Carbon emission monitoring method and device based on real-time power consumption data

Abstract

The invention discloses a carbon emission monitoring method and device based on real-time power consumption data. The method comprises the following steps: acquiring power attribute data of a power consumption object; wherein the power attribute data comprises power consumption data, photovoltaic output and production process data; performing multi-scale load division on the power consumption data in sequence, and determining a meteorological load and a production load of the power consumption object; based on a pre-constructed electric energy model, photovoltaic output and meteorological load and production load of a power utilization object are analyzed and processed, and an energy consumption prediction value is given; and based on the power attribute data, the production load and the energy consumption prediction value, performing carbon emission treatment of different emission sources, and giving the total carbon emission of the power utilization object. The method can achieve the precise decomposition of the meteorological load and the production load, and achieves the real-time and refined estimation of the comprehensive carbon emission.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission prediction technology, specifically relating to a carbon emission monitoring method and device based on real-time electricity consumption data. Background Technology

[0002] With the deepening of the "dual carbon" strategic goals and the comprehensive acceleration of the construction of new power systems, the power system is undergoing a profound transformation from traditional centralized power generation to high-proportion renewable energy integration, generation-grid-load-storage interaction, and distributed smart electricity consumption. In this process, achieving real-time and accurate monitoring and accounting of enterprise-level carbon emissions has become a key technological foundation and urgent need for implementing carbon emission "dual control" assessments, enabling enterprises to carry out refined energy conservation and carbon reduction management, and participating in carbon market trading.

[0003] Currently, carbon emission perception at the enterprise level faces two major challenges: first, traditional accounting methods cannot meet the requirements of real-time and precision; second, existing predictive technologies are insufficient to support process control and real-time decision-making.

[0004] Traditional accounting methods primarily rely on periodic manual calculations based on annual energy ledgers and static emission factors. These methods suffer from poor timeliness, large errors, and an inability to capture real-time fluctuations in carbon emissions. While direct measurement via continuous energy monitoring systems (CEMS) is possible, the high cost and complex maintenance make large-scale deployment difficult across enterprises. Especially with the increasing integration of renewable energy and the growing frequency of green electricity trading, the carbon footprint of electricity exhibits strong spatiotemporal dynamics, rendering regional grid averages or static carbon factors inaccurate in reflecting the true carbon emission intensity of enterprise electricity consumption.

[0005] At the level of existing technical solutions, most focus on macro-level prediction of carbon emissions, which has inherent limitations and cannot meet the core needs of enterprise-level real-time monitoring, as follows: For example, patent CN119721336A discloses a carbon emission prediction method based on big data in the power sector, including the following steps: S1: Collect power production and consumption data to build a data pool; S2: Extract and generate statistical features through statistical analysis and autoencoder technology, extract spatial features using geographic information system technology, and analyze the periodicity and trend of time series data to obtain temporal features, introduce causal inference methods to identify key causal features, and finally construct a comprehensive feature set; S3: Use a regression model as the basic model, introduce LSTM to process complex and long-sequence data, and enhance the model's nonlinear fitting ability, improve the model's generalization ability and accuracy through ensemble learning, construct a carbon emission prediction model, and train it based on the comprehensive feature set; S4: Based on the trained carbon emission prediction model, perform carbon emission prediction according to real-time data; S5: Display the prediction results through visualization tools, supporting user-defined views and analysis. Although it can integrate multi-source big data and utilize advanced machine learning models (such as LSTM and XGBoost) for regional carbon emission prediction, its essence is a "black box" correlation prediction. This technology lacks modeling of the physical causal chain of "electricity consumption → energy consumption → carbon emissions", cannot explain the specific sources of carbon emissions (such as production load or meteorological load), and does not consider the spatiotemporal changes of the electrical carbon factor caused by power grid flow dynamics. Therefore, its output results cannot provide enterprises with traceable and actionable refined carbon management basis.

[0006] For example, patent CN115713149A discloses a method for predicting corporate carbon peak trends. This method selects influencing factors, collects cleaned historical data of these factors, and constructs a predictive model based on the correlation between these factors and carbon emissions. Finally, the cleaned historical data is input into the predictive model to obtain future annual carbon emissions, thus determining the company's carbon peak amount and peak year. This method uses a quantitative model based on macroeconomic socioeconomic indicators (such as the number of employees and GDP structure) for medium- to long-term carbon peak trend analysis. However, this method relies on low-frequency annual statistical data, and the output is a trend forecast for the next few years. It cannot support the real-time carbon emission accounting needs at the hourly or even minute level, and it cannot reflect the immediate impact of operational details such as production scheduling, equipment operation, and weather changes on carbon emissions. This method is severely out of touch with the daily carbon management and real-time decision-making scenarios of enterprises.

[0007] Therefore, there is currently a lack of technical means to penetrate the "black box" of enterprise electricity consumption, couple physical mechanisms with real-time data, dynamically perceive the relationship between electricity and carbon, and thus realize real-time and interpretable accounting of enterprise's comprehensive carbon emissions. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides a carbon emission monitoring method and device based on real-time electricity consumption data. It enables precise decomposition of meteorological load and production load, allowing for real-time and refined estimation of comprehensive carbon emissions.

[0009] In a first aspect, the present invention provides a carbon emission monitoring method based on real-time electricity consumption data, comprising: Obtain the power attribute data of the electricity users; the power attribute data includes electricity consumption data, photovoltaic output and production process data; The electricity consumption data is sequentially divided into multi-scale load segments to determine the meteorological load and production load of the electricity users. Based on a pre-built power model, the photovoltaic output, meteorological load of electricity users, and production load are analyzed and processed to provide a predicted value of energy consumption. Based on electricity attribute data, production load and energy consumption forecasts, carbon emission processing is performed on different emission sources to provide the total carbon emissions of electricity users.

[0010] Furthermore, the electricity consumption data is sequentially divided into multi-scale load segments to determine the meteorological load and production load of the electricity users, including: The current electricity consumption data is decomposed into load characteristics to determine the trend characteristics, periodic characteristics, and residual characteristics. The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. The characteristics of the trend cycle term are decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. Based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load.

[0011] Furthermore, the current electricity consumption data is decomposed into load characteristics to determine trend characteristics, periodic characteristics, and residual characteristics. This includes using the Robust-STL time series decomposition method to decompose the current electricity consumption data into load characteristics and provide trend characteristics, periodic characteristics, and residual characteristics. The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. This includes using the Henderson moving average method to decompose the periodic features by moving average and give the trend cycle features, seasonal features, and random features. The trend cycle feature is decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. This includes using the Hodrick-Prescott smoothing filter method to decompose the trend cycle feature into long-term and short-term components and give the long-term trend component and the short-term fluctuation component.

[0012] Furthermore, based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load, including: The system obtains the current ambient temperature, preset cooling temperature, and preset heating temperature, and combines these with seasonal characteristics to determine the initial meteorological load and non-meteorological load at the current moment. Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load; Determine the meteorological and production attribution weights at the current moment, and combine the initial meteorological load, non-meteorological load, initial production load, and non-production load to give the final meteorological load and production load.

[0013] Furthermore, the current ambient temperature, preset cooling temperature, and preset heating temperature are obtained, and combined with seasonal characteristics, the initial meteorological load and non-meteorological load at the current moment are determined, including: Obtain the current ambient temperature, preset cooling temperature, and preset heating temperature to determine the initial meteorological load at the current moment; Obtain pre-trained meteorological regression models for seasonal features, meteorological load, and non-meteorological load; Based on the difference between the meteorological regression model and the initial meteorological load at the current time, the non-meteorological load at the current time is given; Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load, including: Obtain the number of production load patterns for user objects and determine the number of clusters. Using the initial production load as the centroid and combining the cluster number, cluster analysis is performed on the short-term fluctuation components at the current moment to give the initial production load; The non-production load is determined based on the difference between the trend cycle term characteristics and the initial production load.

[0014] Furthermore, the determination of the meteorological attribution weight and the production attribution weight at the current moment includes: Determine the historical temperature standard deviation and, in conjunction with the current ambient temperature and the preset thermal neutral temperature, give the intensity of the meteorological driving force; Determine the standard deviation of historical capacity intensity, and combine the current capacity intensity with the historical average capacity intensity to give the intensity of production driving force; Based on the intensity of meteorological driving force and the intensity of production driving force, meteorological attribution weights and production attribution weights are given.

