A method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks

Abstract

The present invention discloses a method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks, and relates to the technical field of carbon emission monitoring. The method is aimed at the required direct carbon emissions of enterprise users in high-energy-consuming industrial parks. On the one hand, it uses the wide coverage and real-time nature of electricity and heat data to construct and train a multi-source feature and direct carbon emission mapping model for direct carbon emission calculation. On the other hand, in order to solve the data privacy security issues of the park, park enterprise users and electric power and heat companies, an information interaction method based on the PP3-FL framework and a mapping model construction method based on the FedPSW strategy are introduced. The high-frequency, accurate and efficient calculation of carbon emissions of enterprise users in the park can be achieved under the condition that local data and information are available and invisible, thereby solving the problems of imperfect accounting, limited accuracy of calculation results and failure to consider user privacy in model training and application in existing carbon emission calculation schemes.

Description

Technical Field

[0001] The present invention belongs to the technical field of carbon emission monitoring, and in particular relates to a method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks. Background Art

[0002] Since the Industrial Revolution, global carbon emissions have increased dramatically and climate warming has continued to accelerate. Controlling carbon dioxide emissions to cope with climate change has become an important issue that all mankind cannot avoid. Industry is the main source of carbon dioxide emissions, accounting for about 70% of carbon dioxide emissions, so the potential for carbon emission reduction lies mainly in industry. In recent years, studying the low-carbon development of high-energy-consuming enterprises (also known as high-energy-consuming enterprises, which refer to enterprises that consume natural resources such as electricity, water, natural gas, and materials at a fast rate and have a large demand for resources compared to other enterprises) has become an important direction of industrial development. Carbon emission characteristic analysis technology is the basis for the development of a low-carbon economy. Establishing a simple, accurate and effective carbon emission measurement technology suitable for users of high-energy-consuming industrial parks is of great significance to guiding the low-carbon development of industry and even the entire energy system.

[0003] To promote the implementation of carbon emission accounting, international organizations and countries around the world have introduced relevant carbon emission measurement scope standards. At present, the "Greenhouse Gas Protocol" proposed by the World Business Council for Sustainable Development and the World Resources Institute is a widely recognized international standard for carbon emission measurement. They divide the user carbon emission accounting boundaries into the following three scopes: (1) Direct carbon emissions (i.e., scope 1), which refers to carbon dioxide emissions directly generated by users or organizations, including carbon dioxide generated by the combustion of fossil fuels and carbon emissions caused by the use of carbon-containing materials in industrial production processes, that is, direct emissions refer to emission sources directly controlled by users; (2) Indirect carbon emissions (i.e., scope 2), which refers to greenhouse gas emissions indirectly generated by users or organizations, mainly from emissions in the process of purchasing electricity or heat energy, including carbon emissions generated by part or all of the electricity purchased from power companies; 3) Other indirect carbon emissions (i.e., scope 3), which refers to all other indirect carbon emissions of users or organizations other than direct and indirect carbon emissions; the aforementioned scope 3 carbon emissions are usually generated by other links in the user's value chain, including supply chain, transportation and product life cycle. In general, high energy consuming corporate users are widely required to disclose their scope 1 and scope 2 carbon emissions.

[0004] In the current carbon emission accounting process, the most prominent issues are as follows:

[0005] First, carbon emission accounting is not sound. The existing carbon emission accounting results are generally delayed by 12-18 months, and have shortcomings such as high accounting costs, long cycles and coarse granularity, which cannot meet the needs of relevant policy planning or refined operation of users' carbon emission assets.

[0006] Second, for direct carbon emissions, some domestic and foreign scholars have conducted research on using power big data to build an "electricity-carbon" model to estimate carbon emissions. Although this can effectively solve the problems of low efficiency and timeliness of data collection, its estimation results are uncertain and have poor interpretability. That is, the current research problem is that, on the one hand, existing studies mostly explore the relationship between total electricity consumption and carbon dioxide emissions, and it is difficult to accurately reflect the complex correlation characteristics between electricity and carbon. On the other hand, the machine learning algorithms such as support vector machines used in these studies often have few parameters and simple structures, making it difficult to fully analyze the complex relationship between electricity consumption and carbon dioxide emissions (considering the complexity of this relationship, a more complex deep learning model is needed to complete the measurement task).

[0007] Third, for indirect carbon emissions, the electricity / heat carbon emission factor method is usually used for calculation, but the current calculation of these carbon emission factors does not take into account the electricity / heat transfer between the target park’s region (such as national or provincial regions) and other regions;

[0008] Fourth, the training and application of the "electricity-carbon" model in the existing carbon emission accounting process does not take into account user privacy. Summary of the invention

[0009] The purpose of the present invention is to provide a method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks, so as to solve the problems of incomplete accounting, limited accuracy of calculation results and failure to consider user privacy during model training and application in existing carbon emissions calculation schemes for enterprise users in high-energy-consuming industrial parks.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] In a first aspect, a method for calculating carbon emissions of enterprise users in a high-energy-consuming industrial park is provided, comprising the following steps S1 to S11 executed in sequence:

[0012] S1. The industrial park central server sends a model building request to the terminal devices of each enterprise user located in the high-energy-consuming industrial park;

[0013] S2. After receiving the model building request, the terminal device of each enterprise user uses the first public key in the first key pair generated in advance using homomorphic encryption technology to encrypt the local device information and the total amount of local available data, obtain the first ciphertext and upload it to the industrial park central server;

[0014] S3. The industrial park central server uses the first private key in the first key pair to decrypt the received first ciphertext, obtain the terminal device information of each enterprise user and the total amount of available data, and determine the parameter configuration content of the multi-source feature and direct carbon emissions mapping model based on the terminal device information of each enterprise user and the total amount of available data, and then extract the model construction related information from the parameter configuration content and distribute it to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user can complete the preliminary construction of the local multi-source feature and direct carbon emissions mapping model based on the model construction related information;

[0015] S4. The industrial park central server randomly initializes the model parameters of the multi-source characteristics and direct carbon emissions mapping model and distributes them to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user complete the initialization of the local multi-source characteristics and direct carbon emissions mapping model according to the model parameters;

[0016] S5. The industrial park central server randomly selects K enterprise users located in the high-energy-consuming industrial park, and sends a model training request to the terminal device of each selected user among the K enterprise users, where K represents a positive integer less than or equal to M, and M represents the total number of enterprise users located in the high-energy-consuming industrial park;

[0017] S6. After receiving the model training request, the terminal device of each selected user applies the local multi-source feature time series data and carbon emission time series data to train the local multi-source feature and direct carbon emission mapping model, and in the process of model iteration and update, the local model parameters are encrypted and updated using the federal proximal weight score FedPSW local strategy and differential privacy technology, and then at the end of the model iteration and update, the final parameters of the local model are uploaded to the industrial park central server, wherein the multi-source feature time series data contains the power data and thermal data consumed in each historical unit time period in multiple continuous historical unit time periods, and the carbon emission time series data contains the direct carbon emissions generated in each historical unit time period;

[0018] S7. The industrial park central server adopts the FedPSW global strategy of the federated proximal weight score to aggregate the final model parameters from the terminal devices of the selected users, form the global parameters of the multi-source feature and direct carbon emissions mapping model and distribute them to the terminal devices of the enterprise users, so that the terminal devices of the enterprise users can complete the update of the local multi-source feature and direct carbon emissions mapping model according to the global parameters;

[0019] S8. The industrial park central server determines whether the preset first training end condition is currently met. If so, the multi-source feature and direct carbon emission mapping model is updated according to the global parameters, otherwise, it returns to step S5;

[0020] S9. The industrial park central server sends a carbon emission accounting request to a terminal device of an energy supply company that supplies energy to a target enterprise user located in the high-energy-consuming industrial park, wherein the energy supply refers to electricity and heat supply, and the energy supply company refers to an electric power company and a heat company;

[0021] S10. After receiving the carbon emission accounting request, the terminal device of the energy supply company uses the second public key in the second key pair generated in advance by the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user, obtains the second ciphertext and transmits it back to the central server of the industrial park, wherein the latest energy purchase time series data contains the power data and heat data purchased in each unit time period in at least one recent unit time period;

[0022] S11. The industrial park central server uses the second private key in the second key pair to decrypt the received second ciphertext, obtains the current latest energy purchase time series data of the target enterprise user, imports the multi-source feature and direct carbon emissions mapping model, and outputs the direct carbon emissions of the target enterprise user in the most recent unit time period.

[0023] Based on the above invention content, a new scheme for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks that takes into account the privacy security of multiple parties is provided. That is, for the required direct carbon emissions of enterprise users in high-energy-consuming industrial parks, on the one hand, a multi-source feature and direct carbon emissions mapping model is constructed and trained to perform direct carbon emissions calculation by utilizing the wide coverage and real-time nature of electricity and heat data. On the other hand, in order to solve the data privacy security issues of the park, park enterprise users and power and heat companies, an information interaction method based on the PP3-FL framework (i.e., the application framework of the first key pair, differential privacy technology and the second key pair) and a mapping model construction method based on the FedPSW strategy are introduced. High-frequency, accurate and efficient calculation of carbon emissions of enterprise users in the park can be achieved under the condition that local data and information realization data are available but invisible, thereby solving the problems of imperfect accounting, limited accuracy of calculation results and failure to consider user privacy during model training and application in existing carbon emission calculation schemes for enterprise users in high-energy-consuming industrial parks, which is convenient for practical application and promotion.

