Electric carbon metering method and system based on causal interpretable graph network

Abstract

The invention discloses an electricity carbon metering method and system based on a causal interpretable graph network, and the method comprises the steps: collecting electric power data, meteorological data and topological data, and carrying out the unified packaging of the three types of data; missing repair, space-time alignment and causal variable construction are carried out on the received uniformly packaged multi-source data, and an aligned power-weather-carbon quantity sequence is generated; and explicitly quantifying marginal carbon responsibilities of different nodes in different time periods based on a power-weather-carbon quantity sequence. According to the invention, a set of causal interpretable graph network-driven electricity carbon metering method is constructed, real-time accurate metering, causal transparent tracking, traceable carbon emission responsibility at any moment and any node and automatic strategy generation are realized, and an industrial and commercial carbon reduction optimization scheme is generated based on causal intervention.

Description

Technical Field

[0001] This invention relates to the field of carbon emission metering technology in power systems, specifically to a method and system for measuring carbon emissions in electricity based on causal interpretable graph networks. Background Technology

[0002] Current industrial and commercial electricity carbon metering mainly employs the emission factor method or the grid average factor method, which assumes that the grid carbon emission factor is a static constant. This fails to reflect the dynamic impact of real-time power structure, regional grid interaction, and user behavior on carbon emissions. Recent research indicates that spatiotemporal graph neural networks can model the spatiotemporal coupling relationship between inter-regional power flow and carbon emissions, but they have the following limitations: 1. Existing carbon emission metering terminals mostly use the static emission factor method, which does not take into account the spatiotemporal coupling of real-time power flow, renewable energy output fluctuations and industrial and commercial loads, resulting in large metering errors.

[0003] 2. Recent deep learning-based carbon metrology research focuses only on features of a single time scale and lacks explicit modeling of the "electricity-carbon" causal mechanism, resulting in poor interpretability.

[0004] 3. Graph neural networks have been used for power grid flow prediction, but they do not incorporate causal variables for carbon emissions and cannot directly output carbon factors.

[0005] In summary, existing technologies suffer from three major shortcomings: inaccurate measurement, unclear causal relationships, and imprecise optimization.

[0006] The invention with patent publication number CN120524385A discloses a method and system for tracing the source of carbon emission flow metering errors in power systems. This method identifies and quantifies error sources in distributed carbon metering and performs uncertainty propagation and abnormal node identification / calibration along the power grid topology. Essentially, it's a "source-evaluation-calibration" measurement framework, but it doesn't introduce causal intervention or a dynamic correction mechanism driven by user-side behavior. The invention patent with patent publication number CN118484666A discloses an evaluation method and system for energy storage power stations with multi-application (source-grid-load) architecture. This patent focuses on multi-scenario evaluation and parameter configuration for energy storage power stations, using GNN combined with LSTM to establish a spatiotemporal correlation / causal matrix, and employing reinforcement learning and Monte Carlo methods to optimize strategies and configurations. Its technical focus is on system-level asset optimization, not user-side carbon metering accuracy control, and it doesn't provide an end-to-end closed-loop mechanism for robust metering errors. While both of these existing technologies involve topology error propagation or graph neural networks, they both rely on statistical correlation modeling and lack causal inference mechanisms and dynamic interpretability. CN120524385A only propagates uncertainty through error link analysis and fails to characterize the true causal relationship between variables; CN118484666A, although it integrates GNN and LSTM, has its parameter training based on historical correlation optimization and cannot theoretically distinguish the influence of confounding variables. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for measuring carbon electricity based on causal interpretable graph networks.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for measuring electrocarbon based on causal interpretable graph networks, comprising: Collect power data, meteorological data, and topology data, and encapsulate these three types of data in a unified manner; The received, uniformly packaged multi-source data is subjected to missing data repair, spatiotemporal alignment, and causal variable construction to generate aligned power-meteorological-carbon sequences. Based on the power-meteorological-carbon sequence, the marginal carbon responsibility of different nodes at different time periods is explicitly quantified.

[0009] In this embodiment, generating the aligned power-weather-carbon sequence includes the following steps: An algorithm based on spatiotemporal tensor completion is used to interpolate the occasionally lost measurement values ​​of power and meteorological data at the minute level. The missing and missing power and meteorological data are combined with the topology data and aligned with the same timestamp and spatial node number to ensure that the electrical quantity and meteorological quantity of any node at any time correspond one-to-one. Based on the power grid flow equation and carbon emission factor model, an "electricity-carbon" coupled feature vector for causal inference is generated to obtain an aligned power-meteorological-carbon sequence.

[0010] In this embodiment, an algorithm based on spatiotemporal tensor completion is used to interpolate occasionally lost measurements in multi-source data at the minute level, applying the following formula: ; ; In the formula, For the multi-source data after interpolation and completion, From arrive Spatiotemporal coupling weights, For nodes With nodes Time distance, For nodes With nodes Time distance, For nodes With nodes Spatial electrical distance, For nodes With nodes Spatial electrical distance, For the pre-defined spatiotemporal domain, , All are discrete-time indices. For a historical moment in the neighborhood, These are the node number and the timestamp, respectively. Here, represents the time decay coefficient and the spatial decay coefficient, and exp is an exponential function. For active power tensor, For reactive power tensor, For node voltage tensors, For current tensor, For the photovoltaic irradiance tensor, For wind speed tensor, This is the ambient temperature tensor.