[0015] Furthermore, the pre-construction of the electrical energy model includes: Construct an autoregressive distributed lag model with energy consumption as the autoregressive term and electricity consumption as the distributed lag term; Based on the constraints of autoregressive lag order and distributed lag order, several combinations of autoregressive lag order and distributed lag order are given. Historical training data of user objects is obtained, and combined with the combination scheme and the autoregressive distributed lag model, the model residual variance corresponding to each combination scheme is determined by OLS regression; among which, historical training data includes historical electricity consumption data, historical photovoltaic power output and historical energy consumption. Based on the information criterion method, and combining several combination schemes and the corresponding model residual variance, the optimal autoregressive lag order and distributed lag order of the autoregressive distributed lag model are determined. Based on the optimal autoregressive lag order and distributed lag order, and combined with historical training data, the autoregressive distributed lag model is trained to determine the first model parameters of the autoregressive distributed lag model. Historical meteorological loads, historical production loads, historical training data, and corresponding historical model residuals are obtained. The model residuals are then learned through the Transformer model, and the second model parameters of the Transformer model are given.

[0016] Furthermore, the analysis and processing of photovoltaic power output, meteorological load of electricity users, and production load are used to provide predicted energy consumption values, including: Based on the optimal autoregressive lag order and distributed lag order, the historical electricity consumption data, historical photovoltaic power output, historical energy consumption and corresponding historical model residuals for the predetermined time period before the current moment are determined, and then combined with the current electricity consumption data and photovoltaic power output to form the original prediction data. Based on the autoregressive distributed lag model with determined first model parameters, the original data is analyzed and predicted to provide the predicted energy consumption at the current moment. By determining the parameters of the second model in the Transformer model, the predicted energy consumption, current electricity consumption data, photovoltaic output, meteorological load and production load are predicted, and the model residual at the current moment is determined. The model residuals at the current moment are combined with the predicted energy consumption to determine the final predicted energy consumption value at the current moment.

[0017] Furthermore, based on electricity attribute data, production load, and energy consumption forecasts, carbon emission processing is performed for different emission sources to provide the total carbon emissions of the electricity user, including: Based on the predicted energy consumption and the corresponding carbon emission factor, the direct carbon emissions from energy activities are determined. Based on the difference between current electricity consumption data and photovoltaic output, and combined with the updated real carbon emission factor, determine the indirect carbon emissions from electricity transfer in and out. Based on the production load, and combined with the process ratio and carbon dioxide coefficient in the production process data, the process carbon emissions are determined. The total carbon emissions of electricity users are given by combining the direct carbon emissions from energy activities, the indirect carbon emissions from electricity imports and exports, and the carbon emissions from processes.

[0018] Furthermore, updates to the actual carbon emission factor for electricity include: Obtain the electrical carbon feature vector of the power node where the electricity user is located, as well as the transmission power and weight between the power node and each upstream node; Based on a pre-trained graph neural network model, the carbon emission feature vector of the power node where the power user is located, the transmission power and weight between the power node and each upstream node, and the carbon emission factor of the power node after a predetermined number of propagation rounds are updated to determine the carbon emission factor of the power node after a predetermined number of propagation rounds. The actual carbon emission factor of the electricity user is determined based on the total electricity consumption of the power node and the green electricity consumption of all electricity users within its range, combined with the carbon emission factor of the power node after the predetermined round of propagation.

[0019] Secondly, the present invention also provides a carbon emission monitoring device based on real-time electricity consumption data, employing the above-mentioned carbon emission monitoring method based on real-time electricity consumption data, the device comprising: The data acquisition module is used to acquire the power attribute data of the electricity user; the power attribute data includes electricity consumption data, photovoltaic output and production process data. The load partitioning module is used to sequentially perform multi-scale load partitioning on electricity consumption data to determine the meteorological load and production load of the electricity users. The energy consumption prediction module is used to analyze and process photovoltaic power output, meteorological load of electricity users and production load based on a pre-built power model, and give the predicted value of energy consumption. The carbon emission determination module is used to process carbon emissions from different emission sources based on electricity attribute data, production load and energy consumption forecasts, and to give the total carbon emissions of the electricity user.

[0020] Furthermore, the load sharing module is used for: The current electricity consumption data is decomposed into load characteristics to determine the trend characteristics, periodic characteristics, and residual characteristics. The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. The characteristics of the trend cycle term are decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. Based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load.

[0021] Furthermore, the load sharing module is used for: The system obtains the current ambient temperature, preset cooling temperature, and preset heating temperature, and combines these with seasonal characteristics to determine the initial meteorological load and non-meteorological load at the current moment. Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load; Determine the meteorological and production attribution weights at the current moment, and combine the initial meteorological load, non-meteorological load, initial production load, and non-production load to give the final meteorological load and production load.

[0022] Furthermore, the load sharing module is used for: Obtain the current ambient temperature, preset cooling temperature, and preset heating temperature to determine the initial meteorological load at the current moment; Obtain pre-trained meteorological regression models for seasonal features, meteorological load, and non-meteorological load; Based on the difference between the meteorological regression model and the initial meteorological load at the current time, the non-meteorological load at the current time is given; Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load, including: Obtain the number of production load patterns for user objects and determine the number of clusters. Using the initial production load as the centroid and combining the cluster number, cluster analysis is performed on the short-term fluctuation components at the current moment to give the initial production load; The non-production load is determined based on the difference between the trend cycle term characteristics and the initial production load.

[0023] Furthermore, the load sharing module is also used for: Determine the historical temperature standard deviation and, in conjunction with the current ambient temperature and the preset thermal neutral temperature, give the intensity of the meteorological driving force; Determine the standard deviation of historical capacity intensity, and combine the current capacity intensity with the historical average capacity intensity to give the intensity of production driving force; Based on the intensity of meteorological driving force and the intensity of production driving force, meteorological attribution weights and production attribution weights are given.

[0024] Furthermore, the energy consumption prediction module is used for: Construct an autoregressive distributed lag model with energy consumption as the autoregressive term and electricity consumption as the distributed lag term; Based on the constraints of autoregressive lag order and distributed lag order, several combinations of autoregressive lag order and distributed lag order are given. Historical training data of user objects is obtained, and combined with the combination scheme and the autoregressive distributed lag model, the model residual variance corresponding to each combination scheme is determined by OLS regression; among which, historical training data includes historical electricity consumption data, historical photovoltaic power output and historical energy consumption. Based on the information criterion method, and combining several combination schemes and the corresponding model residual variance, the optimal autoregressive lag order and distributed lag order of the autoregressive distributed lag model are determined. Based on the optimal autoregressive lag order and distributed lag order, and combined with historical training data, the autoregressive distributed lag model is trained to determine the first model parameters of the autoregressive distributed lag model. Historical meteorological loads, historical production loads, historical training data, and corresponding historical model residuals are obtained. The model residuals are then learned through the Transformer model, and the second model parameters of the Transformer model are given.

[0025] Furthermore, the energy consumption prediction module is also used for: Based on the optimal autoregressive lag order and distributed lag order, the historical electricity consumption data, historical photovoltaic power output, historical energy consumption and corresponding historical model residuals for the predetermined time period before the current moment are determined, and then combined with the current electricity consumption data and photovoltaic power output to form the original prediction data. Based on the autoregressive distributed lag model with determined first model parameters, the original data is analyzed and predicted to provide the predicted energy consumption at the current moment. By determining the parameters of the second model in the Transformer model, the predicted energy consumption, current electricity consumption data, photovoltaic output, meteorological load and production load are predicted, and the model residual at the current moment is determined. The model residuals at the current moment are combined with the predicted energy consumption to determine the final predicted energy consumption value at the current moment.