[0024] In a possible design, the latest energy purchase time series data of the target enterprise user is encrypted using the second public key in the second key pair generated in advance using the homomorphic encryption technology to obtain a second ciphertext, including:

[0025] According to the electricity and heat conditions of each region in the most recent unit time period, determine the latest electricity dynamic carbon emission factor and the latest heat dynamic carbon emission factor for each region taking into account cross-regional flows;

[0026] According to the electricity data p purchased by the target enterprise user in the recent unit period T+1 The latest dynamic carbon emission factor DFE of electricity in the region where the target enterprise user belongs and taking into account cross-regional flows P,T+1 , the carbon emissions CE of the purchased electricity of a target enterprise user in the most recent unit period is calculated according to the following formula: indirect,P,T+1 =DFE P,T+1 ×p T+1 ;

[0027] According to the heat data h purchased by a certain target enterprise user in the recent unit period T+1 The latest thermal dynamic carbon emission factor DFE of the region to which the target enterprise user belongs and considering cross-regional flows H,T+1 , the carbon emission CE of purchased heat power of a target enterprise user in the most recent unit period is calculated according to the following formula: indirect,H,T+1 =DFE H,T+1 ×h T+1 ;

[0028] Use the second public key in the second key pair generated in advance by the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user and the carbon emissions of purchased electricity and purchased heat of the target enterprise user in the most recent unit period to obtain a second ciphertext, wherein the latest energy purchase time series data contains the electricity data and heat data purchased in each unit period in at least one most recent unit period;

[0029] After step S10, the method further includes:

[0030] The industrial park central server adds the ciphertext data of carbon emissions from purchased electricity and the ciphertext data of carbon emissions from purchased heat in the second ciphertext, and uses the second private key to decrypt the addition result to obtain the indirect carbon emissions of the target enterprise user in the most recent unit period, and then adds the direct carbon emissions and indirect carbon emissions of the target enterprise user in the most recent unit period to obtain the total carbon emissions of the target enterprise user in the most recent unit period.

[0031] In a possible design, the multi-source feature and direct carbon emissions mapping model is constructed using a variational mode decomposition algorithm, a time-domain convolutional network, and a multi-head attention long short-term memory network algorithm.

[0032] In a possible design, the local multi-source feature time series data and the carbon emission time series data are used to train the local multi-source feature and direct carbon emission mapping model, including the following steps S61-S66:

[0033] S61. Acquire local multi-source characteristic time series data and carbon emission time series data, and then execute step S62, wherein the multi-source characteristic time series data contains power data and heat data consumed in each historical unit period in multiple continuous historical unit periods, and the carbon emission time series data contains direct carbon emissions generated in each historical unit period;

[0034] S62. Using variational mode decomposition algorithm to analyze the power time series data P in the multi-source characteristic time series data = {p1, p2, ..., p t ,…,p T} and thermal time series data H = {h1,h2,…,h t ,…,h T} is decomposed to obtain the multi-source feature sequence {imf P,1 ,imf P,2 ,···,imf P,t ,···,imf P,T ,imf H,1 ,imf H,2 ,···,imf H,t ,···,imf H,T}, then execute step S63, wherein T represents the total number of time periods of the plurality of consecutive historical unit time periods, t represents a positive integer less than or equal to T, and p t represents the power data consumed in the tth historical unit period among the plurality of continuous historical unit periods, h t represents the heat data consumed in the t-th historical unit period, imf P,t Indicates that p t The corresponding modal components, imf H,t Indicates that t The corresponding modal components;

[0035] S63. Based on the multi-source feature sequence, a feature matrix X is formed. t′-1 ,X t′ ,X t′+1 ,X t′-2 ,…,X T}, then execute step S64, where X t′ ={imf P,1 ,imf P,2 ,···,imf P,t′ ,···,imf H,1 ,imf H,2 ,···,imf H,t′}, t′ represents a positive integer less than or equal to T;

[0036] S64. X t′-1 ={imf P,1 ,imf P,2 ,···,imf P,t′-1 ,···,imf H,1 ,imf H,2 ,···,imf H,t′-1} is input into the time domain convolutional network to obtain the modal feature sequence {I1,I2,···,I t′-1}, then execute step S65;

[0037] S65. The modal feature sequence {I1, I2, ···, I t′-1} as a model input item, and the direct carbon emissions generated in the t′-1th historical unit period in the carbon emission time series data and in the multiple continuous historical unit periods as a model output item, train the artificial intelligence model based on the multi-head attention long short-term memory network algorithm, and back-propagate the loss error obtained from the training, and then execute step S66;

[0038] S66. Determine whether the preset second training end condition is currently met. If so, end the training. Otherwise, increment t′ by 1, and then return to step S64.

[0039] In a possible design, during the iterative update process of the model, the local model parameters are encrypted and updated using the FedPSW local strategy and differential privacy technology, including:

[0040] During the iterative update process of the model, the loss error obtained from the training of the mapping model is back-propagated and the parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network are updated according to the following formula:

[0041]

[0042] In the formula, s represents the number of global training rounds currently completed, and w s+1represents the updated parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network in the sth round of model training, η represents the learning rate, ▽loss represents the gradient of the loss error, μ′ represents the proximal term parameter, and w g,s represents the global parameters of the sth round of model training, w s represents the pre-updated parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network in the sth round of model training, represents the encrypted noise function of the differential privacy technique, represents the truth query function, represents the local multi-source characteristics and direct carbon emissions mapping model, represents the standard deviation and ∈ represents the privacy budget, δ represents the model change restriction parameter, e represents the base of the natural logarithm, S represents the real number domain, Pr[] represents the probability function, X t ″ -1 Represents X t′-1 adjacent datasets.

[0043] In a possible design, a FedPSW global strategy is adopted to aggregate the final model parameters from the terminal devices of the selected users to form the global parameters of the multi-source feature and direct carbon emissions mapping model, including:

[0044] According to the final model parameters from the terminal devices of the selected users, the global parameter w of the multi-source feature and direct carbon emission mapping model is formed according to the following formula: g,s+1 :

[0045]

[0046] In the formula, k is a positive integer less than or equal to K, Per k represents the local model performance score corresponding to the k-th selected user among the K enterprise users, D(k) represents the weight score of the k-th selected user in the global model aggregation stage, and w k,s represents the final model parameters from the terminal device of the k-th selected user, s represents the number of global training rounds currently completed, represents a positive integer less than or equal to s, Indicates the user corresponding to the kth selected user and in The final loss error of the model after the round of global training is completed.

[0047] Beneficial effects of the above scheme:

[0048] (1) The present invention creatively provides a new scheme for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks that takes into account the privacy security of multiple parties. That is, for the required direct carbon emissions of enterprise users in high-energy-consuming industrial parks, on the one hand, a multi-source feature and direct carbon emission mapping model is constructed and trained to perform direct carbon emission calculation by utilizing the wide coverage and real-time nature of electricity and heat data. On the other hand, in order to solve the data privacy security issues of the park, the enterprise users in the park, and the electric power and heat company, an information interaction method based on the PP3-FL framework (i.e., the application framework of the first key pair, differential privacy technology, and the second key pair) and a mapping model construction method based on the FedPSW strategy are introduced. The high-frequency, accurate, and efficient calculation of carbon emissions of enterprise users in the park can be achieved under the condition that local data and information realization data are available but invisible. This can solve the problems of imperfect accounting, limited accuracy of calculation results, and failure to consider user privacy during model training and application in existing carbon emission calculation schemes for enterprise users in high-energy-consuming industrial parks, and facilitate practical application and promotion.

[0049] (2) In order to enhance the ability to mine data information, the VMD-TCN-LSTMMA algorithm is also used in the mapping model to decompose and upgrade the dimensions and extract features of power and thermal time series data, thereby enhancing the learning ability to extract key features from time series data and establish correlation analysis with the output sequence;

[0050] (3) For the indirect carbon emissions of enterprise users in industrial parks, a method for calculating the dynamic carbon emission factors of heat and electricity based on cross-domain carbon flows is also proposed. This method can then achieve the purpose of dynamically and accurately calculating the carbon emission factors of regional heat and electricity while taking into account the carbon emissions of electricity and heat transferred between the target park and the national / provincial region, thereby further improving the accuracy of the total carbon emissions calculation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 A schematic flow chart of a method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks provided in an embodiment of the present invention.

[0053] Figure 2 A schematic diagram of the carbon emission accounting boundary for enterprise users in high-energy-consuming industrial parks provided in an embodiment of the present invention.

[0054] Figure 3A schematic diagram of the network architecture of a TCN provided in an embodiment of the present invention.

[0055] Figure 4 A schematic diagram of the network architecture of LSTMMA provided in an embodiment of the present invention.

[0056] Figure 5 An example diagram comparing the evaluation indicators of the VMD-TCN-LSTMMA model provided in the embodiment of the present invention and the other four benchmark models on the hourly carbon emission measurement results of actual steel enterprise users, wherein: Figure 5 (a) shows an example of the RMSE comparison between the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission measurement results of actual steel enterprise users. Figure 5 (b) shows an example of the MAE comparison between the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual steel enterprise users. Figure 5 (c) shows an example of the comparison of the MAPE index of the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual steel enterprise users. Figure 5 (d) shows the R of the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual steel enterprise users. 2 Example chart of indicator comparison.

