used as input, and the future carbon emissions were used as output. The penalty coefficient C and kernel function parameter γ of the model were optimized through grid search. The model performance was evaluated using 5-fold cross validation, and the parameter combination with the smallest average error was selected to train the carbon emission prediction model. The new carbon emission data was input into the feature extraction module and the carbon emission intensity classification model to determine the carbon emission intensity level to which it belongs. The extracted feature vector was then input into the carbon emission prediction model to obtain the predicted value of carbon emissions in the future, providing data support for the formulation of carbon emission reduction strategies.

[0026] S103. According to the timestamp of the carbon emission data, the sliding time window method is used to segment the continuous time series carbon emission data, and the long short-term memory network is used to model the changes in carbon emission patterns in different time periods, explore the nonlinear relationship between carbon emissions and time, and obtain the carbon emission rules in fine-grained time periods.

[0027] The raw carbon emission data was obtained from the carbon emission monitoring system of the distribution network. The box plot method was used to detect outliers. If the value exceeded the upper and lower limits of the box plot, it was considered an outlier and deleted. For missing values, if the missing time period was short, linear interpolation was used to fill it. After completing outlier processing and missing value filling, a normalized carbon emission time series data set was obtained. According to the timestamp of the carbon emission data, the time window length was set to one hour and the sliding step length was set to ten minutes. The continuous time series data was divided into several carbon emission subsequences using the sliding time window method. For each carbon emission subsequence, the mean, variance, maximum value, and minimum value of the data were calculated as statistical features. At the same time, the discrete Fourier transform coefficients were calculated, and the first five low-frequency coefficients were selected as frequency domain features. The statistical features were combined with the frequency domain features to form the feature vector of the carbon emission subsequence. The K-means clustering algorithm was used to cluster the feature vectors of the carbon emission subsequence. The silhouette coefficients under different cluster numbers were calculated, and the number of clusters corresponding to the maximum silhouette coefficient was selected as the optimal number of clusters. According to the value of the cluster center, the clusters are sorted from small to large, and the labels "low emission mode", "medium emission mode" and "high emission mode" are assigned respectively. The cluster result labels are associated with the feature vectors of the carbon emission subsequences. A long short-term memory neural network model is constructed. The feature vectors of the carbon emission subsequences are used as input and the corresponding cluster labels are used as output to train the long short-term memory neural network model to predict the carbon emission mode according to the feature vectors. A self-attention mechanism layer is added before the output layer of the long short-term memory neural network model. The self-attention mechanism layer calculates the attention weight of each time point according to the hidden state of the long short-term memory neural network model, and applies the weight to the hidden state to highlight the information of the key time points. The trained long short-term memory neural network model is combined with the self-attention mechanism model to process the newly collected carbon emission data. The new data is divided into subsequences of the same length as the training data, and the statistical features and frequency domain features of each subsequence are extracted and input into the model to obtain the carbon emission mode prediction results of each subsequence. The carbon emission mode prediction results of all subsequences are counted. If the number of subsequences predicted as "high emission mode" exceeds the preset threshold, it is determined that the overall carbon emissions in the current period may exceed the standard, triggering the early warning mechanism and generating a carbon emission analysis report. Regularly collect updated carbon emission data from all links of the distribution network. Use the new data to recalculate the feature vector, and use the new data and the corresponding carbon emission pattern labels to retrain the long short-term memory neural network model combined with the self-attention mechanism to update the model parameters.

[0028] Specifically, after obtaining carbon emission data from the distribution network carbon emission monitoring system, the Pandas library of Python is used to clean and preprocess the data. By setting the threshold, the data exceeding the normal range by 20% is regarded as anomalies and removed; for missing values, linear interpolation is used to fill in to ensure the continuity of the data. Then, the NumPy library is used to slice the carbon emission data, setting the time window to 1 hour and the sliding step to 10 minutes to generate several carbon emission subsequences. For each subsequence, eight statistical features such as mean, variance, maximum value, and minimum value are calculated, and the SciPy library is used to perform Fourier transform on the data, extract the first 10 frequency domain feature coefficients, and finally obtain an 18-dimensional feature vector. The K-means algorithm in the Scikit-learn library is used to perform cluster analysis on the feature vectors. Through the evaluation of the silhouette coefficient and SSE index, the optimal number of clusters is determined to be 4, representing 4 typical carbon emission patterns. Then, the LSTM neural network model was constructed, and the input layer dimension was set to 18, the hidden layer dimension was set to 64, and the output layer dimension was set to 4, corresponding to 4 carbon emission modes. The feature vector of the carbon emission subsequence was used as input, and the corresponding cluster label was used as output. The LSTM model was trained for 100 epochs to realize the dynamic recognition of carbon emission modes. On the basis of LSTM, an attention mechanism layer was added, and the attention weight of each moment was calculated by the Softmax function. The hidden state of LSTM was dynamically adjusted to explore the nonlinear relationship between carbon emissions and time. The newly collected carbon emission data was predicted using the trained LSTM-Attention model. The new data was divided into several subsequences according to the 1-hour window, and the feature vectors were extracted and input into the model to obtain the carbon emission mode recognition results of each subsequence. The carbon emission mode distribution of each subsequence was counted. If more than 60% of the subsequences were identified as high carbon emission modes, it was determined that the carbon emissions in the current period may exceed the standard, triggering the early warning mechanism, generating a carbon emission analysis report, and notifying relevant personnel by email and SMS. At the same time, the carbon emission data is correlated with the distribution network operation data (such as load curves, equipment status, etc.), and the potential causes of carbon emission anomalies are explored through correlation analysis and causal inference, providing a basis for formulating targeted carbon emission reduction strategies. The LSTM-Attention model is retrained and optimized regularly to continuously improve the accuracy and robustness of carbon emission pattern recognition.

[0029] S104. Based on the electricity load data collected by smart meters, a convolutional neural network is used to extract the power curve characteristics of different types of electrical equipment. The various types of electrical equipment are identified through a softmax classifier. The carbon emission contribution of each device is calculated based on the rated power and usage time of the equipment, and a carbon emission decomposition model for the equipment is established.

[0030] Smart meters are used to collect power load data of users in various links of the distribution network. The load data covers voltage, current, active power, and reactive power information. The data collection frequency is once every five minutes. The collected power load data is transmitted to the data processing center through wired or wireless networks. According to the power load data cleaning strategy, incomplete and obviously abnormal data points are filtered out. If there are vacant values ​​in the data, the nearest neighbor interpolation method is used to complete the data. Complete data cleaning and completion. A one-dimensional convolution filter based on time series is applied to the cleaned power load data. The filter is set to three layers, the window size of each filter layer is three, the step size is one, and the ReLU activation function is used. After each layer of filter processing, thirty-two, sixty-four, and one hundred and twenty-eight feature data are output respectively. The one hundred and twenty-eight feature data output by the last layer of filters are used as the input features of the logistic regression classifier. The processed power load data is classified using a logistic regression classifier. The classification objects include various types of electrical equipment such as air conditioners, lighting, and motors. The logistic regression classifier identifies the power curve characteristics of different types of electrical equipment based on the pre-set classification threshold. Obtain the identification results of various types of electrical equipment. Obtain the rated power and usage time of various types of electrical equipment from the equipment information database. Use the equipment carbon emission calculation formula, combined with the obtained rated power, usage time and the preset equipment carbon emission factor, to calculate the carbon emission contribution of each electrical equipment. For the calculated carbon emission contribution values ​​of each device, build a linear fitting model for the relationship between the carbon emission contribution of the device and the device type. Use the linear fitting tool based on the least squares method for fitting. After obtaining the linear fitting model, extract features from the real-time power load data through a one-dimensional convolution filter based on time series. Use the extracted features as the input of the linear fitting model to calculate the real-time carbon emission contribution of each device. Use the real-time carbon emission contribution data of each device to formulate a carbon emission reduction optimization strategy. The strategy includes adjusting the equipment operation status according to the real-time carbon emission contribution of the equipment, recommending the replacement of high-energy consumption equipment with reference to the linear relationship between the carbon emission contribution of the equipment and the equipment type, and guiding users to optimize their electricity usage habits according to their electricity usage behavior patterns. The ultimate goal is to reduce the overall carbon emissions of the distribution network.