[0026] Furthermore, the carbon emission determination module is used for: Based on the predicted energy consumption and the corresponding carbon emission factor, the direct carbon emissions from energy activities are determined. Based on the difference between current electricity consumption data and photovoltaic output, and combined with the updated real carbon emission factor, determine the indirect carbon emissions from electricity transfer in and out. Based on the production load, and combined with the process ratio and carbon dioxide coefficient in the production process data, the process carbon emissions are determined. The total carbon emissions of electricity users are given by combining the direct carbon emissions from energy activities, the indirect carbon emissions from electricity imports and exports, and the carbon emissions from processes.

[0027] Furthermore, the carbon emission determination module is also used for: Obtain the electrical carbon feature vector of the power node where the electricity user is located, as well as the transmission power and weight between the power node and each upstream node; Based on a pre-trained graph neural network model, the carbon emission feature vector of the power node where the power user is located, the transmission power and weight between the power node and each upstream node, and the carbon emission factor of the power node after a predetermined number of propagation rounds are updated to determine the carbon emission factor of the power node after a predetermined number of propagation rounds. The actual carbon emission factor of the electricity user is determined based on the total electricity consumption of the power node and the green electricity consumption of all electricity users within its range, combined with the carbon emission factor of the power node after the predetermined round of propagation.

[0028] The present invention provides a carbon emission monitoring method and apparatus based on real-time electricity consumption data, which has at least the following beneficial effects: (1) By dividing the load into multiple scales to determine meteorological load and production load, accurate component analysis of complex electricity loads can be achieved, solving the problem of treating the television as a "black box" in traditional methods. By constructing an energy model and combining meteorological load and production load to determine the predicted value of energy consumption, the finely decomposed load components can be used as key feature inputs, realizing a paradigm shift from "extensive total prediction" to "mechanism-driven prediction" and improving prediction accuracy. By combining multiple data, the total carbon emissions of electricity users can be determined, enabling accurate prediction of carbon emissions.

[0029] (2) The total load is decomposed into three components: trend, periodicity and residual through multi-scale load decomposition. The decomposition fit is high and significantly better than the traditional regression decomposition method. It is also robust to irregular disturbances such as abnormal production events and equipment failures, avoiding the distortion effect of outliers on the decomposition results. Each decomposition component has a clear physical meaning (long-term production capacity trend, periodic production law, random disturbance), providing a clear explanatory basis for subsequent analysis.

[0030] (3) By linking meteorological regression and production clustering in a dual-mode approach, and combining dynamic determination of meteorological and production attribution weights, the accuracy of separating meteorological load from production load can be significantly improved. Furthermore, it can automatically adapt to changes in the interaction between environmental regulation and production load under different seasons and production modes, without requiring manual parameter adjustments. For complex scenarios such as seasonal transitions and production adjustment periods, a dynamic weighting mechanism enables reasonable load redistribution, avoiding attribution errors caused by simple rules.

[0031] (4) The autoregressive distributed lag model can clearly depict the dynamic lag relationship between electricity consumption and energy consumption, and accurately and synchronously output the consumption forecasts of various energy sources (coal, gas, oil, etc.). The Transformer model can effectively capture the complex nonlinear patterns and long-range dependencies in the energy system, and reduce the prediction error of the autoregressive distributed lag model. Attached Figure Description

[0032] Figure 1 A flowchart of a carbon emission monitoring method based on real-time electricity consumption data provided by the present invention; Figure 2 A schematic diagram of the carbon emission monitoring method provided in a certain embodiment of the present invention; Figure 3 A flowchart for determining the meteorological load and production load of an electricity user is provided in one embodiment of the present invention; Figure 4 A flowchart for determining meteorological load and production load is provided for one embodiment of the present invention; Figure 5 A flowchart for constructing an electrical energy model is provided in one embodiment of the present invention; Figure 6 A flowchart for providing predicted energy consumption values ​​is provided as an embodiment of the present invention; Figure 7 A flowchart for determining the total carbon emissions of an electricity user, provided as an embodiment of the present invention; Figure 8 A flowchart for updating the real electrical carbon emission factor is provided in one embodiment of the present invention; Figure 9 This is a schematic diagram of a carbon emission monitoring device based on real-time electricity consumption data provided by the present invention. Detailed Implementation

[0033] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0035] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0036] like Figure 1 and Figure 2 As shown, the present invention provides a carbon emission monitoring method based on real-time electricity consumption data, which may include: Acquire the electricity attribute data of electricity users; where electricity users can be enterprises, and the electricity attribute data includes electricity consumption data, photovoltaic output and production process data; electricity consumption data includes the power measured / recorded by the user's main meter, etc., and production process data includes the user's process ratio, carbon emission coefficient, green certificate holdings, etc. The electricity attribute data of the electricity user is data related to carbon emissions, and may also include operating parameters, equipment parameters, energy ledger data, etc. Operating parameters may include ambient temperature, production capacity intensity, etc., and energy ledger data includes historical time series data of various energy consumption, etc. The electricity consumption data is sequentially divided into multi-scale load segments to determine the meteorological load and production load of the electricity users. Based on a pre-built power model, the photovoltaic output, meteorological load of electricity users, and production load are analyzed and processed to provide a predicted value of energy consumption. Based on electricity attribute data, production load and energy consumption forecasts, carbon emission processing is performed on different emission sources to provide the total carbon emissions of electricity users.

[0037] After acquiring the power attribute data of user objects, large model techniques can be used to perform preprocessing operations such as cleaning, formatting, and standardization on the acquired power attribute data. This can eliminate outliers and missing values, ensure data quality, and avoid noise interference in subsequent analysis. Simultaneously, multi-source heterogeneous data (such as electricity consumption data and equipment parameters) can be aligned to a unified time granularity, laying the foundation for multi-dimensional analysis. It also eliminates dimensional differences, making different data (such as electricity consumption data and temperature data) comparable on a unified scale. This provides high-precision and highly consistent input data. The cleaning, formatting, and standardization processes for power attribute data can employ conventional preprocessing steps; achieving the desired preprocessing effect is sufficient and will not be elaborated upon here.

[0038] like Figure 3 As shown, the electricity consumption data is sequentially divided into multi-scale load segments to determine the meteorological load and production load of the electricity users, including: The current electricity consumption data is decomposed into load characteristics to determine the trend characteristics, periodic characteristics, and residual characteristics. The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. The characteristics of the trend cycle term are decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. Based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load.

[0039] In one embodiment, load decomposition is performed on the electricity consumption data at the current moment to determine trend term characteristics, periodic term characteristics, and residual term characteristics, including: using the Robust-STL time series decomposition method to perform load decomposition on the electricity consumption data at the current moment and give trend term characteristics, periodic term characteristics, and residual term characteristics; The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. This includes using the Henderson moving average method to decompose the periodic features by moving average and give the trend cycle features, seasonal features, and random features. The trend cycle feature is decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. This includes using the Hodrick-Prescott smoothing filter method to decompose the trend cycle feature into long-term and short-term components and give the long-term trend component and the short-term fluctuation component.

[0040] In practical applications, the Robust-STL time series decomposition method and the Henderson moving average decomposition method are used to achieve multi-scale load partitioning. Specifically, the Robust-STL time series decomposition method employs a smoothing decomposition method based on local weighted regression. This smoothing decomposition method utilizes the adaptive characteristics of nonparametric regression, performing weighted least squares fitting within the local sample neighborhood to extract three components: trend, period, and residual, respectively, obtaining trend, period, and residual features. This invention, using the Robust-STL time series decomposition method, can maintain strong smoothness and robustness under nonlinear data.

[0041] When performing load decomposition on the current electricity consumption data, let's assume the current time is t, and the electricity consumption data (the measured power of the main meter) is P. t Then the entire timing sequence can be represented as P t =T t +S t +R t Among them, T t The trend term represents the baseline load that varies over long periods of time; S tIt is a periodic term, reflecting regular loads such as periodic production, lighting, and environmental control (air conditioning); R t This is the residual term, which includes irregular disturbances and measurement noise.