[0057] Figure 6 An example diagram comparing the evaluation indicators of the VMD-TCN-LSTMMA model provided in the embodiment of the present invention and the other four benchmark models on the hourly carbon emission measurement results of actual petrochemical enterprise users, wherein: Figure 6 (a) shows an example of the RMSE comparison between the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission measurement results of actual petrochemical enterprise users. Figure 6 (b) shows an example of the MAE comparison between the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual petrochemical enterprise users. Figure 6 (c) shows an example of the comparison of the MAPE index of the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual petrochemical enterprise users. Figure 6 (d) shows the R of the VMD-TCN-LSTMMA model and the other four benchmark models in the hourly carbon emission calculation results of actual petrochemical enterprise users. 2 Example chart of indicator comparison.

[0058] Figure 7 The performance analysis diagram of the improved federated computing framework provided in the embodiment of the present invention and the calculation results of the hourly carbon emissions of actual steel enterprise users without using the federated security computing framework is shown in the figure, wherein: Figure 7 (a) shows the performance analysis diagram of the improved federated computing framework and the non-federated secure computing framework when the number of local clients γ=3 and the hourly carbon emission calculation results of actual steel enterprise users. Figure 7 (b) shows a performance analysis diagram of the improved federated computing framework and the one without the federated secure computing framework when the number of local clients γ=2 and on the hourly carbon emissions calculation results of actual steel enterprise users.

[0059] Figure 8 The performance analysis diagram of the improved federated computing framework provided in the embodiment of the present invention and the computing framework without using the federated security computing framework on the hourly carbon emission calculation results of actual petrochemical enterprise users is shown in the figure, wherein: Figure 8 (a) shows the performance analysis diagram of the improved federated computing framework and the non-federated secure computing framework when the number of local clients γ=5 and the hourly carbon emission calculation results of actual petrochemical enterprise users. Figure 8 (b) shows a performance analysis diagram of the improved federated computing framework and the one without the federated secure computing framework when the number of local clients γ=3 and on the hourly carbon emission calculation results of actual petrochemical enterprise users. DETAILED DESCRIPTION

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structures of the drawings is only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these embodiments without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0061] It should be understood that although the terms first and second, etc. may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another object. For example, a first object can be referred to as a second object, and similarly, a second object can be referred to as a first object without departing from the scope of the exemplary embodiments of the present invention.

[0062] It should be understood that the term "and / or" that may appear in this document is merely a description of the association relationship between associated objects, indicating that there may be three relationships. For example, A and / or B can indicate three situations: A exists alone, B exists alone, or A and B exist at the same time. For another example, A, B and / or C can indicate the existence of any one of A, B and C or any combination of them. The term " / and" that may appear in this document describes another type of association object relationship, indicating that there may be two relationships. For example, A / and B can indicate two situations: A exists alone or A and B exist at the same time. In addition, the character " / " that may appear in this document generally indicates that the previous and next associated objects are in an "or" relationship.

[0063] Example

[0064] like Figure 1 As shown, the carbon emission calculation method for corporate users in high-energy-consuming industrial parks provided in this embodiment is specifically executed interactively by the industrial park central server and the terminal devices of the corporate users and the energy supply company that are respectively communicated with the industrial park central server, and include but are not limited to the following steps S1 to S11 executed in sequence.

[0065] S1. The industrial park central server sends a model building request to the terminal devices of each enterprise user located in the high-energy-consuming industrial park.

[0066] In step S1, the high-energy-consuming industrial park refers to an industrial park where high-energy-consuming enterprises are concentrated, such as a steel industrial park or an electrolytic aluminum industrial park. The industrial park central server is the management center server of the high-energy-consuming industrial park, which is generally located in the high-energy-consuming industrial park. The terminal device of the enterprise user is the management device of the high-energy-consuming enterprise, or it can be a server. The model building request is used to request the terminal device of each enterprise user to establish a multi-source feature and direct carbon emissions mapping model locally and perform model training, which can be generated and sent conventionally.

[0067] Before the step S1, it is necessary to conduct the following carbon emission accounting boundary analysis on the enterprise users in the high energy consumption industrial park and the feasibility verification of the carbon emission calculation method described in this embodiment.

[0068] like Figure 2 As shown in the figure, the carbon emission sources of enterprise users in high-energy-consuming industrial parks are usually divided into direct carbon emissions, indirect carbon emissions and other indirect carbon emissions, and the total carbon emissions of enterprise users are equal to the sum of all direct carbon emissions and indirect carbon emissions within the user boundary, that is, the following formula:

[0069] E=E 燃烧 +E 过程 +E 电力+E 热力

[0070] Where E represents the total carbon emissions of enterprise users in high-energy-consuming industrial parks, in tons of carbon dioxide (tCO2); E 燃烧 It represents the carbon emissions from fossil fuel combustion of enterprise users in high-energy-consuming industrial parks, in tons of carbon dioxide (tCO2); E 过程 It represents the carbon emissions from the industrial production process of enterprise users in high-energy-consuming industrial parks, in tons of carbon dioxide (tCO2); E 电力 +E 热力 Represents the sum of carbon emissions implicitly generated by corporate users in high-energy-consuming industrial parks due to net purchase of electricity and net purchase of heat, in tons of carbon dioxide (tCO2). According to the relevant carbon market, electricity market and European border adjustment mechanism regulations, the carbon emissions that park corporate users are required to report are mainly scope 1 and scope 2 carbon emissions; at the same time, since the carbon emissions in scope 1 and scope 2 are emissions within the user carbon emission accounting boundary calculated by the producer, and the carbon emissions in scope 3 are calculated emissions outside the boundary and the emissions are extremely low. Therefore, this embodiment mainly focuses on the direct carbon emissions (i.e., carbon emissions in scope 1) and indirect carbon emissions (i.e., carbon emissions in scope 2) of park corporate users.

[0071] The direct carbon emissions mainly include two parts, namely, carbon emissions from fossil fuel combustion and carbon emissions from industrial production processes. Carbon emission sources from fossil fuel combustion are defined as the activities of burning different fossil fuels by different combustion equipment within a certain user accounting boundary. This embodiment uses the emission factor method to calculate the historical value of carbon emissions from fossil fuel combustion, that is, the following formula:

[0072]

[0073] In the formula, i represents a positive integer less than or equal to I, I represents the total amount of fossil energy, AD i It indicates the activity level of the enterprise user using the i-th fossil energy, in million kilojoules (GJ); EF i NCV represents the carbon emission factor of the i-th fossil energy, in tons of carbon dioxide per million kilojoules (tCO2 / GJ); i It represents the average lower calorific value of the i-th fossil energy source during the accounting and reporting period, and for solid or liquid fuels, the unit is million kilojoules per ton (GJ / t), and for gaseous fuels, the unit is million kilojoules per 10,000 cubic meters (GJ / 10,000 m 3 );FC i It represents the net consumption of the i-th fossil energy source during the accounting and reporting period, and for solid or liquid fuels, the unit is tons (t), and for gaseous fuels, the unit is ten thousand cubic meters (ten thousand m 3 );CC irepresents the carbon content per unit calorific value of the i-th fossil energy, in tons of carbon / million kilojoules (tC / GJ); i It represents the carbon oxidation rate of the i-th fossil energy, in %; 44÷12 is the chemical mass ratio of carbon dioxide to carbon atoms.

[0074] Carbon emission sources in industrial production processes are defined as chemical reaction activities that produce carbon dioxide emissions in the production process using carbon-containing materials by a certain user. Since data on carbon-containing materials is complex and difficult to obtain, this embodiment adopts the emission factor method proposed by the IPCC (Intergovernmental Panel on Climate Change). Taking the steel industry as an example, the historical values ​​are calculated using the output of major industrial products and the corresponding default emission factors, that is, the following formula:

[0075] E 过程,steel =∑PO steel ×EF steel

[0076] In the formula, E 过程,steel Represents carbon emissions from the industrial production process of the steel industry; PO steel Represents the output of major industrial products in the steel industry, such as steel and pig iron; EF steel Represents the carbon emission factor of the industrial production process corresponding to the product.

[0077] From the above, it can be seen that carbon emissions from fossil fuel combustion and carbon emissions from industrial production processes are the main components of direct carbon emissions. Since power big data such as power big data or thermal big data have the characteristics of high frequency, wide coverage and rich information, they play an important role in macroeconomic trend forecasting or energy consumption monitoring. This embodiment also analyzes the Pearson correlation coefficient (also known as the Pearson coefficient) of the hourly electricity and heat data and carbon emissions of corporate users in the eight major emission industries in the country, as shown in Table 1 below:

[0078] Table 1. Statistical table of correlation between electricity data, thermal data and carbon emission data of typical industries

[0079]

[0080] As shown in the table above, the energy structure of high-energy-consuming enterprise users is relatively stable, and their electricity and heat consumption are strongly correlated with carbon emission data. Therefore, this embodiment provides a "multi-source feature-carbon" mapping model based on power big data and thermal big data, and combines variational mode decomposition algorithm, time domain convolutional network and multi-head attention long short-term memory network algorithm, etc., to achieve the purpose of high-frequency (such as hourly level) direct carbon emission measurement for high-energy-consuming industrial park enterprise users.