[0031] Specifically, the user power load data is collected in the distribution network of a certain industrial park using smart meters. For example, the load data collected from a certain factory A at 2:05 pm is: voltage 380 volts, current 150 amperes, active power 95 kilowatts, and reactive power 30 kilovars. This data is transmitted to the data processing center in real time through the optical fiber network. In order to ensure data quality, the system sets a data threshold range. If the active power data exceeds the range of 10 to 1000 kilowatts, it is regarded as abnormal data and is eliminated. For example, it is found that the active power of a data point is 1200 kilowatts, which exceeds the set range, so it is filtered. For missing data, the nearest neighbor interpolation method is used to fill it. Assuming that the active power data at 2:10 pm is missing, the active power data at the two time points before and after, namely 2:05 and 2:15, are averaged. If 2:05 is 95 kilowatts and 2:15 is 98 kilowatts, the average value is calculated to be 95 kilowatts, and this value is used to supplement the missing data. The cleaned data is input into the one-dimensional convolutional neural network for processing. The input data dimension is set to 1440 points, representing the total number of data points every five minutes in a day. The network contains three convolutional layers. The first layer is configured with 32 convolution kernels, each with a size of 3 and a step size of 1. The activation function is ReLU. Taking the first layer as an example, the input data is convolved with 32 convolution kernels to generate 32 feature maps, each with a dimension of 1438 points. After processing by three convolutional layers, the data features are deeply extracted to prepare for subsequent classification. Then, the convolution-processed data is sent to the softmax classifier for device type identification. The softmax classifier outputs the probability distribution of each category based on the pre-learned power curve features of various types of equipment, and the category with the highest probability is the identification result. Assuming that after a certain power curve is processed by the softmax classifier, the output result is that the probability of air conditioning is 9, the probability of lighting is 1, and the probability of motor is 05, then the device type corresponding to the power curve is identified as air conditioning. Call the device information database to obtain the rated power of the air conditioner is 5 kilowatts and the average daily operating time is 10 hours. The carbon emission calculation formula is used to quantify the carbon emission contribution. Assuming that the carbon emission factor of the air conditioner is 8 kg of carbon dioxide per kilowatt-hour, its daily carbon emissions are 5 kilowatts times 10 hours times 8 kg per kilowatt-hour, which is 40 kg of carbon dioxide. The above carbon emission calculation is performed for all identified electrical equipment. After obtaining the carbon emission contribution values ​​of all equipment, a carbon emission decomposition model is established. Using the linear regression method, the equipment type is used as the independent variable and the carbon emission contribution is used as the dependent variable to construct a linear regression equation. Taking the air conditioner as an example, by collecting multiple air conditioner sample data, a linear relationship model between the carbon emission contribution and the air conditioner operating parameters can be obtained, such as Y=aX+b, where Y represents the carbon emission contribution, X represents the power or operating time of the air conditioner, and a and b are model parameters.With the help of this model, the carbon emission contribution of each air conditioner can be calculated in real time. For example, by inputting the real-time power data of an air conditioner, the corresponding carbon emission value can be calculated. Furthermore, the system proposes a refined carbon emission reduction strategy based on the carbon emission contribution of each device and the carbon emission decomposition model. By analyzing the operating data of the air conditioner, if it is found that a certain air conditioner has been running for too long under low load, it is recommended to adjust the air conditioner operation strategy, such as changing to intermittent operation mode, so as to reduce energy consumption and carbon emissions. In addition, the system can also combine the energy efficiency information of the equipment to recommend a more energy-saving equipment model for replacement. At the same time, energy-saving reminders are pushed through the user interface to encourage users to adjust their electricity usage habits and jointly reduce carbon emissions.

[0032] S105. Build a user electricity consumption behavior profile based on user electricity consumption data, use the attention mechanism to learn the impact weight of user electricity consumption habits on carbon emissions, and use the clustering algorithm to mine the carbon emission patterns under different electricity consumption behaviors to form a user-oriented carbon emission explanation model.

[0033] The modified optimized version is as follows. Obtain the electricity consumption time and electricity consumption data of typical users from the customer information system. Apply the K-means clustering algorithm to cluster the user electricity consumption behavior data, divide the users into peak electricity users and valley electricity users, and form a user electricity consumption portrait. The user electricity consumption portrait contains the characteristics of the user's electricity consumption behavior. Obtain historical operation data from the distribution network database and sensors. The historical operation data includes voltage, current, power, temperature, load type, air temperature and humidity. Form the first data set. Clean the first data set. Use the Lagrange interpolation method to handle missing values. Use the 3σ criterion to eliminate outliers. Form the second data set. Use the principal component analysis method to extract the carbon emission feature quantity of the second data set. Determine the carbon emission feature set based on the principal component with a cumulative contribution rate greater than 80%. Form the third data set. Use StandardScaler to standardize the third data set. Standardization uses Z-score standardization. The calculation formula is: z = (x-μ) / σ. Where x is the original data, μ is the mean, and σ is the standard deviation. Form the fourth data set. Use XGBoost to build a carbon emission prediction model. Grid search and five-fold cross validation are used to optimize the hyperparameters of the XGBoost tool. The root mean square error is used as an evaluation indicator. The optimized carbon emission prediction model is obtained. The optimized carbon emission prediction model is applied to the real-time operation data of the distribution network. The predicted value of carbon emissions in the future period is obtained. The linear regression method is used to analyze the power load characteristics and carbon emission intensity distribution of the distribution network. Determine the key influencing factors of carbon emissions. The key influencing factors of carbon emissions and user power consumption behavior characteristics are used as input. The user power consumption behavior characteristics use the user power consumption behavior characteristics in the user power consumption profile obtained in step 1. The attention mechanism is used to learn the influence weight of user power consumption habits on carbon emissions. Using the influence weight, the K-means clustering algorithm is used to mine the carbon emission patterns under different power consumption behaviors. A user-oriented carbon emission explanation model is formed. The user power consumption profile is associated with the prediction results of carbon emissions and energy consumption. A peak-shifting power consumption strategy is formulated for peak power users. The peak-shifting power consumption strategy guides power consumption during the low load period of the power grid. An energy efficiency management strategy is formulated for high energy consumption users. Energy efficiency management strategies guide the adoption of energy-saving equipment and technologies. The power communication network sends energy-saving and emission reduction strategies to smart meters. Smart meters optimize and control users' electricity consumption behavior according to strategy requirements. Optimization control is to limit non-essential loads during peak hours. Increase energy storage charging during off-peak hours. Smart meters collect users' actual electricity consumption data. Evaluate the implementation effect of energy-saving and emission reduction strategies. If the actual amount of electricity saved does not reach the preset threshold. Adjust the peak time. Improve energy efficiency standards. Use the latest carbon emission data, electricity data and operating parameters to regularly retrain and optimize the parameters of the carbon emission prediction model built by the XGBoost tool.