[0042] The detailed steps for load decomposition using the Robust-STL timing decomposition method are as follows: Step 1: Trend Removal. Initialize trend components. In the k-th iteration, the influence of the trend from the previous round is first eliminated; Obtain the detrended sequence P' t For use in subsequent cycle analysis; Step 2: Smoothing Periodic Subsequences. Aggregate sample points at the same position in each period to form separate subsequences. Assuming the natural period of the data is np, group sample points of the same phase into np subsequences. Perform locally weighted regression on each subsequence using the LOSSES function with a smoothing parameter of ns. Then concatenate and extend the smoothed results to obtain a temporary periodic series, denoted as […]. ; Step 3: Low-pass filtering. This applies to the temporary periodic components. To perform low-pass filtering, a moving average with a window length of np is first applied, followed by a second moving average with a window length of np, and finally a third moving average with a window length of 3. Then, a LOESS regression with a smoothing parameter of ns is used to smooth the results, thus obtaining the low-pass component. ; Step 4: Periodic component purification. After removing low-frequency components, the net periodic component is obtained: ;

[0043] Step 5: Trend Reconstruction. The net periodic component is extracted from the original sequence, and trend smoothing is performed to obtain the trend term characteristics: ; where n t Set the width of the trend smoothing window.

[0044] The Henderson moving average method decomposes the periodic component after Robust-STL time series decomposition into three parts: a trend cyclical feature, a seasonal feature (seasonal fluctuation), and a random feature (irregular disturbance). The mathematical expression of this additive model is as follows: S t =H t +C t +E t H t To represent the characteristics of the trend cycle term after Henderson smoothing, reflecting the long-term direction of change; C tTo represent the characteristics of seasonal terms, it describes regular fluctuations that repeat within a period length p; E t It represents random or irregular terms and is used to characterize short-term disturbances and measurement errors.

[0045] The Hodrick-Prescott smoothing filter method decomposes the trend cycle features obtained from the Henderson moving average method, extracting the trend signal (long-term trend component) representing long-term changes. t The remaining portion is considered as a cyclical fluctuation component (short-term fluctuation component) A. t Satisfying: H t =L t +A t L t For the separated long-term trend component; A t The Hodrick-Prescott smoothing filter method identifies the separated long-term trend component by minimizing its loss function. Then, by minimizing the objective function J, the optimal long-term trend component estimate is obtained, and the corresponding cyclical fluctuation component is determined. The objective function J satisfies the following:

[0046] Among them, the first item The second term represents the fitting error between the trend and the original sequence (trend cycle feature); The second-order difference penalty term for the trend is used to suppress excessive bending; λ is a smoothing coefficient, the larger its value, the smoother the trend curve, and the less cyclical fluctuations A. t The more significant, λ The value is 14400.

[0047] The optimal long-term trend component estimate is determined by minimizing J. The corresponding periodic fluctuation term is: .

[0048] like Figure 4 As shown, based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load, including: The system acquires the current ambient temperature, preset cooling temperature, and preset heating temperature, and, in conjunction with seasonal characteristics, determines the initial meteorological load and non-meteorological load for the current moment. This includes: acquiring the current ambient temperature, preset cooling temperature, and preset heating temperature to determine the initial meteorological load for the current moment; acquiring a pre-trained meteorological regression model for seasonal characteristics, meteorological load, and non-meteorological load; and, based on the difference between the meteorological regression model and the initial meteorological load for the current moment, providing the non-meteorological load for the current moment. Cluster analysis is performed on the short-term fluctuation components at the current moment to determine the initial production load and non-production load. This includes: obtaining the number of production load patterns for user objects and determining the number of clusters; using the initial production load as the centroid and combining it with the number of clusters, cluster analysis is performed on the short-term fluctuation components at the current moment to determine the initial production load; and the non-production load is determined based on the difference between the trend cycle term characteristics and the initial production load. The number of production load patterns refers to the number of different production load patterns possessed by the user object. For example, production load patterns can be simply divided into shift information (morning, afternoon, evening shifts, etc.). Determine the meteorological and production attribution weights at the current moment, and combine the initial meteorological load, non-meteorological load, initial production load, and non-production load to give the final meteorological load and production load. Among them, the meteorological load is the air conditioning load, which is the part of the electrical load used by the electricity user to drive the HVAC system to regulate the indoor ambient temperature and humidity.

[0049] The determination of the current meteorological and production attribution weights includes: Determine the historical temperature standard deviation and, in conjunction with the current ambient temperature and the preset thermal neutral temperature, give the intensity of the meteorological driving force; Determine the standard deviation of historical capacity intensity, and combine the current capacity intensity with the historical average capacity intensity to give the intensity of production driving force; Based on the intensity of meteorological driving force and the intensity of production driving force, meteorological attribution weights and production attribution weights are given.

[0050] In practical applications, when determining meteorological load and production load, the main method is to use the adaptive mechanism of component attribution driven by both meteorology and production, which mainly includes meteorological regression model and periodic clustering model. The meteorological regression model is a meteorological regression model applied to the seasonal features obtained by the Henderson moving average method. The calculation formula is as follows:

[0051] In the formula: β0, β1, β2, and β3 are all regression coefficients. Specifically, β0 is the regression intercept, which represents the baseline load level in the seasonal term when there is no cooling / heating demand and the humidity is at the baseline state; β1 is the regression coefficient for cooling degree-days, which represents the incremental impact of a unit cooling degree-day (for every 1°C increase) on the seasonal characteristic load Ct; β2 is the regression coefficient for heating degree-days, which represents the incremental impact of a unit heating degree-day (for every 1°C decrease) on the seasonal characteristic load Ct; and β3 is the regression coefficient for the humidity term, which represents the impact of a unit change in humidity on the seasonal characteristic load Ct. t The influence of regression coefficients; D c,t To keep cool and meet daily needs ;D h,tTo keep warm and meet daily needs ;T t,env The ambient temperature (in °C) is the observed value at time point t. t,cool T is the preset cooling temperature (cooling equilibrium point temperature). t,heat The preset heating temperature (heating equilibrium point temperature); The measured ambient humidity value at time point t; This represents the random error term of the model.

[0052] Using the above meteorological regression model, the seasonal term C, smoothed by the Henderson moving average method, can be... t The portion of the data that is clearly related to meteorological conditions (mainly temperature) is separated out, which is the meteorological load extracted at time point t. ,satisfy:

[0053] The non-meteorological load in the seasonal characteristics satisfies:

[0054] In the formula, All are coefficient estimates obtained from regression fitting; For meteorological load, C other,t This is a non-meteorological load; The periodic clustering model performs K-means clustering analysis on the periodic fluctuation components (short-term fluctuation components) after applying the Hodrick-Prescott smoothing filter. The K-means clustering analysis formula is as follows:

[0055] In the formula, k is the preset number of clusters; C i Let μ be the i-th cluster, containing a set of time points with similar periodic fluctuation patterns. i Let be the centroid of the i-th cluster, representing the average value of all periodic fluctuation sequence points within that cluster, reflecting a typical production load pattern. A t Let be the value of the periodic fluctuation sequence at time point t; For data point A t to the center of mass μ i The squared Euclidean distance. The clustering results can be matched with shift information in the user's enterprise survey form. Therefore, the initial production load can be obtained. ,satisfy: ; Non-productive load H in trend cycle term t,other for:

[0056] To avoid mutual interference between meteorological load and production load during the transition period, this invention constructs a dynamic attribution weight function model to achieve the allocation of non-production load and non-meteorological load, specifically including: Define a weight α t ∈[0,1], used for dynamically allocating load assignments in the mixed components, satisfying the following relationship:

[0057] In the formula, α t For meteorological attribution weights, α t When α is 1, it means that at time t, non-production loads and non-meteorological loads should be preferentially attributed to air conditioning. t When W is 0, it means that non-production loads and non-meteorological loads at time t should be preferentially attributed to production. weather,t For the intensity of the meteorological driving force, satisfy: W weather,t =|T t -T neutral | / σ T T t Let T be the ambient temperature at time t. neutral | represents the thermal neutral temperature (typically the midpoint of the 18-22°C range), σ T W represents the standard deviation of temperature during the statistical period. production,t For the production driving force strength, the following condition must be met: W production,t =|P intensity,t -P 平均 | / σ P P intensity,t Let P be the production capacity intensity at time t. 平均 σ represents the average production capacity intensity. P γ represents the standard deviation of production capacity intensity; γ is the sensitivity coefficient (γ>0), which controls the steepness of the weighting function and can be obtained through training with historical data.