[0081] In addition, the indirect carbon emissions of enterprise users in high-energy-consuming industrial parks mainly include two parts: indirect carbon emissions implied by the purchase of electricity and indirect carbon emissions implied by the purchase of heat. The IPCC organization and national standards use the emission factor method as the main method for users to calculate carbon emissions from purchased electricity / heat. The currently widely used method for calculating the carbon emission factor of electricity / heat is the average carbon emission factor method, in which the average carbon emission factor represents the carbon emissions generated by the use of unit electricity within a certain area. The calculation principle is the total carbon emissions of the entire power grid divided by the total electricity. However, the aforementioned average carbon emission factor method has the following two major drawbacks: (1) The calculation frequency of the average carbon emission factor method is mainly annual, and it is mainly based on the country as a whole, which leads to the disadvantages of low frequency and coarse granularity in this method; (2) The average carbon emission factor method does not consider the impact of cross-regional electricity / heat transfer in and out on the accuracy of the emission factor. Therefore, it is also necessary for this embodiment to dynamically calculate the regional electricity / heat carbon emission factor based on the carbon flow situation under the consideration of regional electricity / heat transfer.

[0082] S2. After receiving the model building request, the terminal device of each enterprise user uses the first public key in the first key pair generated in advance using the homomorphic encryption technology to encrypt the local device information and the total amount of local available data, obtains the first ciphertext and uploads it to the industrial park central server.

[0083] In step S2, the local device information specifically includes but is not limited to the details of the local server device, and the total amount of local available data specifically includes but is not limited to the total amount of local data sets (i.e., the total amount of multi-source feature time series data and carbon emission time series data described later). The homomorphic encryption technology is a special encryption method that allows users to perform calculations directly on encrypted data without decrypting the data, that is, after the calculation is completed, the result is still in an encrypted state, and only users with decryption keys can see the true value of the calculation result. Specifically, the first key pair can be pre-generated according to, but not limited to, the following steps S21 to S23.

[0084] S21. Randomly generate two large prime numbers p and q according to the following formula:

[0085] gcd(p×q,(p-1)×(q-1))=1

[0086] In the formula, gcd() represents the function of finding the greatest common divisor.

[0087] In step S21, a large prime number is an existing concept, specifically referring to a number that satisfies the following conditions: The prime number p k+1 、p k+2 and pk+3 wait.

[0088] S22. Randomly select an integer that satisfies the following formula

[0089] μ=(L(g λ modn 2 )) -1 modn

[0090] Where n = p × q, λ = lcm(p-1,q-1), lcm() represents the function for finding the least common multiple, and L() represents the predefined function based on the variable x. μ represents the decryption factor, Indicates less than n 2 The set of positive integers .

[0091] S23. Use homomorphic encryption technology to generate a public key (n, g) and a private key (λ, g) to form a set of key pairs.

[0092] In step S2, the first key pair can be pre-generated by the industrial park central server, and the first private key in the first key pair is retained, and the first public key is sent to the terminal device of the enterprise user (different first key pairs need to be generated for different enterprise users), or it can be pre-generated by the terminal device of the enterprise user, and the first public key is retained, and the first private key in the first key pair is sent to the industrial park central server. More specifically, the first ciphertext can be, but is not limited to, encrypted and decrypted according to the following steps S24 to S26.

[0093] S24. Randomly select parameter r1∈Z n , where Z n Represents the set of positive integers less than n.

[0094] S25. According to the following formula, the plaintext To encrypt:

[0095]

[0096] In the formula, Indicates the plain text The encrypted ciphertext.

[0097] S26. According to the following formula, the ciphertext To decrypt:

[0098]

[0099] In the formula, Dec() represents the function for decrypting the ciphertext.

[0100] S3. The industrial park central server uses the first private key in the first key pair to decrypt the received first ciphertext to obtain the terminal device information of each enterprise user and the total amount of available data, and determines the parameter configuration content of the multi-source characteristics and direct carbon emissions mapping model based on the terminal device information and the total amount of available data of each enterprise user, and then extracts the model construction related information from the parameter configuration content and distributes it to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user can complete the preliminary construction of the local multi-source characteristics and direct carbon emissions mapping model based on the model construction related information.

[0101] In step S3, the multi-source feature and direct carbon emission mapping model is used to estimate direct carbon emissions based on multi-source data such as power data and thermal data. In order to fully analyze the complex relationship between power consumption and thermal consumption and direct carbon dioxide emissions, the multi-source feature and direct carbon emission mapping model is preferably constructed using a variational mode decomposition algorithm, a time domain convolutional network, and a multi-head attention long short-term memory network algorithm (the technical details of the aforementioned algorithm and network are shown in subsequent steps). The specific method of determining the parameter configuration content of the multi-source feature and direct carbon emission mapping model based on the terminal device information and the total amount of available data of each enterprise user is an existing conventional method; specifically, the model construction related information includes but is not limited to model hyperparameters and model detailed deployment methods, among which the model hyperparameters include but are not limited to the intrinsic mode components of the variational mode decomposition algorithm, the model learning rate, the local iteration rounds, the number of model layers and / or the number of layer neurons, and the detailed model deployment method includes but is not limited to the requirements and deployment methods for software environments such as Python or Pytorch. In addition, the model construction related information can also be transmitted to the terminal devices of each enterprise user in an encrypted manner, so that the terminal devices of each enterprise user can decrypt and complete the preliminary construction and debugging of the local multi-source characteristics and direct carbon emissions mapping model based on the model construction related information.

[0102] S4. The industrial park central server randomly initializes the model parameters of the multi-source characteristics and direct carbon emissions mapping model and distributes them to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user complete the initialization of the local multi-source characteristics and direct carbon emissions mapping model according to the model parameters.

[0103] In step S4, the model parameters can also be transmitted to the terminal devices of the respective enterprise users in an encrypted manner, so that the terminal devices of the respective enterprise users can decrypt and complete the initialization of the local multi-source feature and direct carbon emission mapping model according to the model parameters.

[0104] S5. The industrial park central server randomly selects K corporate users located in the high-energy-consuming industrial park, and sends a model training request to the terminal device of each selected user among the K corporate users, where K represents a positive integer less than or equal to M, and M represents the total number of corporate users located in the high-energy-consuming industrial park.

[0105] In step S5, each of the K enterprise users is a selected user. The model training request is used to request the terminal device of each selected user to locally train the multi-source feature and direct carbon emission mapping model, which can be conventionally generated and sent.

[0106] S6. After receiving the model training request, the terminal device of each selected user applies the local multi-source feature time series data and carbon emission time series data to train the local multi-source feature and direct carbon emission mapping model, and in the process of model iteration and update, uses the federal proximal weight score FedPSW local strategy and differential privacy technology to encrypt and update the local model parameters, and then uploads the final parameters of the local model to the industrial park central server at the end of the model iteration and update, wherein the multi-source feature time series data includes but is not limited to the electricity data and thermal data consumed in each historical unit time period in multiple continuous historical unit time periods, and the carbon emission time series data includes but is not limited to the carbon emissions generated in each historical unit time period.

[0107] In the step S6, the unit time period can be, but is not limited to, hours, that is, the multi-source feature time series data will include the power data and thermal data consumed in each hour over multiple days, and the carbon emission time series data will include the carbon emissions generated in each hour over the multiple days. Specifically, when the multi-source feature and direct carbon emission mapping model is constructed using a variational mode decomposition algorithm, a time domain convolutional network, and a multi-head attention long short-term memory network algorithm, the local multi-source feature time series data and carbon emission time series data are used to train the local multi-source feature and direct carbon emission mapping model, including but not limited to the following steps S61 to S66.

[0108] S61. Obtain local multi-source characteristic time series data and carbon emission time series data, and then execute step S62, wherein the multi-source characteristic time series data includes but is not limited to the electricity data and thermal data consumed in each historical unit time period in multiple consecutive historical unit time periods, and the carbon emission time series data includes but is not limited to the direct carbon emissions generated in each historical unit time period.

[0109] S62. Using variational mode decomposition algorithm to analyze the power time series data P in the multi-source characteristic time series data = {p1, p2, ..., p t,…,p T} and thermal time series data H = {h1,h2,…,h t ,…,h T} is decomposed to obtain the multi-source feature sequence {imf P,1 ,imf P,2 ,···,imf P,t ,···,imf P,T ,imf H,1 ,imf H,2 ,···,imf H,t ,···,imf H,T}, then execute step S63, wherein T represents the total number of time periods of the plurality of consecutive historical unit time periods, t represents a positive integer less than or equal to T, and p t represents the power data consumed in the tth historical unit period among the plurality of continuous historical unit periods, h t represents the heat data consumed in the t-th historical unit period, imf P,t Indicates that p t The corresponding modal components, imf H,t Indicates that t The corresponding modal components.