[0034] Specifically, the electricity consumption behavior data of 10,000 typical users are obtained through the customer information system. The electricity consumption behavior data includes electricity consumption time and electricity consumption in hours. The K-means clustering algorithm is used to analyze the data and divide the users into peak electricity users and valley electricity users according to their electricity consumption patterns. The number of clusters is set to 2. By calculating the deviation between the electricity consumption of each user in different time periods and the average electricity consumption, 3,000 peak electricity users and 7,000 valley electricity users are identified to form a user electricity consumption profile. At the same time, the electricity consumption profile of typical users is associated with the carbon emissions and energy consumption results predicted by the long short-term memory neural network. The long short-term memory neural network model is trained by collecting the electricity consumption data, carbon emission data and ambient temperature data of users in the past three years. The model contains 3 hidden layers, 128 neurons in each layer, a learning rate of 0.01, and 100 training rounds. Assuming that the long short-term memory neural network model predicts that the carbon emissions of a peak electricity user will exceed the average level by 20% in the next week, an energy-saving and emission reduction strategy is formulated. For these 3,000 peak electricity users, a peak electricity consumption strategy of time-of-use electricity prices is adopted to guide users to use electricity during the low-load period of the power grid, that is, from 23:00 to 7:00 the next day. The electricity price during the low-load period is set at 60% of the peak-load price. For the remaining 7,000 low-load electricity users, 1,000 high-energy-consuming households are promoted to use LED lighting equipment and variable-frequency air conditioners. Compared with traditional equipment, these energy-saving equipment has an energy efficiency improvement of 15%, and government subsidies are provided, with the subsidy amount being 10% of the equipment cost. These policies are sent to users' smart meters through the power communication network. Smart meters optimize and control users' electricity consumption behavior according to policy requirements. For example, during peak hours, smart meters can remotely shut down or limit users' non-essential high-power appliances, such as water heaters and air conditioners. During low-load periods, users' energy storage devices, such as electric vehicles, are controlled to charge, and the maximum charging power is set to 10 kilowatts. Smart meters collect users' actual electricity consumption data every 15 minutes, including voltage, current, active power, reactive power, etc., and calculate the electricity saved per hour. Assuming that the power saving threshold is set to 5% daily power saving, if the actual power saving does not reach this threshold, the system will automatically adjust the peak time period or improve the energy efficiency standard. For example, the peak time period is adjusted to 22:00 to 6:00 the next day, or high-energy-consuming households are required to replace equipment with higher energy efficiency levels, such as air conditioners with first-level energy efficiency. The system will continue to optimize the long-term and short-term memory neural network model and energy-saving and emission reduction strategies. Every quarter, the long-term and short-term memory neural network model is retrained and parameterized using the latest carbon emission data, power data, and equipment operating parameters. By introducing the attention mechanism to optimize the model structure and improve the prediction accuracy, the root mean square error of the predicted carbon emissions is reduced to 3%. The distribution network operation data and carbon emission data are correlated and fused through the data warehouse, and the principal component analysis algorithm is used to extract carbon emission characteristics.By analyzing historical data, 10 possible carbon emission influencing factors were selected, including gross industrial output value, population, vehicle ownership, land use type, temperature, humidity, wind speed, rainfall, vegetation coverage, energy consumption, etc. The analysis results show that the cumulative contribution rate of the first four principal components reached 85%, which was determined as the carbon emission feature set to construct the XGBoost regression model. The carbon emission feature set includes gross industrial output value, population, vehicle ownership and land use type. The grid search algorithm and five-fold cross validation were used to determine the optimal hyperparameter combination of XGBoost, such as the learning rate was set to 1, the maximum tree depth was set to 6, and the subsample ratio was set to 8. The carbon emission trend in the next year was predicted through the optimal carbon emission prediction model. The historical operation data of the past year was obtained from the distribution network database and sensors, with a data collection frequency of once per minute, including voltage, current, power, temperature, load type, temperature and humidity, to form the first data set. The first data set was cleaned, and the missing data was filled using the Lagrangian interpolation method. Assuming that the missing data is less than 1%, the third-order Lagrangian interpolation method was used to fill it. The 3σ criterion is used to remove outliers, and the outlier threshold is set to plus or minus three times the standard deviation. After obtaining the second data set, the random forest algorithm is used for feature selection, and the number of trees is set to 100. According to the feature importance, the top 6 features are selected, including voltage, current, power, load type, air temperature, and temperature, to form a feature subset related to carbon emissions, namely the third data set. The third data set is standardized using the standardization method StandardScaler to eliminate the influence of the dimension, and the fourth data set is obtained, which is input into the carbon emission prediction model constructed by the XGBoost algorithm. The prediction model is hyperparameter tuned using GridSearchCV, and the parameter grid is set, including the learning rate from 01 to 2, the step size is 01, the maximum tree depth from 3 to 10, the step size is 1, and the five-fold cross validation is used. The root mean square error is used as the evaluation indicator to obtain the optimized prediction model. The optimized prediction model is applied to the real-time operation data of the distribution network to predict the carbon emissions in the next 24 hours, with a time interval of 1 hour. Then, the regression analysis method is used to analyze the power load characteristics and carbon emission intensity distribution of the distribution network. For example, the analysis found that the peak hours of electricity consumption are from 9:00 to 12:00 and from 14:00 to 17:00 on weekdays, and the carbon emission intensity increases accordingly, while the carbon emission intensity on holidays is relatively low and evenly distributed. The key factors affecting carbon emissions include load rate, power factor, line loss, etc., which are monitored in real time through the distribution automation system. Energy-saving and emission reduction measures and optimization strategies are formulated based on these key influencing factors. For example, by adjusting the transformer tap, the load rate is controlled between 60% and 80%, and by installing reactive power compensation devices, the power factor is increased to above 95, thereby reducing line loss by 5%.The key influencing factors of carbon emissions and the characteristics of user electricity consumption behavior are used as input, and the attention mechanism is used to learn the weight of the impact of user electricity consumption habits on carbon emissions. For example, the model analysis found that a user used a large number of electric water heaters during the evening peak period, and the impact weight of his electricity consumption behavior on carbon emissions was 8, while another user who used a washing machine during the off-peak period had an impact weight of 2. The clustering algorithm is used to mine the carbon emission patterns under different electricity consumption behaviors. The number of clusters is set to 3, and users are divided into three categories: high carbon emissions, medium carbon emissions, and low carbon emissions. A user-oriented carbon emission explanation model is formed to provide users with personalized energy-saving and emission reduction suggestions, such as suggesting that high-carbon emission users adjust their electricity consumption time periods and suggesting that medium-carbon emission users optimize the configuration of electrical equipment.

[0035] S106. Use the Apriori association rule mining algorithm to extract frequent patterns in carbon emission data from three dimensions: time period, equipment, and user, screen out key factor combinations that have a significant impact on carbon emissions, and embed the rules into the blockchain through smart contracts to perform carbon emission traceability.

[0036] According to the IoT sensors and smart meters deployed in each link of the distribution network, carbon emission data and electricity data are collected in real time to obtain detailed operating parameters and energy consumption information. According to the pre-established data cleaning rules, regular expressions are used to clean non-numeric data and special characters in the collected data set. In the cleaned data set, if there are missing values, the Lagrange interpolation method is used to complete the data. The 3σ criterion is used to remove outliers from the completed data set. The data set after removing outliers is standardized to eliminate the impact of data dimension. The K-means algorithm is used to perform cluster analysis based on the electricity consumption time characteristics and electricity consumption dimensions in the data. The results are used to divide multiple typical user portraits such as industrial users, commercial users, and residential users. For the carbon emission data, electricity data, and operating parameters obtained after cleaning, the pivot table tool is used to perform data statistics according to the time period, equipment, and user dimensions to obtain transactional data. The minimum support and minimum confidence are set, and the association rule algorithm is applied to mine frequent item sets and association rules from transactional data. According to the obtained association rules, the key factor combination that significantly affects carbon emissions is determined. A blockchain network is built on the Ethereum platform to store carbon emission data, electricity data, and mined association rules, and distributed data storage and non-tamperability are achieved through smart contracts. In the blockchain network, the carbon emission intensity of each block is judged based on the carbon emission data and electricity data, and the carbon emission intensity is divided into three levels: high, medium, and low. The preset threshold of carbon emission intensity is determined using a simple moving average algorithm. If the carbon emission intensity of a block exceeds the preset threshold, the corresponding responsible entity is determined to bear a greater carbon emission reduction responsibility. Based on predefined rules, the carbon emission data and electricity data on the chain are traced and audited. During the traceability audit process, the authenticity of the data source is verified, and the integrity and consistency of the data are checked. The results of the data traceability audit are written into the blockchain using smart contracts to ensure that the audit results cannot be tampered with.