[0058] The mixed or unassigned loads (non-production loads and non-meteorological loads) identified by the meteorological regression model and the periodic clustering model are then assigned according to weight α. t Dynamic reallocation is performed while satisfying the following relationship:

[0059] In the formula, Let t be the meteorological load at time t. Let t be the production load at time t.

[0060] Both the final meteorological load and the production load consist of three parts. Taking the meteorological load as an example, the meteorological sensitive load (initial meteorological load) is directly extracted from the seasonal characteristics. The non-meteorological components of the seasonal component are all weighted by α. t The proportion allocated to air conditioning is weighted by α for the non-production portion of the trend cycle term.t The proportion classified as air conditioning.

[0061] like Figure 5 As shown, when forecasting energy consumption, it is necessary to first construct an electrical energy model, which may include: Construct an autoregressive distributed lag model with energy consumption as the autoregressive term and electricity consumption as the distributed lag term, satisfying the following relationship:

[0062] In the formula, EC t Let α be the energy consumption of a user at time t, α be the intercept term of the autoregressive distributed lag model representing the baseline energy consumption level, T be the time trend variable used to capture long-term systematic changes in energy consumption due to technological advancements, energy efficiency improvements, etc., β be the coefficient of the time trend term T, measuring the marginal impact of the time trend on energy consumption, and EC be the energy consumption of a user at time t. t-i This represents the energy consumption at a lag of period i, i.e., the historical energy consumption. For the corresponding EC t-i The autoregressive coefficient; EL t-j Indicates lag j Electricity consumption during the period; θ j For the corresponding EL t-j The distributed lag coefficient, ε t Let p and q be the random error term, and p and q be the autoregressive lag order and the distributed lag order, respectively. Based on the constraints of the autoregressive lag order and the distributed lag order, several combinations of the autoregressive lag order and the distributed lag order are given. The constraints for the autoregressive lag order and the distributed lag order are the maximum autoregressive lag order and the maximum distributed lag order, respectively, and both can be ≤5. For example, when both the maximum autoregressive lag order and the maximum distributed lag order are 4, there are 20 combinations. The minimum value of the autoregressive lag order is 1, and the minimum value of the distributed lag order is 0. Historical training data of user objects is obtained, and combined with the combination scheme and the autoregressive distributed lag model, the model residual variance corresponding to each combination scheme is determined by OLS (Ordinary Least Squares) regression. The historical training data includes historical electricity consumption data, historical photovoltaic power output and historical energy consumption. When processing by OLS regression, the calculation can be simplified, for example, by omitting the intercept term. Based on the information criterion method, and combining several combination schemes and the corresponding model residual variance, the optimal autoregressive lag order and distributed lag order of the autoregressive distributed lag model are determined. Based on the optimal autoregressive lag order and distributed lag order, and combined with historical training data, the autoregressive distributed lag model is trained to determine the first model parameters of the autoregressive distributed lag model. Historical meteorological loads, historical production loads, historical training data, and corresponding historical model residuals are obtained. The model residuals are then learned through the Transformer model, and the second model parameters of the Transformer model are given.

[0063] Among them, the electricity consumption of the autoregressive distributed lag model can be determined by electricity consumption data and photovoltaic output, satisfying the following: , The actual net power consumption of the object at time t is the electricity consumption at time t. This represents the electricity consumption data at time t, i.e., the active power recorded by the main meter. Let be the photovoltaic power output at time t; based on the actual net power consumption, the electricity consumption EL at different times can be determined. t-j .

[0064] This invention constructs a corresponding autoregressive distributed lag model and a Transformer model for different energy sources. When constructing the model, historical energy consumption data is used for each energy source, and historical electricity consumption data and historical photovoltaic output data are used for the overall electricity consumption data, thereby enabling the prediction of energy consumption from the overall level to each energy source.

[0065] like Figure 6 As shown, the analysis and processing of photovoltaic power output, meteorological load of electricity users, and production load are presented to provide predicted energy consumption values, including: Based on the optimal autoregressive lag order and distributed lag order, the historical electricity consumption data, historical photovoltaic power output, historical energy consumption, and corresponding historical model residuals for a predetermined time period before the current moment are determined, and these are combined with the current electricity consumption data and photovoltaic power output to form the original prediction data; wherein, the predetermined time period is determined by the autoregressive lag order and distributed lag order. Based on the autoregressive distributed lag model with determined first model parameters, the original data is analyzed and predicted to provide the predicted energy consumption at the current moment. By determining the parameters of the second model in the Transformer model, the predicted energy consumption, current electricity consumption data, photovoltaic output, meteorological load and production load are predicted, and the model residual at the current moment is determined. The model residuals at the current moment are combined with the predicted energy consumption to determine the final predicted energy consumption value at the current moment.

[0066] In practical applications, energy models are used to analyze and process photovoltaic power output, meteorological loads of electricity users, and production loads to provide predicted energy consumption values, which may include: Based on historical energy consumption data, an ARDL-Transformer model is constructed, which pre-builds the energy model. The corresponding energy model includes an autoregressive distributed lag model (ARDL model) and a Transformer model. The historical energy consumption data considers the real-time output data of the user's photovoltaic system. Since photovoltaic power generation directly participates in the enterprise's internal energy supply, the photovoltaic power generation needs to be allocated to productive electricity consumption segments according to time periods to obtain a net electricity consumption that more accurately reflects actual production activities. This yields the final ARDL model's electricity consumption input, i.e., the electricity consumption, which can be calculated as follows:

[0067] In the formula, The actual net power consumption of the power-consuming object at time t; This represents the active power (electricity consumption data) recorded by the main meter at time t. Let be the photovoltaic power output at time t; There is a close dynamic coupling mechanism between energy consumption and electricity demand during the operation of electricity-consuming entities. To accurately characterize this interaction and simultaneously consider the inertia of energy use and the lag characteristics of electricity consumption behavior, an autoregressive distributed lag (ARDL) model is adopted. The mathematical expression of the ARDL model is as follows:

[0068] In the formula, EC t Let α be the energy consumption of a user at time t, representing the baseline energy consumption level. Let T be the time trend variable, used to capture long-term systematic changes in energy consumption due to technological advancements and energy efficiency improvements. Let β be the coefficient of the time trend term T, measuring the marginal impact of the time trend on energy consumption. t-i This represents the energy consumption corresponding to a lag of i periods, i.e., historical energy consumption. For the corresponding EC t-i The autoregressive coefficient; EL t-j Indicates lag j Electricity consumption during the period; θ j For the corresponding EL t-j The distributed lag coefficient, all θ j The set captures the combined impact of current and historical electricity consumption on current energy consumption (i.e., the distributed lag portion of the model), ε t Let be the random error term, and p and q be the autoregressive lag order and the distributed lag order, respectively.

[0069] The optimal lag order is determined using the information criterion method. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are calculated for OLS regression under different (p,q) combinations:

[0070] In the formula, Let be the residual variance estimate for the corresponding ARDL model, and N be the sample size. The (p,q) value that minimizes either AIC or BIC is chosen as the optimal lag order.

[0071] Specifically, this invention employs an efficient two-stage estimation strategy to accurately estimate the parameters of the ARDL model: The first stage of modeling involves pre-setting the maximum order p and q of the autoregressive and distributed lag terms based on the research question. Subsequently, the OLS method is used to regress the complete ARDL model containing these lagged variables to obtain the estimated values ​​of the model parameters and the model residual sequence.

[0072] The second stage of refinement involves identifying the dynamic characteristics of the residual sequence and constructing a Transformer-based deep time series model to capture these complex patterns. Utilizing the deep feature representation and prediction information provided by this nonlinear model, residual predictions are performed on the linear prediction results from the first stage, ultimately yielding optimized mixed model predictions with superior statistical properties. This process effectively controls estimation biases caused by nonlinear dynamic characteristics, significantly improving the prediction accuracy and reliability of the power model in photovoltaic fluctuation scenarios.

[0073] Specifically, the second-stage Transformer model is used to extract nonlinear features from the residual sequence, and its input feature vector X t ,satisfy:

[0074] In the formula, l =[1,PQ], where PQ is the window length of the Transformer model, which can be determined based on p and q. For example, PQ = p + q.