[0011] In this embodiment, the missing and completed power data and meteorological data are combined with topological data and aligned according to the same timestamp and spatial node number, including: The adjacency matrix in the topological data is spatiotemporally aligned and resampled with the repaired multi-source data; whereby the adjacency matrix is ​​linearly interpolated: ; In the formula, It is an adjacency matrix. For the dispatch center in The original adjacency matrix issued at each time step. For separation recent and ≤ The original topological moment, Leave Recently and > The next original topological time; interval Bundle Sandwiched in the middle, the dynamic adjacency matrix corresponding to that minute is obtained by linear interpolation. ; For the dispatch center in The original adjacency matrix issued at each time step; Variables in power and meteorological data are analyzed based on node-time pairs. spliced ​​tensors With the corresponding adjacency matrix Pack the data into dictionaries to generate a global tensor that corresponds one-to-one with the electrical and meteorological quantities at any given time for any node; including: tensor Expand into a list of node-time pairs: ;in, For nodes At any moment eigenvectors; For each node-time pair , its adjacency vector is extracted from the adjacency matrix : ; in the formula, is the node , the admittance value topology connection at time ; All node-time pairs are sorted, and the electrical quantity and meteorological quantity corresponding to any node at any time are matched to generate the aligned overall tensor .

[0012] In this embodiment, the aligned power-meteorological-carbon quantity sequence is obtained, and the steps include: First, the node injection power is calculated by using the power grid flow equation: ; In the formula, is the node , the complex power at time , is the complex conjugate, is the node , the complex voltage at time , is the node , the complex voltage at time , is the node , and the admittance value between the node at time ; Then, the real part of the node complex power is taken and multiplied by the real-time regional carbon emission factor to obtain the carbon emission amount by using the following formula: ; In the formula, is the estimated carbon emission amount of the node at time , is the real part of the node complex power, is the real-time regional carbon emission factor; Finally, the power data, meteorological data, node power and carbon emission amount are spliced to obtain the power-meteorological-carbon quantity sequence.

[0013] In this embodiment, an external structured knowledge graph is introduced, and the missing edge or node relationship in the causal relationship is completed by using a knowledge graph embedding method, including: The external knowledge graph is , wherein, is an entity set, is a relationship set; The TransE-style embedding method is used to treat entities and relations in the external knowledge graph as vectors, such that the sum of the head entity and relation approximately equals the sum of the tail entities. For each triple... The learning vector satisfies: ; Define the causal completion score function as follows: ; In the formula, For causal completion scoring function, For head entity Embedded vector, Tail entity Embedded vector, , For relationship Learnable parameters The function is a sigmoid function, and its output is the probability of the existence of causal edges. This is the matrix transpose symbol; Based on the probability of the existence of causal edges output by the causal completion score function, construct the completed adjacency matrix: ; In the formula, To complete the adjacency matrix, This is the original adjacency matrix. To set a threshold, For logical OR.

[0014] In this embodiment, the marginal carbon responsibility of different nodes at different time periods is explicitly quantified, including: Map the causally aligned power-meteorological-carbon sequence to a high-dimensional potential causal representation; Based on the high-dimensional potential causal representation, and using a dynamic adjacency matrix, the instantaneous and delayed electrocarbon coupling relationships between nodes are captured. Based on the instantaneous and delayed electrocarbon coupling relationship between nodes, the front-door criterion of causal recommendation is adopted to identify causal effects in the presence of confounding factors. For the identified causal effects, attention weights are used to backtrack the critical path, outputting the real-time carbon factor and visual attribution map for each node; and based on the real-time carbon factor of each node, energy-saving and energy storage optimization schemes are generated and pushed.

[0015] In this embodiment, obtaining a high-dimensional potential causal representation includes the following steps: ; ; In the formula, , These are the trainable weights and biases of the first layer, respectively. For Swish activation function, , These represent the trainable weights and biases of the second layer, respectively. For the first layer of potential representation, For the final potential causal characterization; Specifically, attention embedding is used for the final potential causal representation. The adjacency matrix is ​​explicitly embedded into a multi-head attention mechanism, allowing information to be passed only between nodes with causal relationships, blocking non-causal paths, thereby enhancing the accuracy and interpretability of causal reasoning. ; ; In the formula, , , , These are the trainable projection matrices for the query, key, and value in the attention process, respectively. For element-wise multiplication, For a moment Causal mask of time For a moment Time node , Causal mask between them To complete the adjacency matrix, The dimension of the query and key vector.

[0016] In this embodiment, based on a dynamic adjacency matrix, the instantaneous and delayed electrocarbon coupling relationships between nodes are captured, including: Based on the adjacency matrix, first use neighbor aggregation: x Then, using spacetime gate fusion, the instantaneous and delayed electrocarbon coupling relationships between nodes are obtained: In the formula, For each element in the adjacency matrix, it is represented as a node. , The per-unit electrical admittance between them , The first The layers can be trained with weights; the former processes neighbor information, while the latter processes its own history. To represent the representation of the same node at the previous time step, it is obtained from the output buffer of the previous layer. The GRU is a gated recurrent unit, and its output is... This represents the instantaneous and delayed electrocarbon coupling relationship between nodes. For the first Layer, moment Middle node The neighbor aggregated message vector, , For each node , At any moment The representation of time.