[0037] Specifically, when deploying IoT sensors and smart meters, for example, deploying temperature sensor TH101, humidity sensor HS203, current sensor CS305 and three-phase smart meter ABC987 in the industrial park, respectively, to monitor the operating temperature of the transformer, ambient humidity, line current and overall power consumption. Set the data collection every 10 seconds, the temperature sensor returns a Celsius value, such as 26°C, the humidity sensor returns a percentage value, such as 68%, the current sensor returns an ampere value, such as 152A, and the smart meter returns a kilowatt-hour value, such as 325kWh. These real-time data can be used to analyze the energy consumption mode and equipment operation status of the industrial park at different times. For example, too high current data may mean that the equipment is overloaded, and too high temperature may indicate that the equipment is at risk of overheating. Regular expressions are used to clean non-numeric data and special characters in the data set. For example, for the data collected by smart meters, if there are strings such as "failed to read" or "--", the regular expression "[^0-]" can be used to match and remove non-numeric and non-decimal characters, and "325kWh" can be cleaned to "325" to facilitate subsequent numerical calculations. Lagrange interpolation method is used to handle missing values. Assuming that the temperature data at 10:00 is missing, and the temperature data at 9:50 and 10:10 are 25℃ and 21℃ respectively, the temperature estimate at 10:00 is calculated to be 23℃ using the Lagrange interpolation formula to ensure data continuity. The 3σ criterion is used to eliminate outliers. For example, for a set of current data, the average value μ is calculated to be 150A and the standard deviation σ is 5A. If a current reading exceeds 165A (μ+3σ), the data point is considered to be an outlier and is eliminated to avoid affecting the accuracy of the overall analysis due to individual extreme data. The cleaned data is clustered and analyzed. Taking the daily electricity consumption data of residential users as an example, the K-means clustering algorithm is used, and the number of clusters is set to 3. Based on the characteristics of electricity consumption and electricity consumption time, users are divided into three groups: low electricity consumption (less than 5kWh per day), medium electricity consumption (5-15kWh per day) and high electricity consumption (more than 15kWh per day), which helps power companies provide personalized energy-saving suggestions and services for different user groups. The Apriori algorithm is used to mine association rules for carbon emission data and electricity data. For example, the data of a commercial area is analyzed, and the minimum support is set to 20% and the minimum confidence is 70%. It is found that when the electricity consumption of air-conditioning equipment exceeds 100kWh and the electricity consumption of lighting equipment exceeds 50kWh, the association rule of carbon emissions increases significantly is found. The support of this association rule is 25% and the confidence is 80%, indicating that the rule has a certain universality and high credibility in the data set. By mining such rules, the key factor combination that affects carbon emissions can be identified, such as the simultaneous operation of air-conditioning and lighting systems under high load.A blockchain network is built on the Ethereum platform, and carbon emission data, electricity data, and association rules are deployed on the chain through smart contracts. For example, the above association rule "air conditioning electricity consumption > 100kWh and lighting electricity consumption > 50kWh => carbon emissions increase significantly" is encoded as a smart contract function to ensure data transparency and immutability. Each block contains data within a specific time period. For example, the data per hour is packaged into a block, and the carbon emission intensity of the block is calculated. The preset threshold of carbon emission intensity is determined by an adaptive threshold algorithm. For example, the average carbon emission intensity calculated based on historical data is 8kg / kWh, and the preset threshold is dynamically set to 2kg / kWh (average value plus a certain proportion of standard deviation). If the carbon emission intensity of a block is 3kg / kWh, which exceeds the threshold, it is determined that the responsible entity needs to bear greater emission reduction responsibilities, such as paying higher carbon emission taxes or increasing the purchase of carbon emission quotas. Based on predefined rules, the data on the chain is traced and audited. For example, the rule is set as "the sensor number must be unique and belong to the pre-registered device list". The smart contract code is used to automatically verify this rule to ensure the authenticity and consistency of the data. The audit results are written into the blockchain through the smart contract. For example, the audit contract records "block height 1000, data audit passed, timestamp 2024-05-20 15:00:00". A blockchain-based carbon emission rights trading platform is built, and a two-way auction mechanism is used for quota trading. For example, a company has 100 tons of carbon emission quotas remaining through energy-saving measures, sets the starting price at 50 yuan per ton, and publishes auction information through smart contracts; another company needs to purchase additional carbon emission quotas due to production needs, participates in the auction through smart contracts, and finally trades at 55 yuan per ton, realizing the effective allocation of carbon emission rights and promoting the realization of carbon emission reduction goals.

[0038] S107. Taking the carbon emission influencing factors such as time period, equipment, user, and rules as input features, and carbon emissions as the target variable, the LightGBM regression algorithm is used to establish a carbon emissions prediction model. Using the transfer learning method, the pre-trained carbon emission model is applied to different regions and scenarios to improve the generalization performance of the model.

[0039] Data on carbon emission influencing factors are collected from all links of the distribution network, including time periods, equipment parameters, and user electricity consumption behaviors. The data is cleaned and standardized to obtain a standardized carbon emission feature data set. The Pearson correlation coefficient is used to calculate the correlation coefficient between each feature in the carbon emission feature data set and the carbon emissions, and the feature subset that is highly correlated with carbon emissions is screened out according to the absolute value of the correlation coefficient. The input feature vector of the carbon emission prediction model is constructed using the screened feature subset. With carbon emissions as the target variable and the key feature subset as the input feature, the carbon emission prediction model is trained using the LightGBM regression algorithm. The grid search method is used to find the optimal hyperparameter combination in the preset parameter space, and the optimal parameters of the LightGBM regression algorithm are determined with the minimum mean square error as the goal. The cross-validation method is used to evaluate the generalization performance of the carbon emission prediction model. If the mean square error of the model in cross-validation does not reach the preset threshold, the screening threshold of the Pearson correlation coefficient is adjusted, the feature subset is reselected, and the optimal hyperparameters of the LightGBM regression algorithm are re-searched in the preset parameter space. Repeat this process until the model performance meets the preset mean square error threshold requirement. For distribution networks in different regions and scenarios, obtain the data of carbon emission influencing factors in the source, grid, load and storage links, apply the same cleaning and standardization processing methods as in the training phase, and map the data to a feature space consistent with the training data set. Use the trained LightGBM regression algorithm model to predict the data of new regions and scenarios, and output the predicted carbon emissions of each link. Fix all layer parameters except the output layer in the trained model, and only use the data of the new region and scenario to update the weight of the output layer to obtain the fine-tuned LightGBM regression algorithm model. Use the fine-tuned model to re-predict the data of the new region and scenario to obtain the updated carbon emission prediction value. Obtain the carbon emission data output by the carbon emission tracing model based on the emission factor method. Calculate the absolute error between the carbon emissions predicted by the LightGBM regression algorithm model and the carbon emissions output by the emission factor method model. If the absolute error exceeds the preset threshold, adjust the weight parameters of the output layer in the LightGBM regression algorithm model, and adjust the weights of the relevant features according to the results of the emission factor method model. The adjusted LightGBM regression algorithm model is used to re-predict. Based on the prediction results of the LightGBM regression algorithm model, an accurate assessment report on carbon emissions in each link of the distribution network is generated, providing data support and decision-making basis for the formulation of carbon emission reduction strategies for the distribution network.

[0040] Specifically, data on carbon emission influencing factors are collected from various links of the distribution network, such as electricity consumption and equipment efficiency in different periods, and the data are cleaned and standardized. The Pearson correlation coefficient is used to calculate the correlation between each feature and carbon emissions, and features with a correlation coefficient greater than 8 are selected to construct the input feature vector. Taking carbon emissions as the target variable, the prediction model is trained using the LightGBM regression algorithm, and hyperparameters are optimized through 5-fold cross-validation and grid search. For example, the learning_rate is adjusted from 01 to 05, which improves the prediction accuracy by 2%. The data of the new area is predicted. If the mean square error exceeds 5, the model is fine-tuned with the new data and the parameters are updated. The prediction results are compared with the electric carbon traceability results, and the carbon emission error is calculated. If it exceeds 5%, the traceability and prediction models are optimized. Finally, the results of the two methods are combined to generate an accurate assessment report on carbon emissions in each link of the distribution network, providing data support for the formulation of carbon emission reduction strategies.

[0041] S108. Adopt the model integration strategy to weightedly integrate the time period model, equipment decomposition model, user interpretation model and rule mining model to form a fine-grained multi-dimensional comprehensive carbon emission analysis model. Combined with the reinforcement learning algorithm, the carbon emission control strategy is optimized to achieve coordinated emission reduction in the source, grid, load and storage links.

[0042] S1: Obtain the carbon emission data, electricity data and operating parameters of each link of the distribution network source, grid, load and storage, clean and normalize the data, and store the processed data in a unified format in the database.