[0075] The Transformer model processes the feature vector Xt and predicts the model residuals, specifically including: (1) Multi-head self-attention mechanism:

[0076] In the formula, Q , K , V These are the query matrix, key matrix, and value matrix, respectively. h The number of heads; W oThis is the output weight matrix. Given a vector, and the input vector X t The dimensions are the same, which can capture the relationship between multiple features.

[0077] (2) Feedforward Network (FFN):

[0078] In the formula, W1 is the weight matrix of the first layer. b1 is the bias vector of the first layer. W2 is the weight matrix of the second layer. b2 is the bias vector of the second layer. (3) Transformer encoder layer:

[0079] In the formula, EncoderLayer(X) is a vector with the same dimension as the input, representing the feature vector after processing by the encoder layer; LayerNorm(X+MultiHead(X,X,X)) is the input vector processed by a residual connection. X The output is added to the output of the multi-head self-attention mechanism and then normalized by LayerNorm to stabilize the training process and accelerate convergence. LayerNorm ( X + FFN ( X Input is connected via residual connection. X The result is added to the output of the feedforward network and then normalized through layer normalization to further stabilize the training process and accelerate convergence. (4) The final output of the Transformer:

[0080] In the formula, This represents the prediction model residuals of the Transformer model.

[0081] The final predicted energy consumption value at the current moment is determined by combining the residuals of the Transformer-based prediction model with an autoregressive distributed lag model, satisfying the following relationship:

[0082] In the formula, This represents the predicted energy consumption at the current time t. This represents the predicted energy consumption at the current time t using an autoregressive distributed lag model.

[0083] The total carbon emissions of the electricity-consuming entity determined in this invention are calculated based on the actual carbon emission factor of the electricity-consuming entity, combining the carbon emissions of various energy types and the carbon emissions of the production process. The actual carbon emission factor is calculated using a graph neural network (GNN) model combined with the inference mechanism corresponding to the carbon flow topology graph, so as to achieve dynamic and spatiotemporal updates of the regional carbon emission factor.

[0084] like Figure 7 As shown, based on electricity attribute data, production load, and energy consumption forecasts, carbon emission processing is performed for different emission sources, and the total carbon emissions of the electricity user are given, including: Based on the predicted energy consumption and the corresponding carbon emission factor, the direct carbon emissions from energy activities (corresponding to the carbon emissions from the first emission source) are determined. Based on the difference between the current electricity consumption data and photovoltaic output, and combined with the updated real carbon emission factor, the indirect carbon emissions from electricity transfer in and out are determined (corresponding to the carbon emissions from the second emission source). Based on the production load, and combined with the process ratio and electrical carbon coefficient in the production process data, the process carbon emissions (corresponding to the carbon emissions from the third emission source) are determined. The total carbon emissions of electricity users are given by combining the direct carbon emissions from energy activities, the indirect carbon emissions from electricity imports and exports, and the carbon emissions from processes.

[0085] like Figure 8 As shown, in practical application scenarios, the update of the real-world carbon emission factor includes: Obtain the carbon feature vector of the power node where the electricity user is located, as well as the transmission power and weight between the power node and each upstream node; where upstream node refers to the node that can transmit power to the current power node; Based on a pre-trained graph neural network model, the carbon emission feature vector of the power node where the power user is located, the transmission power and weight between the power node and each upstream node, and the carbon emission factor of the power node after a predetermined number of propagation rounds are updated to determine the carbon emission factor of the power node after a predetermined number of propagation rounds. The actual carbon emission factor of the electricity user is determined based on the total electricity consumption of the power node and the green electricity consumption of all electricity users within its range, combined with the carbon emission factor of the power node after the predetermined round of propagation.

[0086] The graph neural network model is trained based on the power carbon flow topology graph corresponding to the electricity user. Specifically, it includes: Carbon flow topology modeling: The power grid nodes in the area where the electricity user resides are considered as nodes (power nodes) in the power carbon flow topology graph network, and the transmission lines between nodes are considered as directed edges. A power carbon flow topology graph G=(V,E) is constructed, where V is the set of nodes and E represents the power transmission relationships. The carbon emission feature vector of each node includes: the node's initial carbon emission factor and the number of green certificates held by the node. The number of green certificates held refers to the number of green electricity certificates.

[0087] When determining the carbon emission factor of a power node after a predetermined number of propagation rounds based on a pre-trained graph neural network model, the graph neural network (GNN) defines the carbon flow propagation function between nodes to update and determine the carbon emission factor. Specifically, the graph neural network can adopt a graph convolutional network (GCN) that satisfies the following propagation relationship:

[0088] In the formula, This represents the carbon emission factor of node n after propagation in round t+1. P represents the electrical carbon emission factor at node m after the t-th round of propagation; mn A represents the electrical power transmitted from node m to node n. m Let be the eigenvector of the electric carbon at node m; f θ () represents the trainable weight function of the GNN; σ() represents the non-linear activation function; w mn The weights between nodes m and n can be either power transmission weights or carbon flow weights.

[0089] Carbon emission factors of electricity users' power nodes obtained based on GNN inference By combining the total electricity consumption of a node (the sum of the electricity consumption of all electricity users corresponding to that node) and the green electricity consumption of all electricity users, the carbon factor corresponding to each electricity user can be obtained:

[0090] In the formula, E g E represents the amount of green electricity consumed corresponding to the green certificates held by electricity users within the node area. total α represents the total power consumption of the node. g This represents the emission reduction effect coefficient of green certificates. This is the adjusted carbon emission factor for node n after green certificate adjustments. Green certificate holdings are the rights to acquire green electricity consumption, and the green certificate emission reduction effect coefficient is the actual carbon emission reduction efficiency corresponding to a unit of green certificate electricity consumption.

[0091] The total carbon emissions from electricity consumption include direct carbon emissions from energy activities, indirect carbon emissions from electricity generation, and process carbon emissions, specifically including: Ctotal =C energy +C elec +C huaxue In the formula, C total The total carbon emissions of the electricity-consuming object, i.e., by C energy C elec C huaxue The determined total carbon emissions, C energy For direct carbon emissions from energy activities, C elec C represents indirect carbon emissions from electricity imports and exports. huaxue This refers to carbon emissions from industrial processes. Direct carbon emissions from energy activities are mainly generated by the combustion of fossil fuels, while indirect carbon emissions from electricity generation are mainly generated by the generation of purchased electricity. Industrial carbon emissions are mainly generated by chemical reactions during production. By gaining a deeper understanding of the structure, spatial and temporal distribution, and driving factors of carbon emissions, we can not only obtain accurate total carbon emission data but also provide decision support for precise emission reduction.

[0092] Among them, the direct carbon emissions from energy activities are calculated based on enterprise energy consumption data obtained from the ARDL model. The direct carbon emissions from energy activities are calculated as follows:

[0093] In the formula, E k To predict the consumption of the k-th energy source, the energy consumption of each energy source can be predicted separately during the power model forecasting process. This means each energy source corresponds to an autoregressive distributed lag model, which predicts the consumption using the lag term of each energy source and the lag term of the total electricity consumption (using the overall net power consumption). Alternatively, an autoregressive distributed lag model can be directly used to predict the total energy consumption of all energy sources, and the total energy consumption of all energy sources can be determined by combining this with a predetermined energy mix. For example... , where τ k This represents the predetermined energy ratio for the k-th energy source; f fuel,k Let be the carbon emission factor for the kth energy type.

[0094] The calculation of indirect carbon emissions from electricity imports and exports is determined by multiplying the difference between the total electricity consumption data measured by the main electricity meter and the photovoltaic output by the corrected nodal power carbon emission factor, satisfying the following relationship:

[0095] In the formula, C elec For indirect carbon emissions from electricity imports and exports, P load The electricity consumption data measured by the main meter corresponds to the active power recorded by the main switch meter, P PV The power output of photovoltaic power generation; Process carbon emissions are calculated by multiplying the electricity input by the amount of carbon dioxide generated per kilowatt-hour during the process, satisfying the following relationship:

[0096] In the formula, F m The power consumption of the m-th process stage can be determined by the production load obtained from load decomposition and according to the process ratio corresponding to the process stage. For example... , Let be the electrocarbon coefficient of the m-th process step.