[0110] In step S62, the variational mode decomposition algorithm (VMD) is an adaptive and completely non-cyclic mode change and signal processing technology, which combines the Wiener filter, Hilbert transform and the alternating direction method of multipliers (ADMM) technology. VMD can determine the number of mode decompositions of a given sequence according to the specific situation, solving the problems of mode mixing and boundary effects. VMD decomposes the original sequence signal into an internal mode function (IMF) with a limited bandwidth, where each mode component The frequency is concentrated at the center frequency The variational problem in VMD is essentially a constrained optimization problem, the core of which is the construction and solution of the variational problem. The objective function of the variational problem is to minimize the sum of the estimated bandwidths of the IMFs, and the constraint is that the sum of the IMFs is the same as the original signal. t ,…,p T} for decomposition processing as an example, the specific calculation formula is as follows:

[0111]

[0112] Where imfP ={imf P,1 ,imf P,2 ,...,imf P,T} and ω={ω1,ω2,...,ω T} respectively represent the set of all modes and their corresponding center frequencies; T represents the total number of modes; represents the unit impulse function, δ(t) represents the Dirac distribution; represents the convolution operation, e represents the base of the natural logarithm, j represents the imaginary part, Represents the square of the Euclidean norm.

[0113] To simplify the above equations, VMD introduces a quadratic penalty term α and Lagrange multipliers to transform the constrained problem into an unconstrained problem:

[0114]

[0115] In the formula, α is used to reduce the interference of noise and ensure the reconstruction accuracy of the signal, and λ is used to maintain the strictness of the constraint.

[0116] Finally, in order to obtain the optimal solution of the above equation, ADMM is used to optimize the variational equation and find the optimal solution in the iterative optimization sequence, where the update imf P,k ,ω k The iterative formula of and λ is as follows:

[0117]

[0118] In the formula, and Represent p one by one t 、imf P,t , λ(t) and Fourier transform form; n and τ represent the number of iterations and update parameters respectively. The iteration stops when the convergence condition satisfies the following formula:

[0119]

[0120] In the formula, ε represents a preset threshold.

[0121] S63. Based on the multi-source feature sequence, a feature matrix X is formed. t′-1 ,X t′ ,X t′+1 ,X t′-2 ,…,X T}, then execute step S64, where X t′ ={imf P,1 ,imf P,2 ,···,imf P,t′,···,imf H,1 ,imf H,2 ,···,imf H,t′}, t′ represents a positive integer less than or equal to T.

[0122] S64. X t′-1 ={imf P,1 ,imf P,2 ,···,imf P,t′-1 ,···,imf H,1 ,imf H,2 ,···,imf H,t′-1} is input into the time domain convolutional network to obtain the modal feature sequence {I1,I2,···,I t′-1}, then execute step S65.

[0123] In step S64, the temporal convolutional network (TCN) is a variant of the convolutional neural network (CNN). The TCN network uses causal convolution, dilated convolution and residual blocks to extract sequence data with large receptive field and temporality, such as Figure 3 As shown. The TCN network used in this embodiment adopts dilated causal convolution with dilation factors d=1, 2, 4, 8 and filter coefficient k=3. In order to ensure training efficiency, TCN also introduces multiple residual blocks to accelerate the mapping model, where each residual block consists of two dilated causal convolution layers with the same dilation factor, a normalization layer, a rectifier linear unit (Rectifier Linear Units, ReLU) activation layer and a discarded Dropout layer. Normalization is performed after the causal convolution operation, ReLU is used for linear correction, and a Dropout layer is added after correction to regularize the network.

[0124] S65. The modal feature sequence {I1, I2, ···, I t′-1} as the model input item, and the direct carbon emissions generated in the carbon emission time series data and in the t′-1th historical unit time period in the multiple continuous historical unit time periods as the model output item, train the artificial intelligence model based on the multi-head attention long short-term memory network algorithm, and back-propagate the loss error obtained from the training, and then execute step S66.

[0125] In step S65, the overall framework of the Long Short-Term Memory Network with Multi-head Attention (LSTMMA) algorithm is as follows: Figure 4 As shown in the figure, the specific calculation process is as follows: The result obtained by TCN is passed through the linear transformation layer to map out three tensors Q, K and V of the same size as the input. These three tensors are put into the scaled dot product attention calculation part to obtain the attention weights and the reconstructed attention matrix. The calculation method is as follows:

[0126]

[0127] In the formula, Q and K are divided by the inner product operation Apply the normalization function (Softmax) to get an attention weight matrix of appropriate size, and then multiply it by V to get the reconstructed attention weight matrix, whose size is still equal to {e1,e2,···,e t-1}. Finally, Attention is the calculation result of scaled dot product attention.

[0128] The above calculation process is the process of obtaining the attention matrix in the subspace. In order to obtain the multi-head attention matrix, the tensors Q, K and V are divided into h equal-sized segments in the same direction (usually the last dimension of the tensor), and each segment of the tensor is placed in a different subspace to calculate the attention matrix. After the calculation is completed, the matrices of these subspaces are spliced ​​together in the division direction, as shown in the following formula:

[0129] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0130] In the formula, head i Represents the attention matrix calculated in subspace i, Concat represents the concatenation operation of these sub-attention matrices in the split direction, and finally multiplies them by the parameter matrix W O Perform a linear transformation to obtain the output matrix reconstructed by the multi-head attention mechanism, which has the same scale as the input. MultiHead represents the calculation result of the multi-head attention mechanism.

[0131] After coupling the multi-head attention mechanism with the LSTM (Long Short-Term Memory) network, since the multi-head attention module is included in the forget gate and input gate structure, the output f t and i t After attention reconstruction, it becomes a tensor and participates in the current state c t and output state ht Calculation. tanh and σ represent the hyperbolic tangent activation function and sigmoid activation function respectively. The specific operation formula is as follows:

[0132]

[0133] Where W f , W i , W z and W o The weight matrices of the input, input gate, forget gate, and output gate of the LSTM memory cell are represented one by one, respectively, b i 、b z 、b f and b o Then they represent the corresponding deviation vectors respectively.

[0134] S66. Determine whether the preset second training end condition is currently met. If so, end the training. Otherwise, increment t′ by 1, and then return to step S64.

[0135] In step S66, the second training end condition is specifically but not limited to that the local multi-source feature and direct carbon emission mapping model reaches the training upper limit or converges.

[0136] In step S6, the Federated proximal score weight FedPSW is a method for updating model parameters by scoring model weights in a federated learning algorithm, and the differential privacy technology is a means in cryptography that aims to provide a method for maximizing the accuracy of data queries when querying from a statistical database while minimizing the chance of identifying its records. Specifically, during the iterative update of the model, the local model parameters are encrypted and updated using the Federated proximal score weight FedPSW local strategy and differential privacy technology, including but not limited to the following steps: During the iterative update of the model, the loss error obtained from the training of the mapping model is back-propagated and the parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network are updated according to the following formula:

[0137]

[0138] In the formula, s represents the number of global training rounds currently completed, and w s+1represents the updated parameters of the multi-head attention long short-term memory network algorithm and the time-domain convolutional network in the sth round of model training, η represents the learning rate, ▽loss represents the gradient of the loss error, μ′ represents the proximal term parameter (an existing term that plays an important role in optimization algorithms, especially when dealing with non-smooth optimization problems; the proximal term, also known as the proximal operation, is a method for dealing with non-differentiable or complex functions, which allows optimization without directly calculating the entire function; the proximal term parameter is usually used in conjunction with the gradient descent method to solve certain specific types of optimization problems), w g,s represents the global parameters of the sth round of model training, w s represents the pre-updated parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network in the sth round of model training, represents the encrypted noise function of the differential privacy technique, represents the truth query function, represents the local multi-source characteristics and direct carbon emissions mapping model, represents the standard deviation and ∈ represents the privacy budget, δ represents the model change restriction parameter, e represents the base of the natural logarithm, S represents the real number domain, Pr[] represents the probability function, and X′ t′-1 Represents X t′-1 In addition, the aforementioned loss error can be obtained by conventional calculation using, but not limited to, a root mean square error function.

[0139] S7. The industrial park central server adopts the federal proximal weight score FedPSW global strategy to aggregate the final model parameters from the terminal devices of the selected users, form the global parameters of the multi-source characteristics and direct carbon emissions mapping model and distribute them to the terminal devices of the corporate users, so that the terminal devices of the corporate users can complete the update of the local multi-source characteristics and direct carbon emissions mapping model according to the global parameters.

[0140] In the step S7, specifically, the FedPSW global strategy is adopted to aggregate the final model parameters from the terminal devices of the selected users to form the global parameters of the multi-source feature and direct carbon emissions mapping model, including but not limited to the following steps: Based on the final model parameters from the terminal devices of the selected users, the global parameters w of the multi-source feature and direct carbon emissions mapping model are formed according to the following formula: g,s+1 :

[0141]

[0142] In the formula, k is a positive integer less than or equal to K, Per krepresents the local model performance score corresponding to the k-th selected user among the K enterprise users, D(k) represents the weight score of the k-th selected user in the global model aggregation stage, and w k,s represents the final model parameters from the terminal device of the k-th selected user, s represents the number of global training rounds currently completed, represents a positive integer less than or equal to s, Indicates the user corresponding to the kth selected user and in The final loss error of the model after the round of global training is completed. The aforementioned round of global training is to execute a round of steps S5 to S7.

[0143] S8. The industrial park central server determines whether the preset first training end condition is currently met. If so, the multi-source feature and direct carbon emission mapping model is updated according to the global parameters, otherwise, returns to step S5.

[0144] In step S8, the first training end condition is specifically but not limited to that the number of global training rounds currently completed reaches a preset number of rounds.