[0017] In this embodiment, the front-door criterion for causal recommendation is used to identify causal effects in the presence of confounding factors, including: Counterfactual construction: ; Carbon emissions readout: ; Display the calculated marginal responsibility factor: ; In the formula, For nodes exist Photovoltaic irradiance at any given time The proportion of human-induced disturbances. This is represented as a value range of {-0.3, -0.15, 0, 0.15, 0.3}, used only for a single forward computation, and does not retain gradients. The photovoltaic irradiance after applying a virtual photovoltaic perturbation, To refeed the photovoltaic irradiance after applying a virtual photovoltaic perturbation into the same convolutional network, the resulting counterfactual representation is... This represents the counterfactual carbon emissions after photovoltaic perturbation. This represents the real-time regional carbon emission factor. For activation function, For linear readout weights, This is the matrix transpose symbol. The sampling interval is a fixed constant. This serves as the baseline zero value for all subsequent perturbation comparisons. This can be expressed as replacing the average difference with a differential approximation, thus obtaining the instantaneous marginal carbon responsibility. The marginal responsibility coefficient characterizes the identification of causal effects in the presence of confounding factors.

[0018] In this embodiment, the generation and push of energy-saving and energy storage optimization schemes are obtained through the following steps: The instantaneous and delayed electrocarbon coupling relationships between nodes are used to generate variable-attribution weights for readability and visualization: In the formula, For variable-attribution weight, It is an exponential function. This is a trainable projection matrix, where each row corresponds to one original variable channel. The instantaneous and delayed electrocarbon coupling relationship between nodes; Based on marginal responsibility coefficients and variable-attribution weights, a three-dimensional action of "load shifting - energy storage - photovoltaic" is generated online to achieve energy saving and quantifiable carbon emission reduction. The continuous actions are: ; Objective function: ; In the formula, In order to increase the proportion of photovoltaic capacity, The input features are the marginal responsibility coefficient and the variable-attribution weight concatenation. , These represent the weights and biases of the last layer of the policy network. For element-wise tangent, The three rows are independently mapped to load translation π1, energy storage charging and discharging π2, and photovoltaic capacity expansion π3, generating three-dimensional actions simultaneously in one forward pass; Represented as three hyperparameters, Let be the objective function. This is the marginal responsibility coefficient. Indicates the load shift ratio. Indicates the ratio of energy storage charge / discharge power. This indicates the proportion of photovoltaic capacity that can be increased.

[0019] This invention also provides a system for an electrocarbon metering method based on a causal interpretable graph network, comprising: The data acquisition layer is used to collect power data, meteorological data, and topology data, and then encapsulates these three types of data in a unified manner before transmitting the values ​​to the output processing layer. The data processing layer is used to perform missing data repair, spatiotemporal alignment, and causal variable construction on the received multi-source data to generate aligned power-meteorological-carbon sequences. The causal relationship inference model layer is used to explicitly quantify the marginal carbon responsibility of different nodes at different time periods based on the power-meteorological-carbon sequence.

[0020] Compared with the prior art, the beneficial effects of the present invention are: Real-time node-level carbon factor calculation: This invention is the first to synchronously use real-time regional carbon emission factors with completed node electrical quantities to generate minute-level node carbon factors, ensuring that carbon emission measurement is consistent with the actual operating state of the power grid and significantly reducing errors caused by static factors. Furthermore, this invention constructs dynamic adjacency and, in conjunction with spatiotemporal tensor completion / alignment and causal masking attention, suppresses fluctuations caused by equipment errors / faults at the data level.

[0021] Causal Explainability Mechanism: Through virtual perturbation and attention backtracking, the system explicitly quantifies the carbon responsibility of each node and each moment, and gives the contribution ratio of factors such as load, photovoltaic, and meteorology, so as to realize the readability and traceability of the causes of carbon emissions.

[0022] End-to-end strategy generation: While outputting carbon factors, the system automatically generates three-dimensional actions such as load shifting, energy storage charging and discharging, and photovoltaic capacity expansion based on marginal carbon responsibility and attribution weights. Users can implement carbon reduction plans without manual intervention.

[0023] After the system is powered on, the edge IoT terminal polls data from various sensors and the dispatch center at fixed intervals, encapsulating electrical quantities, meteorological quantities, and topological quantities into a unified message, which is then sent to the data processing layer via the MQTT protocol. The data processing layer first performs minute-level interpolation on missing or abnormal measurement values, then strictly aligns the electrical quantities, meteorological quantities, and topological quantities with the same minute cycle and node number, and finally generates an electricity-carbon coupling feature vector for causal inference based on the power grid flow equation and the real-time regional carbon emission factor, completing the transformation from the original multi-source flow to a high-dimensional causal alignment tensor.

[0024] In the causal relationship inference model layer, the aligned tensor is read, and the electrical, meteorological, and topological information is first mapped into potential causal representations through a causal encoder. Then, a spatiotemporal graph convolutional network slides along the time axis and combines dynamic adjacency relationships to propagate spatial information, forming a node embedding that simultaneously contains spatiotemporal coupling and causal semantics. The causal intervention layer applies a virtual perturbation to the photovoltaic irradiance and compares the difference in carbon emissions before and after the perturbation to obtain the node-level marginal carbon responsibility. The interpretable generator maps the potential representations back to the original variable space and outputs the contribution ratio of each variable to the current carbon emissions. Based on the marginal responsibility and contribution ratio, the policy optimizer automatically generates three-dimensional actions of load shifting, energy storage charging and discharging, and photovoltaic capacity expansion, completing the closed loop from data to decision.