[0043] S2: Based on the time characteristics, the K-means clustering algorithm is used to cluster the carbon emission data and build carbon emission models for different time periods. Based on the equipment type and parameters, the decision tree algorithm is used to classify the carbon emission data and build the equipment carbon emission decomposition model. Based on the user attribute characteristics, the FP-Growth algorithm is used to mine the frequent patterns of user carbon emission behavior and build the user carbon emission explanation model. S3: The Apriori algorithm is used to mine the association rules of the pre-processed carbon emission data, power data and operating parameters to find the association rules between the data, such as "the carbon emission of equipment type A is higher than that of equipment type B", etc., to determine the key factors affecting carbon emissions and build a carbon emission rule mining model. S4: The weighted average method is used to fuse the output results of the time period model, equipment decomposition model, user explanation model and rule mining model to obtain a comprehensive carbon emission analysis model. The weights of each model are traversed through grid search, and the model performance of each group of weights is calculated using cross-validation. The weight combination with the best performance is selected to improve the accuracy and generalization ability of the model. S5: The carbon emission data, electricity data and operating parameters are integrated, and the carbon emission influencing factors are used as input features and the carbon emission amount is used as the output indicator to construct a long short-term memory neural network model. The network weights are optimized using the gradient descent method, and the prediction error is minimized through the back propagation algorithm to train the carbon emission prediction model. S6: The deep reinforcement learning algorithm is combined with the long short-term memory neural network model, and the grid operation state is used as the reinforcement learning environment state, the optimization control instructions of the distribution equipment are used as the reinforcement learning action, and the reduction of carbon emissions is used as the reward function. Through continuous interactive learning with the environment, the optimal carbon emission control strategy is generated. S7: The optimal carbon emission control strategy is converted into specific control instructions for the distribution equipment, including adjusting the transformer tap, optimizing the switching of reactive compensation capacitors, balancing load distribution, etc., and sent to the distribution automation terminal RTU through the power communication network, and the RTU optimizes the control of the distribution equipment according to the instructions. S8: The carbon emission-related data before and after the optimization control are collected through the smart meters and environmental sensors on the distribution equipment, the changes in the carbon emission intensity per unit of electricity are calculated, and the emission reduction effect of the carbon emission control strategy is evaluated. If the emission reduction effect does not reach the expected target, return to step S6 and re-perform strategy optimization iteration. S9: Continue to collect the latest carbon emission data and operation data of the source, grid, load and storage links, and regularly use new data to retrain and optimize the comprehensive carbon emission analysis model and control strategy. Combined with the latest carbon emission policies and carbon trading market prices, dynamically adjust the distribution network emission reduction targets and reward and punishment mechanisms to achieve adaptive closed-loop optimization of distribution network carbon emission analysis and control, and support the economical and efficient operation and low-carbon sustainable development of the distribution network.

[0044] Specifically, the carbon emission, electricity and operation data of each link of the distribution network are first collected and preprocessed, and the data are cleaned and standardized using methods such as maximum and minimum normalization. Then the K-means clustering algorithm is used to divide the carbon emission data into three categories: peak, flat and valley according to the time characteristics, and the C5 decision tree algorithm is used to decompose the carbon emissions into each device according to the device type and parameters. Then the Apriori algorithm is used to mine the association rules between user electricity consumption behavior and carbon emissions. For example, the confidence of the association between the air conditioner use period and the peak of carbon emissions reaches 85%. Then the results of each model are weighted averaged and fused. The weights are obtained by grid search optimization. The accuracy of the comprehensive model on the test set reaches 92%. A carbon emission prediction model is constructed based on the LSTM neural network, and the operation parameters are optimized by combining the reinforcement learning algorithm. For example, adjusting the air conditioner temperature setting from 26℃ to 28℃ can reduce carbon emissions by 5%. Finally, the optimization strategy is converted into a control instruction and sent to the distribution automation terminal for execution. The emission reduction effect is evaluated through sensor data collection, and the model and strategy are updated regularly.

[0045] S109. Generate adversarial networks to expand the diversity of carbon emission data and simulate carbon emission conditions under different extreme scenarios. Visual traceability of the entire carbon emission process data is achieved through blockchain browsers, providing a decision-making basis for accurately calculating carbon emissions and formulating emission reduction policies.

[0046] The modified optimized version is as follows. The distribution network operation data is obtained from the distribution network operation management system, and the historical carbon emission data is obtained from the historical database. After obtaining the historical carbon emission data, the Kolmogorov-Smirnov test is used to determine whether the historical carbon emission data conforms to the normal distribution. If the Kolmogorov-Smirnov test result shows that the historical carbon emission data does not conform to the normal distribution, the data distribution form is adjusted by logarithmic transformation to make it fit the normal distribution. The TensorFlow library is used to implement the generative adversarial network to generate simulated data that conforms to the actual carbon emission data distribution and expand the diversity of carbon emission data. The simulated data is integrated into the original data set to form an expanded data set, the mean of the expanded data set is calculated using the mean function, and the variance of the expanded data set is calculated using the variance function. According to the mean and variance results of the expanded data set, the temperature and humidity data are obtained from the meteorological database as environmental factor data. For the expanded data set, the extreme high temperature scene data set is screened out according to the records with temperatures higher than the preset high temperature threshold. The ARIMA model is used to perform time series analysis using the extreme high temperature scene data set to predict the carbon emission trend of the distribution network under the extreme high temperature scenario. The blockchain browser is deployed using the Hyperledger Fabric platform to present the carbon emission data of each link of the distribution network and the forecast trend obtained by the ARIMA model in the blockchain browser. The blockchain browser uses the data traceability function of Hyperledger Fabric to trace the source of carbon emission data and its change trajectory, and realize the transparent display of carbon emission data association information and the logical relationship between data.

[0047] Specifically, we first obtained the operation data and carbon emission data of the last year from the distribution network operation monitoring system and the carbon emission monitoring platform, totaling 30,000. The carbon emission data was tested using the Shapiro-Wilk normality test method, and the p value was 03, which was less than the significance level of 05, so it was judged that the data did not obey the normal distribution. The data was logarithmically transformed, and the transformed data passed the normality test. Then the CGAN network was used to expand the data to 100,000, in which the discriminator used a 4-layer fully connected network and the generator used a 6-layer deconvolution network. The mean of the expanded data set was 2 and the variance was 4, which was consistent with the actual data distribution. Combined with the extreme weather data provided by the meteorological department, 10 extreme weather scenarios such as extreme high temperature and typhoon were constructed, and carbon emission data under the corresponding scenarios were generated as specific scenario data sets. The data set was used to predict the trend of carbon emissions under extreme weather. The prediction results showed that carbon emissions in high temperature weather increased by 20% compared with normal temperature. Finally, the data traceability information was presented on the blockchain browser, including the time, equipment, and value of data generation, so that the entire process of data flow could be transparent and traceable. The blockchain browser links data-related information through hash pointers, making it easier for users to explore the logical relationship between data.

[0048] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A distribution network carbon emission analysis method, characterized in that: The method comprises: According to the topology of the distribution network and the type of carbon emission sources, IoT sensors and smart meters are used to collect carbon emission values, timestamps, geographic locations and other multi-dimensional carbon emission raw data at different time and space granularities. Distributed ledger technology is used to store heterogeneous carbon emission data in a unified format in the blockchain network, and a multi-node consensus mechanism is used to verify the authenticity and reliability of carbon emission data. The raw carbon emission data in the blockchain is cleaned, denoised and repaired, and a structured carbon emission data set is constructed by combining the carbon emission source type, intensity level and other attribute annotations. In view of the high-dimensional characteristics of carbon emission data, an autoencoder neural network is used to extract the low-dimensional semantic features of the data set to achieve dimensionality reduction and compression of carbon emission data; According to the timestamp of carbon emission data, the sliding time window method is used to segment the continuous time series carbon emission data, and the long short-term memory network is used to model the changes in carbon emission patterns in different time periods, explore the nonlinear relationship between carbon emissions and time, and obtain the carbon emission rules in fine-grained time periods; Based on the power load data collected by smart meters, a convolutional neural network is used to extract the power curve features of different types of power equipment. The softmax classifier is used to identify various types of power equipment. The carbon emission contribution of each device is calculated based on the rated power and usage time of the equipment, and a carbon emission decomposition model for the equipment is established. Build a user electricity consumption behavior profile based on user electricity data, use the attention mechanism to learn the impact weight of user electricity consumption habits on carbon emissions, and use clustering algorithms to mine carbon emission patterns under different electricity consumption behaviors to form a user-oriented carbon emission explanation model; The Apriori association rule mining algorithm is used to extract frequent patterns in carbon emission data from three dimensions: time period, equipment, and user, and to screen out key factor combinations that have a significant impact on carbon emissions. The rules are then embedded in the blockchain through smart contracts to perform carbon emission traceability. Taking carbon emission influencing factors such as time period, equipment, user, and rules as input features, and carbon emissions as the target variable, the LightGBM regression algorithm is used to establish a carbon emissions prediction model. The pre-trained carbon emission model is applied to different regions and scenarios using the transfer learning method to improve the generalization performance of the model. Adopting the model integration strategy, the time period model, equipment decomposition model, user interpretation model and rule mining model are weighted and integrated to form a fine-grained and multi-dimensional comprehensive carbon emission analysis model. Combined with the reinforcement learning algorithm, the carbon emission control strategy is optimized to achieve coordinated emission reduction in the source, grid, load and storage links. Generative adversarial networks are used to expand the diversity of carbon emission data, simulate carbon emission conditions under different extreme scenarios, and realize visual traceability of the entire carbon emission process data through blockchain browsers, providing a decision-making basis for accurately calculating carbon emissions and formulating emission reduction policies.