[0145] S9. The industrial park central server sends a carbon emission accounting request to the terminal device of an energy supply company that supplies energy to a target corporate user located in the high-energy-consuming industrial park, wherein the energy supply refers to electricity and heat, and the energy supply company refers to an electric power company and a heat company.

[0146] In step S9, the terminal device of the energy supply company is a management device of an electric power company or a heat company, or may be a server. The carbon emission accounting request is used to request the terminal device of the energy supply company to return the latest energy purchase time series data / and the latest indirect carbon emissions of a certain target enterprise user, which can be generated and sent conventionally.

[0147] S10. After receiving the carbon emission accounting request, the terminal device of the energy supply company uses the second public key in the second key pair generated in advance by the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user, obtains the second ciphertext and transmits it back to the industrial park central server, wherein the latest energy purchase time series data includes but is not limited to the electricity data and heat data purchased for each unit time period in at least one recent unit time period.

[0148] In the step S10, for example, if the unit time period is an hour, the latest energy purchase time series data may include the electricity data and thermal data purchased in each hour in the last 24 hours. The specific generation method of the second public key pair is consistent with that of the first public key pair. For details, please refer to the aforementioned steps S21 to S23, which will not be repeated here. The second key pair can be pre-generated by the industrial park central server, and the second private key in the second key pair is retained, and the second public key is sent to the terminal device of the energy supply company. It can also be pre-generated by the terminal device of the energy supply company, and the second public key is retained, and the second private key in the second key pair is sent to the industrial park central server. The specific encryption and decryption process of the second ciphertext is consistent with that of the first ciphertext. For details, please refer to the aforementioned steps S24 to S26, which will not be repeated here.

[0149] In the step S10, if the carbon emission accounting request is also used to request the terminal device of the energy supply company to return the latest indirect carbon emissions of the target corporate user, then in view of the problem that the existing carbon emission factor calculation does not take into account the electricity / heat transfer between the region to which the target park belongs and other regions, preferably, the latest energy purchase time series data of the target corporate user is encrypted using the second public key in the second key pair generated in advance using the homomorphic encryption technology to obtain a second ciphertext, including but not limited to the following steps S101 to S104.

[0150] S101. According to the electricity and heat conditions of each region in the most recent unit time period, determine the latest electricity dynamic carbon emission factor and the latest heat dynamic carbon emission factor for each region taking into account cross-regional flows.

[0151] In step 101, the region is exemplified but not limited to a provincial region, that is, the latest dynamic carbon emission factor of electricity and the latest dynamic carbon emission factor of heat need to be calculated for each provincial region, taking into account the cross-regional flow. Specifically, according to the power and heat conditions of each region in the most recent unit time period, the latest dynamic carbon emission factor of electricity and the latest dynamic carbon emission factor of heat are determined for each region, taking into account the cross-regional flow, including but not limited to the following steps S1011 to S1012.

[0152] S1011. Based on the electricity and heat conditions of each region in the most recent unit time period, calculate the total electricity production, carbon emissions from electricity production, total heat production and carbon emissions from heat production of each region in the most recent unit time period, as well as the electricity transfer ratio and heat transfer ratio between each pair of regions, wherein the electricity transfer ratio refers to the ratio of the amount of electricity flowing from the first region in the corresponding region pair to the second region in the corresponding region pair to the total electricity production of the first region, and the heat transfer ratio refers to the ratio of the amount of heat flowing from the first region to the second region to the total heat production of the first region.

[0153] In step S1011, the specific calculation method of the total electricity production, the carbon emissions of electricity production, the total heat production, the carbon emissions of heat production, the electricity transfer ratio and the heat transfer ratio is the existing conventional method. For example, the carbon emissions CE of electricity production can be calculated according to the following formula:

[0154]

[0155] In the formula, Indicates the total number of generating units in the corresponding region, Indicates less than or equal to A positive integer, Indicates the first The carbon emission intensity per unit of electricity of each generator set, Indicates the The total power generation of each generator set.

[0156] S1012. According to the total electricity production, carbon emissions from electricity production, total heat production and carbon emissions from heat production of each region in the most recent unit period, and the proportion of electricity transfer and heat transfer of each pair of regions in the most recent unit period, the latest dynamic carbon emission factor of electricity and the latest dynamic carbon emission factor of heat in each region taking into account cross-regional flows are calculated according to the following formula:

[0157]

[0158] Where n is a positive integer, DFE n represents the latest capacity dynamic carbon emission factor of the nth region and considering cross-regional flow, i and j represent positive integers, e ij N represents the ratio of the capacity flowing from the ith region to the jth region in the most recent unit period to the total capacity production of the ith region. n It represents the total capacity production of the nth region in the most recent unit period, CE nIt represents the carbon emission produced by the capacity of the nth region in the most recent unit time period, where the capacity refers to electricity or heat.

[0159] S102. Based on the electricity data p purchased by the target enterprise user in the recent unit period T+1 The latest dynamic carbon emission factor DFE of electricity in the region where the target enterprise user belongs and taking into account cross-regional flows P,T+1 , the carbon emissions CE of the purchased electricity of a target enterprise user in the most recent unit period is calculated according to the following formula: indirect,P,T+1 =DFE P,T+1 ×p T+1 .

[0160] S103. Based on the heat data h purchased by the target enterprise user in the recent unit period T+1 The latest thermal dynamic carbon emission factor DFE of the region to which the target enterprise user belongs and considering cross-regional flows H,T+1 , the carbon emission CE of purchased heat power of a target enterprise user in the most recent unit period is calculated according to the following formula: indirect,H,T+1 =DFE H,T+1 ×h T+1 .

[0161] S104. Use the second public key in the second key pair generated in advance using the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user and the carbon emissions of purchased electricity and purchased heat of the target enterprise user in the most recent unit period to obtain a second ciphertext, wherein the latest energy purchase time series data includes but is not limited to electricity data and heat data purchased for each unit period in at least one recent unit period.

[0162] S11. The industrial park central server uses the second private key in the second key pair to decrypt the received second ciphertext, obtains the current latest energy purchase time series data of the target enterprise user, imports the multi-source feature and direct carbon emissions mapping model, and outputs the direct carbon emissions of the target enterprise user in the most recent unit time period.

[0163] In step S11, if the second ciphertext also contains the encrypted carbon emissions of purchased electricity and purchased heat of the target enterprise user in the most recent unit time period, in order to obtain the total carbon emissions of the target enterprise user in the most recent unit time period, preferably, after step S10, the method also includes the following steps: the industrial park central server adds the ciphertext data of carbon emissions of purchased electricity and the ciphertext data of carbon emissions of purchased heat in the second ciphertext, and uses the second private key to decrypt the addition result to obtain the indirect carbon emissions of the target enterprise user in the most recent unit time period, and then adds the direct carbon emissions and indirect carbon emissions of the target enterprise user in the most recent unit time period to obtain the total carbon emissions of the target enterprise user in the most recent unit time period.

[0164] In order to verify the effectiveness of the carbon emission calculation method proposed in this embodiment, this embodiment also uses a data set of 31 corporate users in Tianjin and belonging to 4 high-energy-consuming industrial industries (which comes from the Tianjin Carbon Emission Monitoring Demonstration Project, and selects typical high-energy-consuming corporate users with obvious energy consumption characteristics, including hourly electricity data, thermal data and energy consumption data from January 1, 2021 to December 1, 2023) for example analysis: the data set is divided into training set, verification set and test set in a ratio of 6:2:2, and two scenarios of typical high-energy-consuming industries: steel industry and electrolytic aluminum industry are selected for display to test the accuracy of the carbon emission calculation method proposed in this embodiment.

[0165] In the case analysis process, the VMD-TCN-LSTMMA algorithm (i.e., the abbreviation of the variational mode decomposition algorithm, the time domain convolutional network, and the multi-head attention long short-term memory network algorithm) is used to construct a direct carbon emission measurement model for high-energy-consuming enterprise users (i.e., the multi-source feature and direct carbon emission mapping model), wherein the VMD algorithm is used to decompose time series data, and the parameter that needs to be set is the number of decomposed intrinsic mode components; the TCN network is used for feature extraction, and the parameters that need to be set are the expansion factor and the filter coefficient; the LSTMMA algorithm is used for model training, and the specific hyperparameter settings are shown in Table 2 below:

[0166] Table 2. Basic parameter setting table of VMD-TCN-LSTMMA algorithm

[0167]

[0168] In order to solve the privacy and security issues of user data and promote data fusion, this embodiment adopts a horizontal federated learning training framework based on differential privacy protection to train the above-mentioned carbon emission estimation model: based on the distributed concept of federated learning, during the training process, after completing a round of local training through the local VMD-TCN-LSTM model, the key model control parameters are transmitted to the control center (i.e., the central server of the industrial park), and the global parameter update is completed by aggregation to achieve global model optimization. This process does not require direct data transmission, which greatly improves the security and privacy of the training effect. Through this advanced distributed learning method, it can meet the efficiency and flexibility requirements of model training, and can also ensure data privacy and information security. Based on the number of typical industry users who respond to federated learning invitations in actual environments, the key parameters in the federated learning framework are shown in Table 3 below.