[0025] The application layer displays node-level carbon factors in real time, pushes executable instructions to the user-side energy management system, and provides a policy evaluation interface to the regulatory side, achieving a rapid closed loop among users, parks, and regulators. Attached Figure Description

[0026] Figure 1 This is a flowchart of an electric carbon metering method based on a causal interpretable graph network according to an embodiment of the present invention.

[0027] Figure 2 This is a block diagram of an electric carbon metering system based on a causal interpretable graph network, according to an embodiment of the present invention.

[0028] Figure 3 This is a data processing flowchart of an embodiment of the present invention.

[0029] Figure 4This is a block diagram of the causal relationship inference model layer in an embodiment of the present invention. Detailed Implementation

[0030] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0031] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0032] Please see Figure 1 , 2 As shown, this invention provides a method for measuring electrocarbon based on causal interpretable graph networks, including... S10 collects power data, meteorological data, and topology data, and encapsulates these three types of data in a unified manner.

[0033] In this embodiment, the present invention is an intelligent terminal and optimization suggestion system for industrial and commercial electricity carbon metering based on causal interpretable graph networks, which consists of four core layers: data acquisition layer, data processing layer, causal relationship inference model layer, and application layer.

[0034] The data acquisition layer mainly consists of three sub-modules: intelligent measurement module, meteorological sensing module, and topology synchronization module. These modules are used to collect power data, meteorological data, and topology data, and to encapsulate these three types of data in a unified manner.

[0035] The intelligent measurement module synchronously collects active power P, reactive power Q, node voltage V, and current I at 1-minute intervals using smart meters and transformers installed on the industrial and commercial user side. The meteorological sensing module uses a miniature weather station to obtain photovoltaic irradiance G, wind speed W, and ambient temperature T for the same time period, which are used for subsequent renewable energy output estimation. The topology synchronization module subscribes to network topology files periodically published by the distribution automation system via MQTT. This file is automatically generated by the distribution master station based on real-time telemetry and telecontrol data from on-site ring main units, switching stations, and distributed photovoltaic grid-connected points, and is updated every 5 minutes. The file includes a node list, branch connection relationships, and the details of each branch. Impedance value. Where R is the impedance. Reactance, which includes inductive reactance and capacitive reactance. The unit is the imaginary number. After the edge IoT terminal receives the file, ① if , If it is a per-unit value, then directly calculate the per-unit value of the electrical admittance using the modulus. ② If , For values ​​with given names, first normalize them according to the corresponding node reference voltage, then calculate the admittance magnitude. The per-unit values ​​of the electrical admittance can be directly used to construct the dynamic adjacency matrix at the current moment. , of which elements For nodes With nodes The per-unit electrical admittance of the closed branch; if the switch is open or the equipment is disconnected from the grid, the corresponding... =0. The entire process requires no local power flow calculation and relies entirely on the telemetry and telemetry data sources of the distribution automation system. After being uniformly encapsulated by the edge IoT terminal, the three types of data are sent to the data processing layer via the MQTT protocol, providing high-precision, low-latency multi-source input for the subsequent causal interpretable graph network.

[0036] MQTT (Message Queuing Telemetry Transport) is a lightweight, publish / subscribe (Pub / Sub) IoT communication protocol designed for low-bandwidth, high-latency, or unstable network environments. It transmits P, ... Q Small data packets with a duration of 1 minute, such as V, I, G, W, and T, are reliably delivered to the edge IoT terminal and then forwarded to the data processing layer.

[0037] S20 performs missing data repair, spatiotemporal alignment, and causal variable construction on the received unified encapsulated multi-source data to generate an aligned power-meteorological-carbon sequence.

[0038] In this embodiment, the data processing layer mainly consists of three sub-modules: a missing data repair module, a spatiotemporal alignment module, and a causal variable construction module, which uniformly process the multi-source inputs sent from the data acquisition layer.

[0039] First, the missing data repair module uses a spatiotemporal tensor-based completion algorithm to interpolate occasionally lost measurements at the minute level. Then, the spatiotemporal alignment module strictly aligns all multi-source data according to the same timestamp and spatial node number, ensuring a one-to-one correspondence between electrical quantities and meteorological quantities at any node at any given time. Finally, the causal variable construction module generates an "electricity-carbon" coupled feature vector for causal inference based on the power grid flow equation and carbon emission factor model. This provides a directly input, high-dimensional, low-noise, causally aligned dataset for the causal interpretable graph network of the causal relationship inference model layer. Specifically: In this embodiment, generating the aligned power-weather-carbon sequence includes the following steps: S21, an algorithm based on spatiotemporal tensor completion is used to interpolate the occasionally lost measurement values ​​of power data and meteorological data at the minute level.

[0040] In this embodiment, the missing data repair module reads the original tensors of three types of data pushed by the data acquisition layer at once: active power tensor. Reactive power tensor Node electricity Compression tensor, current tensor and meteorological tensor .in, This refers to the number of metered nodes in the entire industrial and commercial power distribution network. This represents the total number of minute-level sampling points included in a single inference window. It is a real number. For the photovoltaic irradiance tensor, For wind speed tensor, This is the ambient temperature tensor.