2. The method according to claim 1, characterized in that According to the distribution network topology and carbon emission source types, IoT sensors and smart meters are used to collect carbon emission values, timestamps, geographic locations and other multi-dimensional carbon emission raw data of different time and space granularities, and the heterogeneous carbon emission data are stored in a unified format in the blockchain network through distributed ledger technology. The authenticity and reliability of the carbon emission data are verified using a multi-node consensus mechanism, including: Deploy IoT sensors and smart meters in all links of the distribution network, and determine the placement and collection frequency of sensors and smart meters based on the topology of the distribution network and the type of carbon emission sources; Collect information such as carbon emission values, timestamps, and geographic locations at different time and space granularities of the distribution network; Obtaining raw data on carbon emissions from distribution networks from sensors and smart meters; Clean and compress the raw data of carbon emissions from the distribution network obtained on the edge computing device; Anomaly detection algorithms are used to identify and remove abnormal data from the original data of carbon emissions from the distribution network, thus obtaining a high-quality data set of carbon emissions from the distribution network. Upload the processed distribution network carbon emission dataset to the blockchain network; Using distributed ledger technology to store distribution network carbon emissions datasets in a unified format; Smart contract technology is used to define the storage structure and access rules of the distribution network carbon emission dataset to ensure the security and privacy of the distribution network carbon emission dataset; Deploy a multi-node consensus mechanism in the blockchain network to verify the carbon emission data set of the distribution network uploaded to the chain; If the distribution network carbon emission dataset passes consensus verification, it will be permanently recorded in the blockchain ledger; If the distribution network carbon emission dataset fails to pass verification, it will be refused to be written into the ledger; For the distribution network carbon emission dataset recorded in the blockchain ledger, the temporal and spatial characteristics of distribution network carbon emissions are analyzed using time series decomposition and spatial autocorrelation analysis methods; Obtain distribution network carbon emission datasets from blockchain ledgers; Use SPSS tools to perform time series decomposition on the distribution network carbon emission data set to obtain the peak periods of distribution network carbon emission; Use R language to perform spatial autocorrelation analysis on the distribution network carbon emission dataset to obtain high-incidence areas of distribution network carbon emissions; Based on the time series decomposition results and spatial autocorrelation analysis results, the key influencing factors of carbon emissions in the distribution network are determined; A distribution network carbon emission prediction model is established using the long short-term memory neural network (LSTM) algorithm; Based on the distribution network carbon emission data set, the multi-step prediction model is trained using historical carbon emission data as training samples and the key influencing factors of distribution network carbon emission as model input; According to the trained distribution network carbon emission prediction model, the predicted value of the distribution network carbon emission in the future period is obtained; Compare the predicted carbon emissions of the distribution network with the actual carbon emissions data of the distribution network obtained from sensors and smart meters; According to the comparison results of the distribution network carbon emission prediction value and the actual carbon emission data of the distribution network, the accuracy of the distribution network carbon emission prediction model is evaluated; If the prediction error of the distribution network carbon emission prediction model exceeds the preset threshold, the distribution network carbon emission prediction model is fine-tuned and optimized using the newly collected distribution network carbon emission data; According to the predicted value of carbon emissions of the distribution network in the future, the genetic algorithm is used to formulate the energy-saving and emission reduction strategy of the distribution network; Taking minimizing carbon emissions of the distribution network and maximizing power supply reliability as the optimization goals, the optimal grid dispatching plan and equipment operating parameters are solved in the MATLAB optimization toolbox to obtain the optimal energy-saving and emission reduction strategy for the distribution network. Send the optimal energy-saving and emission-reduction strategy for the distribution network to the distribution automation master station; Optimize the control of the distribution network by controlling switches and adjusting equipment parameters through the distribution automation system; Based on the real-time carbon emission data of the distribution network obtained from sensors and smart meters, the implementation effect of the optimal energy-saving and emission reduction strategy of the distribution network is evaluated.

3. The method according to claim 1, characterized in that The raw carbon emission data in the blockchain is cleaned, denoised and repaired, and a structured carbon emission data set is constructed by combining attribute annotations such as carbon emission source type and intensity level; in view of the high-dimensional characteristics of carbon emission data, an autoencoder neural network is used to extract low-dimensional semantic features of the data set to achieve dimensionality reduction and compression of carbon emission data, including: Obtain raw carbon emission data from the blockchain network, clean the data using regular expressions to remove illegal characters and outliers, and use the nearest neighbor interpolation method to repair missing values ​​to obtain a cleaned carbon emission data set; According to the attributes of carbon emission sources such as type, industry and scale, the cleaned carbon emission dataset is labeled to construct a structured carbon emission dataset; The structured carbon emission dataset is divided into a training set and a test set with a ratio of 8:2; The categorical attributes in the training set are converted into numerical features using One-Hot encoding; In view of the high-dimensional characteristics of the carbon emission data set, the principal component analysis method is used for feature selection, and the principal components with cumulative variance contribution rate greater than 80% are selected as new feature subsets to achieve dimensionality reduction of carbon emission data; Based on the carbon emission training dataset after dimensionality reduction, an autoencoder neural network model is used to extract low-dimensional semantic features of the data; Set the encoder to contain two fully connected layers, and the decoder to be symmetrical with the encoder. Use the Adam optimizer and the MSE loss function to train the model until convergence. Extract the encoder part of the autoencoder neural network model as a feature extraction module for carbon emission data; Use the feature extraction module to extract features from the carbon emission training data after dimension reduction to obtain a compressed carbon emission feature vector; The K-means clustering algorithm is used to perform cluster analysis on the compressed carbon emission feature vectors, and the optimal cluster number K is determined by the elbow rule. According to the clustering results, the carbon emission data is divided into K carbon emission intensity levels, and the cluster labels are used as the labels of carbon emission intensity; A carbon emission intensity classification model is established using a decision tree algorithm, with the compressed carbon emission feature vector as input and the carbon emission intensity level as output. The optimal feature is selected through the information gain ratio to generate a carbon emission intensity classification decision tree. Evaluate the accuracy, precision, and recall of the decision tree classification model on the test set; A carbon emission prediction model was constructed using the support vector regression algorithm. The Gaussian kernel function was selected, with the compressed carbon emission feature vector and historical carbon emissions as input and future carbon emissions as output. The model hyperparameters were optimized through grid search, and the carbon emission prediction model was trained. Evaluate the support vector regression model's mean square error and mean absolute error on the test set; For new carbon emission data, we first use the feature extraction module to extract its low-dimensional semantic features, and then input the extracted features into the carbon emission intensity classification model to determine the carbon emission intensity level to which it belongs; Finally, the carbon emission feature vector is input into the carbon emission prediction model to obtain the predicted value of carbon emissions in the future period; Comprehensive carbon emission intensity levels and carbon emission forecast results provide data support for formulating carbon emission reduction strategies.

4. The method according to claim 1, characterized in that: According to the timestamp of carbon emission data, the sliding time window method is used to segment the continuous time series carbon emission data, and the long short-term memory network is used to model the changes in carbon emission patterns in different time periods, and the nonlinear relationship between carbon emissions and time is mined to obtain the carbon emission rules in fine-grained time periods, including: Obtaining raw carbon emission data from the distribution network carbon emission monitoring system; The box plot method is used to detect outliers. If the value exceeds the upper and lower limits of the box plot, it is considered an outlier and deleted; For missing values, if the missing time period is short, linear interpolation is used to fill it; After completing outlier processing and missing value filling, a normalized carbon emission time series data set is obtained; According to the timestamp of carbon emission data, the time window length is set to one hour, the sliding step is set to ten minutes, and the continuous time series data is divided into several carbon emission subsequences using the sliding time window method; For each carbon emission subsequence, the data mean, variance, maximum and minimum values ​​are calculated as statistical features; At the same time, the discrete Fourier transform coefficients are calculated, and the first five low-frequency coefficients are selected as frequency domain features; Combine the statistical features with the frequency domain features to form the feature vector of the carbon emission subsequence; The K-means clustering algorithm is used to cluster the feature vectors of carbon emission subsequences; Calculate the silhouette coefficient under different cluster numbers, and select the cluster number corresponding to the maximum silhouette coefficient as the optimal cluster number; According to the value of the cluster center, sort from small to large, assign labels "low emission mode", "medium emission mode" and "high emission mode" respectively, and associate the cluster result labels with the feature vectors of the carbon emission subsequence; Construct a long short-term memory neural network model; Using the feature vector of the carbon emission subsequence as input and the corresponding cluster label as output, the long short-term memory neural network model is trained to predict the carbon emission pattern based on the feature vector; Add a self-attention mechanism layer before the output layer of the long short-term memory neural network model; The self-attention mechanism layer calculates the attention weight of each time point according to the hidden state of the long short-term memory neural network model, and applies the weight to the hidden state to highlight the information of the key time points; Use the trained long short-term memory neural network model combined with the self-attention mechanism model to process the newly collected carbon emission data; Divide the new data into subsequences of the same length as the training data, extract the statistical features and frequency domain features of each subsequence, and input them into the model to obtain the carbon emission pattern prediction results of each subsequence; Count the carbon emission pattern prediction results of all subsequences; If the number of subsequences predicted as "high emission mode" exceeds the preset threshold, it is determined that the overall carbon emissions in the current period may exceed the standard, triggering the early warning mechanism and generating a carbon emissions analysis report; Regularly collect updated carbon emission data from all links of the distribution network; The feature vectors are recalculated using the new data, and the long short-term memory neural network model combined with the self-attention mechanism is retrained using the new data and the corresponding carbon emission pattern labels to update the model parameters.