[0097] This invention fundamentally changes the traditional carbon emission accounting paradigm, achieving a technological leap from "static, lagging, and extensive" to "dynamic, real-time, and refined." Specifically, it shortens the carbon emission accounting cycle from the traditional monthly and annual periods to the hourly or even minute-level, enabling carbon flow monitoring synchronized with enterprise production and operations. It can respond in real-time to the instantaneous impact of dynamic factors such as production plan adjustments, equipment start-up and shutdown, and weather changes on carbon emissions. Through multi-level technical means, it significantly reduces systematic errors caused by load aliasing, static factors, and process ambiguity in traditional accounting methods. Furthermore, the accounting granularity can be refined from the enterprise as a whole to process links, production units, and even key equipment, achieving "microscopic" observation of carbon emissions. It establishes a complete and traceable mapping relationship from electricity consumption to energy consumption and then to carbon emissions, enabling direct and quantitative evaluation of the actual effects of emission reduction measures such as technological upgrades, energy substitution, and production scheduling, supporting precise investment decisions. This approach overcomes the limitations of traditional static electricity carbon factors in scenarios with high proportions of renewable energy access, achieving spatiotemporal dynamic perception of the electricity-carbon relationship. Furthermore, through a green certificate correction mechanism, it accurately reflects the real emission reduction contributions of enterprises using green electricity, supporting the healthy development of the green electricity trading market. Further, by determining meteorological and production loads through multi-scale load decomposition, it enables precise component analysis of complex electricity loads, solving the problem of traditional methods treating electricity as a "black box." By constructing an energy model and combining meteorological and production loads to determine predicted energy consumption values, it uses finely decomposed load components as key feature inputs, achieving a paradigm shift from "extensive total prediction" to "mechanism-driven prediction," and improving prediction accuracy. By combining multiple data sources, it determines the total carbon emissions of electricity users, enabling accurate carbon emission prediction. The high-precision results of multi-scale load decomposition serve as high-quality input for energy consumption prediction, and the prediction results, in turn, serve as a reliable basis for carbon accounting, forming a positive accuracy enhancement cycle and reducing overall error. The carbon emission accounting results can provide feedback to verify the rationality of load decomposition and energy consumption prediction, guide model parameter adjustments, achieve a closed-loop self-optimization system, and continuously improve performance. From load insights to energy consumption forecasting and carbon emission accounting, it provides a complete decision-making chain, supporting multi-level decision-making from real-time scheduling to strategic planning. It enables precise load management to reduce energy waste, accurate forecasting to optimize procurement, and precise accounting to lower carbon costs, achieving comprehensive cost reduction and efficiency improvement.

[0098] like Figure 9 As shown, the present invention also provides a carbon emission monitoring device based on real-time electricity consumption data. Employing the aforementioned carbon emission monitoring method based on real-time electricity consumption data, the device includes: The data acquisition module is used to acquire the power attribute data of the electricity user; the power attribute data includes electricity consumption data, photovoltaic output and production process data. The load partitioning module is used to sequentially perform multi-scale load partitioning on electricity consumption data to determine the meteorological load and production load of the electricity users. The energy consumption prediction module is used to analyze and process photovoltaic power output, meteorological load of electricity users and production load based on a pre-built power model, and give the predicted value of energy consumption. The carbon emission determination module is used to process carbon emissions from different emission sources based on electricity attribute data, production load and energy consumption forecasts, and to give the total carbon emissions of the electricity user.

[0099] Furthermore, the load sharing module is used for: The current electricity consumption data is decomposed into load characteristics to determine the trend characteristics, periodic characteristics, and residual characteristics. The periodic features are decomposed by moving average to determine the trend cycle features, seasonal features, and random features. The characteristics of the trend cycle term are decomposed into long-term and short-term components to determine the long-term trend component and the short-term fluctuation component. Based on the long-term trend component and the short-term fluctuation component, the seasonal characteristics are divided to determine the meteorological load and the production load.

[0100] Furthermore, the load sharing module is used for: The system obtains the current ambient temperature, preset cooling temperature, and preset heating temperature, and combines these with seasonal characteristics to determine the initial meteorological load and non-meteorological load at the current moment. Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load; Determine the meteorological and production attribution weights at the current moment, and combine the initial meteorological load, non-meteorological load, initial production load, and non-production load to give the final meteorological load and production load.

[0101] Furthermore, the load sharing module is used for: Obtain the current ambient temperature, preset cooling temperature, and preset heating temperature to determine the initial meteorological load at the current moment; Obtain pre-trained meteorological regression models for seasonal features, meteorological load, and non-meteorological load; Based on the difference between the meteorological regression model and the initial meteorological load at the current time, the non-meteorological load at the current time is given; Cluster analysis is performed on the short-term fluctuation components at the current moment to provide the initial production load and non-production load, including: Obtain the number of production load patterns for user objects and determine the number of clusters. Using the initial production load as the centroid and combining the cluster number, cluster analysis is performed on the short-term fluctuation components at the current moment to give the initial production load; The non-production load is determined based on the difference between the trend cycle term characteristics and the initial production load.

[0102] Furthermore, the load sharing module is also used for: Determine the historical temperature standard deviation and, in conjunction with the current ambient temperature and the preset thermal neutral temperature, give the intensity of the meteorological driving force; Determine the standard deviation of historical capacity intensity, and combine the current capacity intensity with the historical average capacity intensity to give the intensity of production driving force; Based on the intensity of meteorological driving force and the intensity of production driving force, meteorological attribution weights and production attribution weights are given.

[0103] Furthermore, the energy consumption prediction module is used for: Construct an autoregressive distributed lag model with energy consumption as the autoregressive term and electricity consumption as the distributed lag term; Based on the constraints of autoregressive lag order and distributed lag order, several combinations of autoregressive lag order and distributed lag order are given. Historical training data of user objects is obtained, and combined with the combination scheme and the autoregressive distributed lag model, the model residual variance corresponding to each combination scheme is determined by OLS regression; among which, historical training data includes historical electricity consumption data, historical photovoltaic power output and historical energy consumption. Based on the information criterion method, and combining several combination schemes and the corresponding model residual variance, the optimal autoregressive lag order and distributed lag order of the autoregressive distributed lag model are determined. Based on the optimal autoregressive lag order and distributed lag order, and combined with historical training data, the autoregressive distributed lag model is trained to determine the first model parameters of the autoregressive distributed lag model. Historical meteorological loads, historical production loads, historical training data, and corresponding historical model residuals are obtained. The model residuals are then learned through the Transformer model, and the second model parameters of the Transformer model are given.

[0104] Furthermore, the energy consumption prediction module is also used for: Based on the optimal autoregressive lag order and distributed lag order, the historical electricity consumption data, historical photovoltaic power output, historical energy consumption and corresponding historical model residuals for the predetermined time period before the current moment are determined, and then combined with the current electricity consumption data and photovoltaic power output to form the original prediction data. Based on the autoregressive distributed lag model with determined first model parameters, the original data is analyzed and predicted to provide the predicted energy consumption at the current moment. By determining the parameters of the second model in the Transformer model, the predicted energy consumption, current electricity consumption data, photovoltaic output, meteorological load and production load are predicted, and the model residual at the current moment is determined. The model residuals at the current moment are combined with the predicted energy consumption to determine the final predicted energy consumption value at the current moment.

[0105] Furthermore, the carbon emission determination module is used for: Based on the predicted energy consumption and the corresponding carbon emission factor, the direct carbon emissions from energy activities are determined. Based on the difference between current electricity consumption data and photovoltaic output, and combined with the updated real carbon emission factor, determine the indirect carbon emissions from electricity transfer in and out. Based on the production load, and combined with the process ratio and carbon dioxide coefficient in the production process data, the process carbon emissions are determined. The total carbon emissions of electricity users are given by combining the direct carbon emissions from energy activities, the indirect carbon emissions from electricity imports and exports, and the carbon emissions from processes.