[0041] In this embodiment, the algorithm based on spatiotemporal tensor completion for minute-level interpolation of sporadic lost measurement values ​​in multi-source data specifically includes: when a certain moment is detected... a node When any element contains a null value or an anomaly marker, the missing element repair module immediately invokes the spatiotemporal tensor completion algorithm, whereby... Number the nodes. , All are discrete-time indices, corresponding to a 1-minute sampling interval. At the current sampling time, For a historical moment in the neighborhood, For the preset spatiotemporal neighborhood, weights Indicates from arrive The spatiotemporal coupling weights, therefore, after summing, naturally complete the process from... arrive Mapping: ; ; In the formula, For the multi-source data after interpolation and completion, From arrive Spatiotemporal coupling weights, For nodes With nodes Time distance, For nodes With nodes Time distance, For nodes With nodes Spatial electrical distance, For nodes With nodes Spatial electrical distance. A smaller spatial electrical distance value indicates greater relevance, and it is given higher weight in interpolation or attention weighting. For the pre-defined spatiotemporal domain, , All are discrete-time indices. For a historical moment in the neighborhood, These are the node number and the timestamp, respectively. The node number is the traversal node number during normalized summation, and... Belonging to the same set of observation nodes, Here, represents the time decay coefficient and the spatial decay coefficient, and exp is the exponential function. After this step, all tensors are padded to be complete. .

[0042] S22 combines the missing and completed power and meteorological data with the topology data and aligns them with the same timestamp and spatial node number to ensure that the electrical quantity and meteorological quantity of any node at any time correspond one-to-one.

[0043] In this embodiment, the next step is the spatiotemporal alignment module, which uses a uniform 1-minute timeframe to process the dynamic adjacency tensor transmitted from the topology synchronization module. Perform a spatiotemporal alignment resampling with the repaired electrical-meteorological tensor.

[0044] When scheduling topology time interval When linear interpolation is used: ; In the formula, It is an adjacency matrix. For the dispatch center in The original adjacency matrix issued at each time step. For separation recent and ≤ The original topological moment, Leave Recently and > The next original topological time; interval Bundle Sandwiched in the middle, the dynamic adjacency matrix corresponding to that minute is obtained by linear interpolation. This is for use in subsequent lexicographical packaging.

[0045] Then the seven variable node-time pairs were... ,Will spliced ​​together , , and the corresponding adjacent tensor Packaged in lexicographical order: tensor Expand into a list of node-time pairs: ;in, For nodes At any moment eigenvectors.

[0046] For each node-time pair From the adjacency matrix Extract its adjacency vector: ; In the formula, For nodes At any moment Admittance values ​​and topological connections; All node-time pairs The electrical quantities at any node at any time are sorted and mapped to meteorological quantities, generating an aligned global tensor. .

[0047] In this embodiment, an aligned global tensor is generated. A dynamic adjacency matrix is ​​generated by linearly interpolating to minute-level resolution and subscribing to the MQTT from the distribution master station every 5 minutes. In this way, the electrical quantity and meteorological quantity of any node at any time are strictly one-to-one corresponding and synchronized with the power grid topology.

[0048] S23, based on the power grid flow equation and carbon emission factor model, generates an “electricity-carbon” coupled feature vector for causal inference and obtains an aligned power-meteorological-carbon sequence.

[0049] In this embodiment, the final causal variable construction module reads the aligned... Provides "node-time-causal features and provides" "Node-node-spatial correlation" is first calculated using the power flow equations at the nodes: ; In the formula, For nodes At any moment The power of the power recovery at that time For complex conjugate, For nodes At any moment Complex voltage at time, For nodes At any moment Complex voltage at time, For nodes With nodes Between moments The admittance value.

[0050] Then, take the real part of the node's complex power and multiply it by the real-time regional carbon emission factor, and apply the following formula to obtain the carbon emissions: ; In the formula, For nodes At any moment The estimated carbon emissions, Let be the real part of the node's complex power. This represents the real-time regional carbon emission factor.

[0051] Furthermore, externally structured knowledge graphs can be introduced, and knowledge graph embedding methods can be used to complete missing edges or node relationships in causal relationships, including: Let the external knowledge graph be... ,in This is a collection of entities, including industrial and commercial user nodes, distributed photovoltaic nodes, weather station nodes, and carbon emission factor value instances. This is a set of relationships, including: User-Grid-Connected to-Distributed Photovoltaic Node (representing the electrical connection between the user's distribution box and the photovoltaic grid connection point, used to establish a "user → photovoltaic" causal edge, supporting intervention and attribution from photovoltaic output to user carbon emissions); Distributed Photovoltaic Node-Irradiance-Weather Station Node (representing that photovoltaic output is jointly determined by local irradiance, temperature, and wind speed, used to introduce weather confounding control to ensure that weather common factors are eliminated when intervening in photovoltaics); Regional Carbon Emission Factor Value Instance-Affected to-Commercial User Node (representing that the real-time carbon factor is directly multiplied by the user's active power, forming a "electricity → carbon" conversion causal edge, used for marginal carbon responsibility calculation); Industrial and Commercial User Node-Belonging to-Feeder Node (representing which distribution feeder or switching station the user is connected to, used to trace carbon factors upwards and provide topological hierarchy); Node i-Connected to-Node j: indicating that there is a closed branch between the two nodes (admittance per unit value ≠ 0), corresponding to the dynamic adjacency matrix. =1, used for graph convolution message passing, load node - impact carbon emissions - regional carbon emission factor instance (representing that node power changes affect carbon factors in reverse through demand-side response, used to support bidirectional causal edges between demand side and grid side, and for use by the policy evaluation interface).