[0169] Table 3. Federated learning framework parameter setting table

[0170] Parameter name Settings Framework Environment Flower Number of clients 5 Local training rounds 8 Global training rounds 100 Learning Rate 0.001 Dropout 0.1

[0171] The specific program of the carbon emission calculation method proposed in this embodiment is written in Python 3.7 and implemented on the TensorFlow 2.2 platform, using Keras version 2.3.1; and running on a computer with an AMD Ryzen 75800H CPU, a main frequency of 3.20 GHz and a memory of 32G. In addition, the deployment mode of the federated algorithm is single-machine deployment.

[0172] This embodiment also uses the following four error statistics to evaluate model performance: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and determination coefficient R 2 , their calculation formulas are as follows:

[0173]

[0174] In the formula, and They represent the actual and calculated values ​​of the total carbon emissions of enterprise user i at hour t respectively; represents the mean of the actual value of the user's carbon emissions, and the other formula symbols are derived based on the corresponding indicator formula. For the accuracy evaluation of the carbon emission calculation method proposed in this embodiment, the smaller the values ​​of MAE, RMSE and MAPE, the higher the accuracy of the carbon emission calculation result; the determination coefficient R 2 The closer the value is to 1, the better the carbon emission calculation results will be.

[0175] This embodiment also selects four benchmark models, namely LSTM, LSTMMA, TCN, and TCN-LSTMMA, which have been verified in the field of time series analysis, for comparison and verification with the multi-source feature and direct carbon emission mapping model proposed in this embodiment (specifically, it can be represented by VMD-TCN-LST MMA); Figure 5 and Figure 6 As shown in the figure, it can be clearly seen that the mapping model proposed in this embodiment is superior to the other four benchmark models in terms of the four evaluation index dimensions and can better fit the curve change trend. Compared with the other four benchmark models, the mapping model proposed in this embodiment has better feature capture ability, not only with a low error rate, but also with a high determination coefficient R 2 This shows that the mapping model avoids singularities caused by data mutations. This proves that this embodiment can effectively help the LSTM network better capture the time series change characteristics implied by the input features by adding VMD data decomposition, TCN feature extraction steps and multi-head attention mechanism. In summary, the mapping model proposed in this embodiment has the learning ability to extract key features from time series data such as electricity and heat and establish correlation analysis with the output sequence, so as to accurately estimate the hourly carbon emissions of high-energy-consuming enterprise users. The performance results of VMD-TCN-LSTMMA on the four indicators of the other four benchmark models are shown in Table 4 below:

[0176] Table 4. Statistical table of performance improvement results of VMD-TCN-LSTMMA and other four benchmark models

[0177]

[0178]

[0179] This example also verifies the role of the proposed horizontal federated learning framework and the impact of the number of federated participants on the model training results: by controlling the parameters (i.e., the number of local clients, i.e., the number of selected users K), the accuracy of model calculations trained in different scenarios is obtained, such as Figure 7 and Figure 8As shown, it can be observed that the increase in the number of local clients for steel enterprise users from 2 to 3 did not produce a significant improvement in the training effect, while the increase in the number of local clients for petrochemical enterprise users from 3 to 5 improved the model calculation accuracy by 2.1%. This is because the number of clients increased by a small amount and the amount of model training data failed to increase to the extent that affected the training accuracy. At the same time, it can also be observed that the carbon emission measurement error of the target user's decentralized training model is relatively large, and the average deviation is concentrated in the range of 7% to 9% as a whole. The model accuracy of the horizontal federated training framework proposed in this embodiment is concentrated around 95% to 97%. This is because the model based on horizontal federated learning can benefit from the shared knowledge of multiple clients (these users participate in the model training process locally without compromising the privacy of their data). By aggregating the knowledge learned from multiple federated participants, the model based on horizontal federated learning can achieve higher calculation accuracy without the risk of data leakage.

[0180] Based on the above example analysis, the relative error of the carbon emission calculation method proposed in this embodiment and the comparison between the calculation result and the actual value of carbon emissions are shown in Table 5 below:

[0181] Table 5. Comparison statistics of the errors between the calculation results of the carbon emission calculation method proposed in this embodiment and the actual value of carbon emissions

[0182]

[0183] In summary, the carbon emission calculation method for enterprise users in high-energy-consuming industrial parks provided by this embodiment has the following technical effects:

[0184] (1) This embodiment provides a new solution for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks that takes into account the privacy and security of multiple parties. That is, for the required direct carbon emissions of enterprise users in high-energy-consuming industrial parks, on the one hand, a multi-source feature and direct carbon emissions mapping model is constructed and trained to perform direct carbon emissions calculation by utilizing the wide coverage and real-time nature of electricity and heat data. On the other hand, in order to solve the data privacy and security issues of the park, enterprise users in the park, and electric power and heat companies, an information interaction method based on the PP3-FL framework (i.e., the application framework of the first key pair, differential privacy technology, and the second key pair) and a mapping model construction method based on the FedPSW strategy are introduced. The high-frequency, accurate, and efficient calculation of carbon emissions of enterprise users in the park can be achieved under the condition that local data and information are available but invisible. This can solve the problems of imperfect accounting, limited accuracy of calculation results, and failure to consider user privacy during model training and application in existing carbon emissions calculation solutions for enterprise users in high-energy-consuming industrial parks, and facilitate practical application and promotion.

[0185] (2) In order to enhance the ability to mine data information, the VMD-TCN-LSTMMA algorithm is also used in the mapping model to decompose and upgrade the dimensions and extract features of power and thermal time series data, thereby enhancing the learning ability to extract key features from time series data and establish correlation analysis with the output sequence;

[0186] (3) For the indirect carbon emissions of enterprise users in industrial parks, a method for calculating the dynamic carbon emission factors of heat and electricity based on cross-domain carbon flows is also proposed. This method can then achieve the purpose of dynamically and accurately calculating the carbon emission factors of regional heat and electricity while taking into account the carbon emissions of electricity and heat transferred between the target park and the national / provincial region, thereby further improving the accuracy of the total carbon emissions calculation results.

[0187] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks, characterized in that: The method comprises the following steps S1 to S11 which are performed in sequence: S1. The industrial park central server sends a model building request to the terminal devices of each enterprise user located in the high-energy-consuming industrial park; S2. After receiving the model building request, the terminal device of each enterprise user uses the first public key in the first key pair generated in advance using homomorphic encryption technology to encrypt the local device information and the total amount of local available data, obtain the first ciphertext and upload it to the industrial park central server; S3. The industrial park central server uses the first private key in the first key pair to decrypt the received first ciphertext, obtain the terminal device information of each enterprise user and the total amount of available data, and determine the parameter configuration content of the multi-source feature and direct carbon emissions mapping model based on the terminal device information of each enterprise user and the total amount of available data, and then extract the model construction related information from the parameter configuration content and distribute it to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user can complete the preliminary construction of the local multi-source feature and direct carbon emissions mapping model based on the model construction related information; S4. The industrial park central server randomly initializes the model parameters of the multi-source characteristics and direct carbon emissions mapping model and distributes them to the terminal devices of each enterprise user, so that the terminal devices of each enterprise user complete the initialization of the local multi-source characteristics and direct carbon emissions mapping model according to the model parameters; S5. The industrial park central server randomly selects K enterprise users located in the high-energy-consuming industrial park, and sends a model training request to the terminal device of each selected user among the K enterprise users, where K represents a positive integer less than or equal to M, and M represents the total number of enterprise users located in the high-energy-consuming industrial park; S6. After receiving the model training request, the terminal device of each selected user applies the local multi-source feature time series data and carbon emission time series data to train the local multi-source feature and direct carbon emission mapping model, and in the process of model iteration and update, the local model parameters are encrypted and updated using the federal proximal weight score FedPSW local strategy and differential privacy technology, and then at the end of the model iteration and update, the final parameters of the local model are uploaded to the industrial park central server, wherein the multi-source feature time series data contains the power data and thermal data consumed in each historical unit time period in multiple continuous historical unit time periods, and the carbon emission time series data contains the direct carbon emissions generated in each historical unit time period; S7. The industrial park central server adopts the FedPSW global strategy of the federated proximal weight score to aggregate the final model parameters from the terminal devices of the selected users, form the global parameters of the multi-source feature and direct carbon emissions mapping model and distribute them to the terminal devices of the enterprise users, so that the terminal devices of the enterprise users can complete the update of the local multi-source feature and direct carbon emissions mapping model according to the global parameters; S8. The industrial park central server determines whether the preset first training end condition is currently met. If so, the multi-source feature and direct carbon emission mapping model is updated according to the global parameters, otherwise, it returns to step S5; S9. The industrial park central server sends a carbon emission accounting request to a terminal device of an energy supply company that supplies energy to a target enterprise user located in the high-energy-consuming industrial park, wherein the energy supply refers to electricity and heat supply, and the energy supply company refers to an electric power company and a heat company; S10. After receiving the carbon emission accounting request, the terminal device of the energy supply company uses the second public key in the second key pair generated in advance by the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user, obtains the second ciphertext and transmits it back to the central server of the industrial park, wherein the latest energy purchase time series data contains the power data and heat data purchased in each unit time period in at least one recent unit time period; S11. The industrial park central server uses the second private key in the second key pair to decrypt the received second ciphertext, obtains the current latest energy purchase time series data of the target enterprise user, imports the multi-source feature and direct carbon emissions mapping model, and outputs the direct carbon emissions of the target enterprise user in the most recent unit time period.

2. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 1 is characterized in that: The latest energy purchase time series data of the target enterprise user is encrypted using the second public key in the second key pair generated in advance using the homomorphic encryption technology to obtain a second ciphertext, including: According to the electricity and heat conditions of each region in the most recent unit time period, determine the latest electricity dynamic carbon emission factor and the latest heat dynamic carbon emission factor for each region taking into account cross-regional flows; According to the electricity data p purchased by the target enterprise user in the recent unit period T+1 The latest dynamic carbon emission factor DFE of electricity in the region where the target enterprise user belongs and taking into account cross-regional flows P,T+1 , the carbon emissions CE of the purchased electricity of a target enterprise user in the most recent unit period is calculated according to the following formula: indirect,P,T+1 =DFE P,T+1 ×p T+1 ; According to the heat data h purchased by a certain target enterprise user in the recent unit period T+1 The latest thermal dynamic carbon emission factor DFE of the region to which the target enterprise user belongs and considering cross-regional flows H,T+1 , the carbon emission CE of purchased heat power of a certain target enterprise user in the most recent unit period is calculated according to the following formula: indirect,H,T+1 =DFE H,T+1 ×h T+1 ; Use the second public key in the second key pair generated in advance by the homomorphic encryption technology to encrypt the latest energy purchase time series data of the target enterprise user and the carbon emissions of purchased electricity and purchased heat of the target enterprise user in the most recent unit period to obtain a second ciphertext, wherein the latest energy purchase time series data contains the electricity data and heat data purchased in each unit period in at least one most recent unit period; After step S10, the method further includes: The industrial park central server adds the ciphertext data of carbon emissions from purchased electricity and the ciphertext data of carbon emissions from purchased heat in the second ciphertext, and uses the second private key to decrypt the addition result to obtain the indirect carbon emissions of the target enterprise user in the most recent unit period, and then adds the direct carbon emissions and indirect carbon emissions of the target enterprise user in the most recent unit period to obtain the total carbon emissions of the target enterprise user in the most recent unit period.

3. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 2 is characterized in that: According to the power and heat conditions of each region in the most recent unit period, the latest power dynamic carbon emission factor and the latest heat dynamic carbon emission factor of each region taking into account cross-regional flows are determined, including: According to the power situation and heat situation of each region in the most recent unit time period, the total power production, carbon emissions of power production, total heat production and carbon emissions of heat production of each region in the most recent unit time period, as well as the power transfer ratio and heat transfer ratio between each pair of regions are calculated, wherein the power transfer ratio refers to the ratio of the amount of power flowing from the first region in the corresponding region pair to the second region in the corresponding region pair to the total power production of the first region, and the heat transfer ratio refers to the ratio of the amount of heat flowing from the first region to the second region to the total heat production of the first region; According to the total electricity production, carbon emissions from electricity production, total heat production and carbon emissions from heat production of each region in the most recent unit period, and the proportion of electricity transfer and heat transfer of each pair of regions in the most recent unit period, the latest dynamic carbon emission factor of electricity and the latest dynamic carbon emission factor of heat in each region taking into account cross-regional flows are calculated according to the following formula: Where n is a positive integer, DFE n represents the latest capacity dynamic carbon emission factor of the nth region and considering cross-regional flows, i and j represent positive integers, e ij N represents the ratio of the capacity flowing from the ith region to the jth region in the most recent unit period to the total capacity production of the ith region. n It represents the total capacity production of the nth region in the most recent unit period, CE n It represents the carbon emission produced by the capacity of the nth region in the most recent unit time period, where the capacity refers to electricity or heat.

4. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 1 is characterized in that: The first key pair or the second key pair is pre-generated in the following manner: Two large prime numbers p and q are randomly generated according to the following formula: gcd(p×q,(p-1)×(q-1))=1 In the formula, gcd() represents the function of finding the greatest common divisor; Randomly select an integer that satisfies the following formula μ=(L(g λ towards n 2 )) -1 towards n Where n = p × q, λ = lcm(p-1,q-1), lcm( ) represents the function for finding the least common multiple, and L( ) represents the predefined function based on the variable x. μ represents the decryption factor, Indicates less than n 2 The set of positive integers ; Homomorphic encryption technology is used to generate a public key (n, g) and a private key (λ, g) to form a key pair.

5. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 4 is characterized in that: The first ciphertext or the second ciphertext is encrypted and decrypted in the following manner: Randomly select parameter r1∈Z n , where Z n represents the set of positive integers less than n; According to the following formula, the plaintext To encrypt: In the formula, Indicates the plain text The encrypted ciphertext; According to the following formula, the ciphertext To decrypt: In the formula, Dec() represents the function for decrypting the ciphertext.

6. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 1, characterized in that: The multi-source feature and direct carbon emission mapping model is constructed using a variational mode decomposition algorithm, a time domain convolutional network, and a multi-head attention long short-term memory network algorithm.

7. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 6 is characterized in that: Using local multi-source feature time series data and carbon emission time series data to train a local multi-source feature and direct carbon emission mapping model includes the following steps S61 to S66: S61. Acquire local multi-source characteristic time series data and carbon emission time series data, and then execute step S62, wherein the multi-source characteristic time series data contains power data and heat data consumed in each historical unit period in multiple continuous historical unit periods, and the carbon emission time series data contains direct carbon emissions generated in each historical unit period; S62. Using variational mode decomposition algorithm to analyze the power time series data P in the multi-source characteristic time series data = {p1, p2, ..., p t ,…,p T } and thermal time series data H = {h1,h2,…,h t ,…,h T } is decomposed to obtain the multi-source feature sequence {imf P,1 ,imf P,2 ,···,imf P,t ,···,imf P,T ,imf H,1 ,imf H,2 ,···,imf H,t ,···,imf H,T }, then execute step S63, wherein T represents the total number of time periods of the plurality of consecutive historical unit time periods, t represents a positive integer less than or equal to T, and p t represents the power data consumed in the tth historical unit period among the plurality of continuous historical unit periods, h t represents the heat data consumed in the t-th historical unit period, imf P,t Indicates that p t The corresponding modal components, imf H,t Indicates that t The corresponding modal components; S63. Based on the multi-source feature sequence, a feature matrix X is formed. t′-1 ,X t′ ,X t′+1 ,X t′-2 ,…,X T }, then execute step S64, where X t′ ={imf P,1 ,imf P,2 ,···,imf P,t′ ,···,imf H,1 ,imf H,2 ,···,imf H,t′ }, t′ represents a positive integer less than or equal to T; S64. X t′-1 ={imf P,1 ,imf P,2 ,···,imf P,t′-1 ,···,imf H,1 ,imf H,2 ,···,imf H,t′-1 } is input into the time domain convolutional network to obtain the modal feature sequence {I1,I2,···,I t′-1 }, then execute step S65; S65. The modal feature sequence {I1, I2, ···, I t′-1 } as a model input item, and the direct carbon emissions generated in the t′-1th historical unit period in the carbon emission time series data and in the multiple continuous historical unit periods as a model output item, train the artificial intelligence model based on the multi-head attention long short-term memory network algorithm, and back-propagate the loss error obtained from the training, and then execute step S66; S66. Determine whether the preset second training end condition is currently met. If so, end the training. Otherwise, increment t′ by 1, and then return to step S64.

8. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 7, characterized in that: During the iterative update process of the model, the local model parameters are encrypted and updated using the FedPSW local strategy and differential privacy technology, including: During the iterative update process of the model, the loss error obtained from the training of the mapping model is back-propagated and the parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network are updated according to the following formula: In the formula, s represents the number of global training rounds currently completed, and w s+1 represents the updated parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network in the sth round of model training, η represents the learning rate, represents the gradient of the loss error, μ′ represents the proximal term parameter, and w g,s represents the global parameters of the sth round of model training, w s represents the pre-updated parameters of the multi-head attention long short-term memory network algorithm and the time domain convolutional network in the sth round of model training, represents the encrypted noise function of the differential privacy technique, represents the truth query function, represents the local multi-source characteristics and direct carbon emissions mapping model, represents the standard deviation and ∈ represents the privacy budget, δ represents the model change restriction parameter, e represents the base of the natural logarithm, S represents the real number domain, Pr[] represents the probability function, X t ″ -1 Represents X t′-1 adjacent datasets.

9. The method for calculating carbon emissions of enterprise users in high-energy-consuming industrial parks as claimed in claim 1, characterized in that: The FedPSW global strategy is adopted to aggregate the final model parameters from the terminal devices of each selected user to form the global parameters of the multi-source feature and direct carbon emissions mapping model, including: According to the final model parameters from the terminal devices of the selected users, the global parameter w of the multi-source feature and direct carbon emission mapping model is formed according to the following formula: g,s+1 : In the formula, k is a positive integer less than or equal to K, Per k represents the local model performance score corresponding to the k-th selected user among the K enterprise users, D(k) represents the weight score of the k-th selected user in the global model aggregation stage, and w k,s represents the final model parameters from the terminal device of the k-th selected user, s represents the number of global training rounds currently completed, represents a positive integer less than or equal to s, Indicates the user corresponding to the kth selected user and in The final loss error of the model after the round of global training is completed.

10. The method for calculating carbon emissions of enterprise users in high energy consumption industrial parks according to claim 1, characterized in that: The model building related information includes model hyperparameters and detailed model deployment methods.

Images