5. The method according to claim 1, characterized in that The power load data collected by the smart meter is used to extract the power curve features of different types of power equipment using a convolutional neural network, identify various types of power equipment through a softmax classifier, and calculate the carbon emission contribution of each device in combination with the rated power and usage time of the device, and establish a carbon emission decomposition model for the device, including: Use smart meters to collect electricity load data from users in all links of the distribution network; Load data includes voltage, current, active power and reactive power information; The frequency of data collection is once every five minutes; Transmit the collected power load data to the data processing center via a wired or wireless network; According to the power load data cleaning strategy, incomplete and obviously abnormal data points are filtered out; If there are missing values ​​in the data, use the nearest neighbor interpolation method to fill them; Complete data cleaning and completion; Applying a time series based one-dimensional convolution filter on the cleaned electricity load data; The filter is set to three layers, the filter window size of each layer is three, the step size is one, and the ReLU activation function is used; After each layer of filter processing, thirty-two, sixty-four, and one hundred and twenty-eight feature data are output respectively; The 128 feature data output by the last layer of filters are used as input features of the logistic regression classifier; Use logistic regression classifier to classify the processed power load data; The classification objects cover various types of electrical equipment such as air conditioners, lighting and electric motors; The logistic regression classifier identifies the power curve characteristics of different types of electrical equipment based on a pre-set classification threshold; Obtain identification results of various types of electrical equipment; Obtain the rated power and usage time of various electrical equipment from the equipment information database; Use the equipment carbon emission calculation formula, combined with the obtained rated power and usage time and the preset equipment carbon emission factor to calculate the carbon emission contribution of each electrical equipment; Based on the calculated carbon emission contribution values ​​of each device, a linear fitting model for the relationship between the carbon emission contribution of the device and the device type is constructed; Fitting was performed using a linear fitting tool based on the least squares method; After obtaining the linear fitting model, the real-time power load data is passed through a one-dimensional convolution filter based on time series to extract features; The extracted features are used as inputs to the linear fitting model to calculate the real-time carbon emission contribution of each device; Utilize the real-time carbon emission contribution data of each device to develop carbon emission reduction optimization strategies; The strategies include adjusting the equipment operation status according to the equipment's real-time carbon emission contribution, recommending the replacement of high-energy-consuming equipment based on the linear relationship between the equipment's carbon emission contribution and the equipment type, and guiding users to optimize their electricity usage habits based on their electricity usage behavior patterns; The ultimate goal is to reduce the overall carbon emissions of the distribution grid.

6. The method according to claim 1, characterized in that The user electricity consumption behavior portrait is constructed based on the user electricity consumption data, the influence weight of the user's electricity consumption habits on carbon emissions is learned by using the attention mechanism, and the carbon emission patterns under different electricity consumption behaviors are mined by the clustering algorithm to form a user-oriented carbon emission explanation model, including: The modified optimized version is as follows; Obtain electricity usage time and electricity consumption data of typical users from the customer information system; The K-means clustering algorithm is used to cluster the user electricity consumption behavior data, dividing the users into peak electricity users and off-peak electricity users to form a user electricity consumption portrait; The user electricity usage profile includes the user's electricity usage behavior characteristics; Acquire historical operating data from distribution network databases and sensors; Historical operating data includes voltage, current, power, temperature, load type, and air temperature and humidity; forming a first data set; Cleaning the first data set; Lagrange interpolation method is used to handle missing values; The 3σ criterion was used to eliminate outliers; forming a second data set; The principal component analysis method is used to extract the carbon emission characteristic quantity of the second data set; Determine the carbon emission feature set based on the principal components with cumulative contribution rates greater than 80%; forming a third data set; StandardScaler was used to standardize the third dataset; The standardization process uses Z-score standardization; The calculation formula is: z = (x-μ) / σ; Where x is the original data, μ is the mean, and σ is the standard deviation; forming a fourth data set; Use XGBoost tool to build carbon emission prediction model; Apply grid search and 5-fold cross validation to optimize the hyperparameters of the XGBoost tool; Root mean square error was used as the evaluation metric; Obtaining an optimized carbon emission prediction model; Apply the optimized carbon emission prediction model to the real-time operation data of the distribution network; Obtain the predicted value of carbon emissions in the future; Linear regression method is used to analyze the power load characteristics and carbon emission intensity distribution of the distribution network; Identify the key factors affecting carbon emissions; The key influencing factors of carbon emissions and the characteristics of users’ electricity consumption behavior are used as input; The user electricity usage behavior characteristics use the user electricity usage behavior characteristics in the user electricity usage portrait obtained in step 1; Apply the attention mechanism to learn the impact weight of users’ electricity usage habits on carbon emissions; Using impact weights, the K-means clustering algorithm is used to mine carbon emission patterns under different electricity consumption behaviors; Form a user-oriented carbon emission explanation model; Correlate user electricity consumption profiles with carbon emissions and energy consumption forecasts; Formulate peak electricity consumption strategies for peak electricity users; The peak-shifting electricity consumption strategy guides electricity consumption during the low-load period of the power grid; Formulate energy efficiency management strategies for high energy consumption households; Energy efficiency management strategies guide the adoption of energy-saving equipment and technologies; The power communication network sends energy-saving and emission-reduction strategies to smart meters; Smart meters optimize and control users’ electricity usage behavior according to policy requirements; Optimal control is to limit non-essential loads during peak hours; Increase energy storage charging during off-peak hours; Smart meters collect users’ actual electricity usage data; Evaluate the effectiveness of energy conservation and emission reduction strategies; If the actual power saving does not reach the preset threshold; Adjust peak hours; Improving energy efficiency standards; Use the latest carbon emission data, electricity data and operating parameters to regularly retrain and optimize the parameters of the carbon emission prediction model built by the XGBoost tool.

7. The method according to claim 1, characterized in that The Apriori association rule mining algorithm is used to extract frequent patterns in carbon emission data from three dimensions: time period, equipment, and user, and to screen out key factor combinations that have a significant impact on carbon emissions. The rules are embedded in the blockchain through smart contracts to perform carbon emission traceability, including: According to the IoT sensors and smart meters deployed in each link of the distribution network, carbon emission data and electricity data are collected in real time to obtain detailed operating parameters and energy consumption information; According to the pre-established data cleaning rules, regular expressions are used to clean non-numeric data and special characters in the collected data set; In the cleaned data set, if there are missing values, the Lagrange interpolation method is used to complete the data; Using the 3σ criterion, outliers are removed from the completed data set; Standardize the data set after removing outliers to eliminate the impact of data dimension; Using the K-means algorithm, cluster analysis is performed based on the electricity consumption time characteristics and electricity consumption dimensions in the data. The results are used to divide multiple typical user portraits such as industrial users, commercial users, and residential users. For the carbon emission data, electricity data and operating parameters obtained after cleaning, use the pivot table tool to collect data statistics according to time period, equipment and user dimensions to obtain transaction data; Set the minimum support and minimum confidence, apply the association rule algorithm, and mine frequent item sets and association rules from transactional data; According to the obtained association rules, determine the key factor combination that significantly affects carbon emissions; Build a blockchain network on the Ethereum platform to store carbon emission data, electricity data and mined association rules, and achieve distributed data storage and non-tamperability through smart contracts; In the blockchain network, the carbon emission intensity of each block is determined based on carbon emission data and electricity data, and the carbon emission intensity is divided into three levels: high, medium and low; A simple moving average algorithm is used to determine the preset threshold of carbon emission intensity; If the carbon emission intensity of a block exceeds the preset threshold, the corresponding responsible entity will be judged to bear a greater carbon emission reduction responsibility; Based on predefined rules, trace and audit the carbon emission data and electricity data on the chain; During the traceability audit process, verify the authenticity of the data source and check the integrity and consistency of the data; Smart contracts are used to write the results of data traceability audits into the blockchain to ensure that the audit results cannot be tampered with.