[0052] The TransE-style embedding method is used to treat entities and relations in the external knowledge graph as vectors, such that the sum of the head entity and relation approximately equals the tail entity. For each triple... The learning vector satisfies: ; Define the causal completion score function as follows: ; In the formula, For causal completion scoring function, For head entity Embedded vector, Tail entity Embedded vector, , For relationship Learnable parameters The function is a sigmoid function, and its output is the probability of the existence of causal edges. This is the matrix transpose symbol.

[0053] Set threshold Construct the completed causal adjacency matrix and update it simultaneously. Adjacency matrix within for Specifically: Based on the probability of the existence of causal edges output by the causal completion score function, construct the completed adjacency matrix: ; In the formula, To complete the adjacency matrix, This is the original adjacency matrix. To set a threshold, For logical OR. When When the value is 1, it indicates that in the original adjacency matrix... The value is already 1, meaning the switch is closed, the branch truly exists, or the scoring function is complete. Threshold .in This represents a logical OR operation; the value is set to 1 if either condition is met, and 0 otherwise. (Completed adjacency matrix) Used for graph convolution message passing, determining whether information exchange is allowed between nodes i and j.

[0054] Finally, the power data, meteorological data, node power, and carbon emissions are concatenated to obtain a power-meteorological-carbon sequence: ; Final output As the direct input to the causal relationship inference model layer, it completes the full-link processing from the original multi-source flow to the high-dimensional causal alignment tensor.

[0055] S30, based on the power-meteorological-carbon sequence, explicitly quantifies the marginal carbon responsibility of different nodes at different times.

[0056] Please see Figure 3 and Figure 4As shown, in this embodiment, the causal relationship inference model layer is the "causal hub" of the invention, mainly composed of five modules connected in series: a causal encoder module, a spatiotemporal graph convolutional network module, a causal intervention layer module, an interpretable generator, and a policy optimizer. The causal encoder module first maps the aligned power-meteorological-carbon sequence from the data processing layer into a high-dimensional latent causal representation; the spatiotemporal graph convolutional network module captures the instantaneous and delayed electro-carbon coupling relationships between nodes based on a dynamic adjacency matrix; the causal intervention layer module applies a virtual "carbon price perturbation" using the front-door criterion to explicitly quantify the marginal carbon responsibility of different nodes at different times. This invention employs the front-door criterion in causal inference to identify causal effects in the presence of confounding factors; the interpretable generator uses attention weights to backtrack the critical path, outputting real-time carbon factors and a visualized attribution map for each industrial and commercial user's 0.4 km²-level node; and the policy optimizer automatically generates and pushes energy consumption and storage optimization schemes that save more than 8% of electricity based on the intervention results. The entire model layer completes the "data → causality → decision" closed loop in an end-to-end manner, realizing high-precision measurement and the generation of interpretable carbon reduction strategies.

[0057] In this embodiment, the marginal carbon responsibility of different nodes at different time periods is explicitly quantified, including: S31 maps the causally aligned power-weather-carbon sequence to a high-dimensional potential causal representation.

[0058] In this embodiment, the causal encoder module compresses the power-weather-carbon sequence from the data processing layer into a unified-dimensional latent causal representation, providing a differentiable, same-scale input to the downstream spatiotemporal graph convolutional network module. ; ; In the formula, , These are the trainable weights and biases of the first layer, respectively. For Swish activation function, , These represent the trainable weights and biases of the second layer, respectively. For the first layer of potential representation, This represents the final potential causal characterization.

[0059] Specifically, attention embedding is used to embed the adjacency matrix into a multi-head attention mechanism for the final potential causal representation. This allows information transfer only between nodes with causal relationships, blocking non-causal paths to enhance the accuracy and interpretability of causal reasoning. The causal encoder module outputs a node representation matrix. The causal mask attention is defined as follows: ; ; In the formula, , , , These are the trainable projection matrices for the query, key, and value in the attention process, respectively. Element-wise multiplication is used to introduce causal masks. , For a moment The causal mask of time, by It is derived from element-wise mapping, without further learning or processing, and is used to mask the attention weights between non-causal nodes. For a moment Time node , Causal mask between them To complete the adjacency matrix, The dimension of the query and key vector.

[0060] S32, based on a high-dimensional latent causal representation and a dynamic adjacency matrix, captures the instantaneous and delayed electrocarbon coupling relationships between nodes.

[0061] In this embodiment, the spatiotemporal graph convolutional network module first uses neighbor aggregation, wherein... These are the original per-unit electrical admittance values, derived from the dynamic adjacency matrix issued by the distribution automation system. The larger the value, the tighter the electrical connection between nodes ij. The normalized electrical attention weights, i.e., the electrical relevance of node i to neighbor j at time t, are used for weighting the neighbor information in the spatiotemporal graph convolution. The sum of the electrical outgoing admittances of node i at the same time is represented by l, which iterates through all neighbors of i and is used as the normalized denominator. ; ; Then, using spacetime gate fusion, the instantaneous and delayed electrocarbon coupling relationships between nodes are obtained: ; In the formula, The elements in the adjacency matrix are represented as nodes. , The per-unit electrical admittance between them , The first The layers can be trained with weights; the former processes neighbor information, while the latter processes its own history. To represent the representation of the same node at the previous time step, it is obtained from the output buffer of the previous layer. The GRU (Gated Recurrent Unit) is a gated recurrent unit, and its output is... This represents the instantaneous and delayed electrocarbon coupling relationship between nodes. For the first Layer, moment Middle node The neighbor aggregated message vector, , For each node , At any moment The representation of time. After 3 layers of convolution, we get Enter the causal intervention layer module.