[0106] Furthermore, the carbon emission determination module is also used for: Obtain the electrical carbon feature vector of the power node where the electricity user is located, as well as the transmission power and weight between the power node and each upstream node; Based on a pre-trained graph neural network model, the carbon emission feature vector of the power node where the power user is located, the transmission power and weight between the power node and each upstream node, and the carbon emission factor of the power node after a predetermined number of propagation rounds are updated to determine the carbon emission factor of the power node after a predetermined number of propagation rounds. The actual carbon emission factor of the electricity user is determined based on the total electricity consumption of the power node and the green electricity consumption of all electricity users within its range, combined with the carbon emission factor of the power node after the predetermined round of propagation.

[0107] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A carbon emission monitoring method based on real-time electricity consumption data, characterized in that, The method comprises the following steps: obtaining power attribute data of the power consumer; wherein the power attribute data comprises power consumption data, photovoltaic output and production process data; performing multi-scale load division on the power consumption data in sequence to determine the meteorological load and production load of the power consumer; based on the pre-constructed power model, analyzing and processing the photovoltaic output, meteorological load and production load of the power consumer to give the energy consumption prediction value; based on the power attribute data, production load and energy consumption prediction value, processing the carbon emissions of different emission sources to give the total carbon emissions of the power consumer.

2. The carbon emission monitoring method of claim 1, wherein, The method for performing multi-scale load division on the power consumption data in sequence to determine the meteorological load and production load of the power consumer comprises the following steps: performing load decomposition on the power consumption data at the current time to determine the trend item feature, periodic item feature and residual item feature; performing moving average decomposition on the periodic item feature to determine the trend cycle item feature, seasonal item feature and random item feature; performing long-short term decomposition on the trend cycle item feature to determine the long-term trend component and short-term fluctuation component; based on the long-term trend component and short-term fluctuation component, dividing the seasonal item feature to determine the meteorological load and production load.

3. The carbon emission monitoring method of claim 2, wherein, The method for dividing the seasonal item feature based on the long-term trend component and short-term fluctuation component to determine the meteorological load and production load comprises the following steps: obtaining the ambient temperature, preset cooling temperature and preset heating temperature at the current time, and combining the seasonal item feature to determine the initial meteorological load and non-meteorological load at the current time; performing cluster analysis on the short-term fluctuation component at the current time to give the initial production load and non-production load; determining the meteorological attribution weight and production attribution weight at the current time, and combining the initial meteorological load, non-meteorological load, initial production load and non-production load to give the final meteorological load and production load.

4. The carbon emission monitoring method of claim 3, wherein, The method for obtaining the ambient temperature, preset cooling temperature and preset heating temperature at the current time, and combining the seasonal item feature to determine the initial meteorological load and non-meteorological load at the current time comprises the following steps: obtaining the ambient temperature, preset cooling temperature and preset heating temperature at the current time to determine the initial meteorological load at the current time; obtaining the pre-trained meteorological regression model about the seasonal item feature, meteorological load and non-meteorological load; based on the difference between the meteorological regression model and the initial meteorological load at the current time, giving the non-meteorological load at the current time; The method for performing cluster analysis on the short-term fluctuation component at the current time to give the initial production load and non-production load comprises the following steps: obtaining the number of production load modes of the user object to determine the number of cluster clusters; taking the initial production load as the centroid, combining the number of cluster clusters, and performing cluster analysis on the short-term fluctuation component at the current time to give the initial production load; based on the difference between the trend cycle item feature and the initial production load, determining the non-production load.

5. The carbon emission monitoring method of claim 4, wherein, The determination of the meteorological attribution weight and production attribution weight at the current time comprises the following steps: determining the historical temperature standard deviation, and combining the ambient temperature at the current time and the preset thermal neutral temperature to give the meteorological driving force intensity; determining the historical production capacity intensity standard deviation, and combining the production capacity intensity at the current time and the historical average production capacity intensity to give the production driving force intensity; The meteorological attribution weight and the production attribution weight are given based on the meteorological driving force intensity and the production driving force intensity.

6. The carbon emission monitoring method according to any one of claims 1 to 5, characterized in that, The pre-construction of the electricity model includes: An autoregressive distributed lag model is constructed with the energy consumption as the autoregressive term and the electricity consumption as the distributed lag term; Based on the constraint conditions of the autoregressive lag order and the distributed lag order, a number of combination schemes of the autoregressive lag order and the distributed lag order are given; The historical training data of the user object is obtained, and the combination scheme and the autoregressive distributed lag model are combined to determine the model residual variance corresponding to each combination scheme through OLS regression; wherein the historical training data includes historical electricity data, historical photovoltaic output and historical energy consumption; Based on the information criterion method, the optimal autoregressive lag order and the optimal distributed lag order of the autoregressive distributed lag model are determined by combining a number of combination schemes and corresponding model residual variances; Based on the optimal autoregressive lag order and the optimal distributed lag order, the autoregressive distributed lag model is trained based on the historical training data to determine the first model parameters of the autoregressive distributed lag model; The historical meteorological load, historical production load, historical training data and corresponding historical model residuals are obtained, and the model residuals are learned through the Transformer model to give the second model parameters of the Transformer model.

7. The carbon emission monitoring method of claim 6, wherein, The photovoltaic output, meteorological load and production load of the electricity consumption object are analyzed and processed to give the energy consumption prediction value, including: Based on the optimal autoregressive lag order and the optimal distributed lag order, the historical electricity data, historical photovoltaic output, historical energy consumption and corresponding historical model residuals in a predetermined time period before the current time are determined, and the prediction original data is formed with the electricity data and photovoltaic output at the current time; Based on the autoregressive distributed lag model with the first model parameters determined, the prediction original data is analyzed and predicted to give the predicted energy consumption at the current time; The Transformer model with the second model parameters determined is used to predict the predicted energy consumption, electricity data, photovoltaic output, meteorological load and production load at the current time to determine the model residual at the current time; The final energy consumption prediction value at the current time is determined by combining the model residual at the current time with the predicted energy consumption.

8. The carbon emission monitoring method according to any one of claims 1 to 5, characterized by, Based on the electricity attribute data, production load and energy consumption prediction value, the carbon emissions of different emission sources are processed to give the total carbon emissions of the electricity consumption object, including: Based on the energy consumption prediction value of the energy and the corresponding carbon emission factor, the direct carbon emissions of energy activities are determined; Based on the difference between the electricity data and the photovoltaic output at the current time, and combined with the updated real electricity carbon emission factor, the indirect carbon emissions between power import and export are determined; Based on the production load and combined with the process ratio in the production process data and the electricity carbon coefficient, the process carbon emissions are determined; The total carbon emissions of the electricity consumption object are given by combining the direct carbon emissions of energy activities, the indirect carbon emissions between power import and export, and the process carbon emissions.

9. The carbon emission monitoring method of claim 8, wherein, The update of the real electricity carbon emission factor includes: The electricity carbon feature vector of the electricity consumption object is obtained, as well as the transmission power and weight between the electricity consumption object and each upstream node. The power carbon feature vector of the power node where the power consumer is located, the transmission power and weight between the power node and each upstream node, and the power carbon emission factor of the power node when the power node propagates for a predetermined round are updated based on a pre-trained graph neural network model, to determine the power carbon emission factor of the power node after the power node propagates for the predetermined round; According to the total power consumption of the power node and the green power consumption of all power consumers within the range, and in combination with the power carbon emission factor of the power node after the power node propagates for a predetermined round, the real power carbon emission factor of the power consumer is determined.

10. A carbon emission monitoring device based on real-time electricity consumption data, characterized by, The carbon emission monitoring method and device based on real-time power consumption data according to any one of claims 1-9 comprises: A data acquisition module for acquiring power attribute data of the power consumer; wherein the power attribute data includes power consumption data, photovoltaic output, and production process data; A load division module for sequentially performing multi-scale load division on the power consumption data to determine the meteorological load and production load of the power consumer; An energy consumption prediction module for analyzing and processing the photovoltaic output, the meteorological load and the production load of the power consumer based on a pre-constructed power model to give an energy consumption prediction value; A carbon emission determination module for processing carbon emissions of different emission sources based on the power attribute data, the production load, and the energy consumption prediction value to give the total carbon emission of the power consumer.

Images