8. The method according to claim 1, characterized in that The carbon emission influencing factors such as time period, equipment, user, and rules are used as input features, carbon emissions are used as target variables, and the LightGBM regression algorithm is used to establish a carbon emission prediction model; the pre-trained carbon emission model is applied to different regions and scenarios using the transfer learning method to improve the model generalization performance, including: Collect data on carbon emission influencing factors from all aspects of the distribution network, including time period, equipment parameters and user electricity consumption behavior; Clean and standardize the data to obtain a standardized carbon emission characteristic data set; The Pearson correlation coefficient is used to calculate the correlation coefficient between each feature in the carbon emission feature data set and the carbon emissions, and the feature subset that is highly correlated with carbon emissions is screened out based on the absolute value of the correlation coefficient; The input feature vector of the carbon emission prediction model is constructed using the filtered feature subset; Taking carbon emissions as the target variable and the key feature subset as the input features, the LightGBM regression algorithm is used to train the carbon emissions prediction model; The grid search method is used to find the optimal hyperparameter combination in the preset parameter space, and the optimal parameters of the LightGBM regression algorithm are determined with the goal of minimizing the mean square error. The cross-validation method is used to evaluate the generalization performance of the carbon emission prediction model; If the mean square error of the model in cross validation does not reach the preset threshold, adjust the screening threshold of the Pearson correlation coefficient, reselect the feature subset, and re-search the optimal hyperparameters of the LightGBM regression algorithm in the preset parameter space; Repeat this process until the model performance meets the preset mean square error threshold requirement; For distribution networks in different regions and scenarios, obtain data on carbon emission influencing factors in their source, grid, load and storage links, apply the same cleaning and standardization methods as in the training phase, and map the data to a feature space consistent with the training data set; Use the trained LightGBM regression algorithm model to predict the data of new regions and scenarios, and output the predicted carbon emissions of each link; Fix all layer parameters except the output layer in the trained model, and only use the data of the new area and scene to update the weight of the output layer to obtain the fine-tuned LightGBM regression algorithm model; Use the fine-tuned model to re-predict data for new regions and scenarios to obtain updated carbon emissions forecasts; Obtain carbon emission data output by the carbon emission source tracing model based on the emission factor method; Calculate the absolute error between the carbon emissions predicted by the LightGBM regression algorithm model and the carbon emissions output by the emission factor method model; If the absolute error exceeds the preset threshold, the weight parameters of the output layer in the LightGBM regression algorithm model are adjusted, and the weights of related features are adjusted according to the results of the emission factor method model; Re-predict using the adjusted LightGBM regression algorithm model; Based on the prediction results of the LightGBM regression algorithm model, an accurate assessment report on carbon emissions in each link of the distribution network is generated, providing data support and decision-making basis for the formulation of carbon emission reduction strategies for the distribution network.

9. The method according to claim 1, characterized in that: The model integration strategy is adopted to weightedly integrate the time period model, equipment decomposition model, user interpretation model and rule mining model to form a fine-grained and multi-dimensional comprehensive carbon emission analysis model; combined with the reinforcement learning algorithm to optimize the carbon emission control strategy, to achieve coordinated emission reduction in the source, grid, load and storage links, including: S1: Obtain carbon emission data, electricity data and operating parameters of each link of the distribution network source, grid, load and storage, clean and normalize the data, and store the processed data in a unified format in the database; S2: Based on the time characteristics, the K-means clustering algorithm is used to perform cluster analysis on the carbon emission data and build carbon emission models for different periods; According to the equipment type and parameters, the decision tree algorithm is used to classify the carbon emission data and build an equipment carbon emission decomposition model; According to the user attribute characteristics, the FP-Growth algorithm is used to mine the frequent patterns of user carbon emission behavior and build a user carbon emission explanation model; S3: Use the Apriori algorithm to mine association rules between the pre-processed carbon emission data, power data and operating parameters, discover the association rules between the data, such as "the carbon emission of equipment type A is higher than that of equipment type B", etc., determine the key factors affecting carbon emissions, and build a carbon emission rule mining model; S4: The output results of the time period model, equipment decomposition model, user interpretation model and rule mining model are integrated using the weighted average method to obtain a comprehensive carbon emission analysis model; Through grid search, the weights of each model are traversed, and the model performance of each group of weights is calculated using cross-validation. The weight combination with the best performance is selected to improve the accuracy and generalization ability of the model. S5: Carry out data fusion of carbon emission data, electricity data and operation parameters, take carbon emission influencing factors as input features and carbon emission as output indicators, and build a long short-term memory neural network model; The network weights were optimized using the gradient descent method, and the prediction error was minimized through the back propagation algorithm to train a carbon emission prediction model. S6: Combine the deep reinforcement learning algorithm with the long short-term memory neural network model, use the grid operation state as the reinforcement learning environment state, the optimization control instructions of the distribution equipment as the reinforcement learning action, and the reduction of carbon emissions as the reward function, and generate the optimal carbon emission control strategy through continuous interactive learning with the environment; S7: Convert the optimal carbon emission control strategy into specific control instructions for distribution equipment, including adjusting transformer taps, optimizing reactive compensation capacitor switching and balancing load distribution, etc., and send it to the distribution automation terminal RTU through the power communication network. The RTU optimizes and controls the distribution equipment according to the instructions. S8: Collect carbon emission data before and after optimization and control through smart meters and environmental sensors on power distribution equipment, calculate the change in carbon emission intensity per unit of electricity, and evaluate the emission reduction effect of carbon emission control strategies; If the emission reduction effect does not reach the expected target, return to step S6 and iterate the strategy optimization again; S9: Continue to collect the latest carbon emission data and operation data of the source, grid, load and storage links, and regularly use the new data to retrain and optimize the comprehensive carbon emission analysis model and management strategy; In combination with the latest carbon emission policies and carbon trading market prices, the distribution network emission reduction targets and reward and punishment mechanisms are dynamically adjusted to achieve adaptive closed-loop optimization of distribution network carbon emission analysis and management, supporting the economical and efficient operation and low-carbon sustainable development of the distribution network.

10. The method according to claim 1, characterized in that The generative adversarial network is used to expand the diversity of carbon emission data and simulate carbon emission conditions under different extreme scenarios; the blockchain browser is used to realize the visual traceability of the entire process of carbon emission data, providing a decision-making basis for accurately calculating carbon emissions and formulating emission reduction policies, including: The modified optimized version is as follows; Obtain distribution network operation data from the distribution network operation management system, and obtain historical carbon emission data from the historical database; After obtaining the historical carbon emission data, the Kolmogorov-Smirnov test is used to determine whether the historical carbon emission data conforms to the normal distribution; If the Kolmogorov-Smirnov test results show that the historical carbon emission data do not conform to the normal distribution, the data distribution shape is adjusted using logarithmic transformation to make it fit the normal distribution; Use the TensorFlow library to implement a generative adversarial network to generate simulated data that matches the actual carbon emission data distribution and expand the diversity of carbon emission data; The simulated data is integrated into the original data set to form an extended data set, the mean of the extended data set is calculated using the mean function, and the variance of the extended data set is calculated using the variance function; According to the mean and variance results of the expanded data set, the temperature and humidity data are obtained from the meteorological database as environmental factor data; For the expanded data set, the extreme high temperature scene data set is screened out based on the records with temperatures higher than the preset high temperature threshold; Using the ARIMA model, we conduct time series analysis using the extreme high temperature scenario dataset to predict the carbon emission trend of the distribution network under extreme high temperature scenarios; Use the Hyperledger Fabric platform to deploy a blockchain browser, and present the carbon emission data of each link of the distribution network and the predicted trend obtained by the ARIMA model in the blockchain browser; The blockchain browser uses the data traceability function of Hyperledger Fabric to trace the source of carbon emission data and its change trajectory, and realize the transparent display of carbon emission data correlation information and the logical relationship between data.

Images