[0062] S33, based on the instantaneous and delayed electrocarbon coupling relationship between nodes, adopts the front-door criterion of causal recommendation to identify causal effects in the presence of confounding factors.

[0063] In this embodiment, the causal intervention layer module can apply virtual photovoltaic perturbations within the front door frame to explicitly calculate the marginal responsibility coefficient for "how much CO2 can be reduced for every additional 1kW of photovoltaic power". The spatiotemporal graph convolutional network module can only provide the correlation between photovoltaics and carbon emissions, while The net effect after removing confounding factors was quantified: Fact Construction: ; Carbon emissions readout: ; Display the calculated marginal responsibility factor: ; In the formula, For nodes exist The photovoltaic irradiance at any given time comes from the alignment tensor. , The proportion of human-induced disturbances. This is represented as a value range of {-0.3, -0.15, 0, 0.15, 0.3}, used only for a single forward computation, and does not retain gradients. The photovoltaic irradiance after applying a virtual photovoltaic perturbation, To refeed the photovoltaic irradiance after applying virtual photovoltaic perturbation into the same convolutional network, namely three layers of spatiotemporal convolution and GRU, the resulting counterfactual representation is... This represents the counterfactual carbon emissions after photovoltaic perturbation. This represents the real-time regional carbon emission factor. For activation function, For linear readout weights, This is the matrix transpose symbol. The sampling interval is a fixed constant, specifically... =1 / 60h, This serves as the baseline zero value for all subsequent perturbation comparisons. This can be expressed as replacing the average difference with a differential approximation, thus obtaining the instantaneous marginal carbon responsibility. The marginal responsibility coefficient characterizes the identification of causal effects in the presence of confounding factors and is output to the interpretable generator and policy optimizer.

[0064] S34 uses attention weights to backtrack the critical path for the identified causal effects, outputs the real-time carbon factor and visual attribution map for each node, and generates and pushes energy-saving and energy storage optimization schemes based on the real-time carbon factor of each node.

[0065] In this embodiment, the real-time carbon factor and visual attribution map of each node are output, including: The interpretable generator maps the 32-dimensional latent representation back to the 9-dimensional original variable space, that is, the instantaneous and lagged electrocarbon coupling relationship between nodes, generating variable-attribution weights for readability and visualization: ; In the formula, As a variable—attribution weight—representing the proportion of contribution to current carbon emissions, it is fed into the front-end visualization interface. It is an exponential function. This is a trainable projection matrix, where each row corresponds to one original variable channel. This represents the instantaneous and delayed electrocarbon coupling relationship between nodes.

[0066] In this embodiment, generating and pushing energy-saving and energy storage optimization schemes includes: a strategy optimizer based on marginal responsibility coefficients. With variable-attribution weight The system generates three-dimensional actions online, including load shifting, energy storage, and photovoltaics, achieving both energy savings and quantifiable carbon emission reductions. The continuous actions are: ; Objective function: ; In the formula, In order to increase the proportion of photovoltaic capacity, The input features are the marginal responsibility coefficient and the variable-attribution weight concatenation. , These represent the weights and biases of the last layer of the policy network. To ensure element-wise tangent, guarantee the output of each dimension. , The three rows are independently mapped to load translation π1, energy storage charging and discharging π2, and photovoltaic capacity expansion π3, generating three-dimensional actions simultaneously in one forward pass; Let be the objective function. This represents the marginal responsibility coefficient. This indicates the load shift ratio: -1 delays the load by 1 hour, +1 advances the load by 1 hour. Indicates the ratio of energy storage charge / discharge power. Indicates the proportion of photovoltaic capacity that can be increased. This is represented by three hyperparameters with values ​​of 10, 0.5, and 0.3. The first term maximizes emission reduction, while the latter two penalize excessive actions. After gradient backpropagation, The policy is distributed to industrial and commercial EMS via MQTT push.

[0067] In this embodiment, the application layer, deployed on the edge gateway and user-side energy management system (EMS), mainly consists of three lightweight modules: "real-time carbon metering display," "optimization suggestion push," and "policy evaluation interface." The real-time carbon metering display module refreshes the node-level carbon factors generated by the model layer in the form of charts every second, allowing industrial and commercial users to view them intuitively. The optimization suggestion push module automatically generates JSON instructions based on the "load shifting-energy storage-photovoltaic" three-dimensional actions output by the strategy optimizer and sends them to the EMS via the MQTT protocol. The policy evaluation interface provides governments and industrial parks with one-click scenario simulation. By inputting different carbon prices or incentive policies, the overall emission reduction effect can be evaluated, thereby completing a rapid closed loop between the user side, the industrial park side, and the regulatory side.

[0068] Please see Figures 2 to 4 As shown, the present invention also provides a system for an electrocarbon metering method based on a causal interpretable graph network, comprising: The data acquisition layer is used to collect power data, meteorological data, and topology data, and then encapsulates these three types of data in a unified manner before transmitting the values ​​to the output processing layer. The data processing layer is used to perform missing data repair, spatiotemporal alignment, and causal variable construction on the received multi-source data to generate aligned power-meteorological-carbon sequences. The causal relationship inference model layer is used to explicitly quantify the marginal carbon responsibility of different nodes at different time periods based on the power-meteorological-carbon sequence.

[0069] In this embodiment, the system also includes an application layer, which mainly consists of three lightweight modules: "real-time carbon metering display," "optimization suggestion push," and "policy evaluation interface." The real-time carbon metering display refreshes the node-level carbon factors output by the result-relationship inference model layer in the form of charts every second. The optimization suggestion push automatically sends executable instructions to the mobile phones of industrial and commercial users or energy management systems based on the energy-saving schemes generated by the strategy optimizer. The policy evaluation interface provides governments and industrial parks with one-click scenario simulation to evaluate the overall emission reduction effect under different carbon prices or incentive policies, realizing a rapid closed loop between the user side, the industrial park side, and the regulatory side.

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0071] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for measuring carbon electricity based on causal interpretable graph networks, characterized in that, include: Collect power data, meteorological data, and topology data, and encapsulate these three types of data in a unified manner; The received, uniformly packaged multi-source data is subjected to missing data repair, spatiotemporal alignment, and causal variable construction to generate aligned power-meteorological-carbon sequences. Based on the power-meteorological-carbon sequence, the marginal carbon responsibility of different nodes at different time periods is explicitly quantified.

2. The method for metering carbon based on causal interpretable graph networks according to claim 1, characterized in that, Generating aligned power-weather-carbon sequences includes the following steps: An algorithm based on spatiotemporal tensor completion is used to interpolate the occasionally lost measurement values ​​of power and meteorological data at the minute level. The missing and missing power and meteorological data are combined with the topology data and aligned with the same timestamp and spatial node number to ensure that the electrical quantity and meteorological quantity of any node at any time correspond one-to-one. Based on the power grid flow equation and carbon emission factor model, an "electricity-carbon" coupled feature vector for causal inference is generated to obtain an aligned power-meteorological-carbon sequence.

3. The method for measuring carbon electricity based on causal interpretable graph networks according to claim 2, characterized in that, An algorithm based on spatiotemporal tensor completion performs minute-level interpolation on sporadically lost measurements in multi-source data, applying the following formula: ； ； In the formula, For the multi-source data after interpolation and completion, From arrive Spatiotemporal coupling weights, For nodes With nodes Time distance, For nodes With nodes Time distance, For nodes With nodes Spatial electrical distance, For nodes With nodes Spatial electrical distance, For the pre-defined spatiotemporal domain, , All are discrete-time indices. For a historical moment in the neighborhood, These are the node number and the timestamp, respectively. Here, represents the time decay coefficient and the spatial decay coefficient, and exp is an exponential function. For active power tensor, For reactive power tensor, For node voltage tensors, For current tensor, For the photovoltaic irradiance tensor, For wind speed tensor, This is the ambient temperature tensor.

4. The method for measuring carbon electricity based on causal interpretable graph networks according to claim 2, characterized in that, The missing and completed power and meteorological data, combined with topological data, are aligned using the same timestamp and spatial node number, including: The adjacency matrix in the topological data is spatiotemporally aligned and resampled with the repaired multi-source data; whereby the adjacency matrix is ​​linearly interpolated: ； In the formula, It is an adjacency matrix. For the dispatch center in The original adjacency matrix issued at each time step. For separation recent and ≤ The original topological moment, Leave Recently and > The next original topological time; interval Bundle Sandwiched in the middle, the dynamic adjacency matrix corresponding to that minute is obtained by linear interpolation. ; For the dispatch center in The original adjacency matrix issued at each time step; Variables in power and meteorological data are analyzed based on node-time pairs. spliced ​​tensors With the corresponding adjacency matrix Pack the data into dictionaries to generate a global tensor that corresponds one-to-one with the electrical and meteorological quantities at any given time for any node; including: tensor Expand into a list of node-time pairs: ;in, For nodes At any moment eigenvectors; For each node-time pair From the adjacency matrix Extract its adjacency vector: In the formula, For nodes At any moment Admittance values ​​and topological connections; All node-time pairs The electrical quantities at any node at any time are sorted and mapped to meteorological quantities, generating an aligned global tensor. .

5. The method for metering carbon based on causal interpretable graph networks according to claim 2, characterized in that, To obtain the aligned power-meteorological-carbon sequence, the steps include: First, use the power flow equations to calculate the node injection power: ； In the formula, For nodes At any moment The power of the power recovery at that time For complex conjugate, For nodes At any moment Complex voltage at time, For nodes At any moment Complex voltage at time, For nodes With nodes Between moments The admittance value; Then, take the real part of the node's complex power and multiply it by the real-time regional carbon emission factor, and apply the following formula to obtain the carbon emissions: ； In the formula, For nodes At any moment The estimated carbon emissions, Let be the real part of the node's complex power. Real-time regional carbon emission factor; Finally, the power data, meteorological data, node power, and carbon emissions are spliced ​​together to obtain the power-meteorological-carbon sequence.

6. The method for measuring carbon electricity based on causal interpretable graph networks according to claim 4, characterized in that, By introducing an externally structured knowledge graph, missing edges or node relationships in causal relationships are completed using knowledge graph embedding methods, including: Let the external knowledge graph be... ,in, For a collection of entities, For a set of relations; The TransE-style embedding method is used to treat entities and relations in the external knowledge graph as vectors, such that the sum of the head entity and relation approximately equals the tail entity. For each triple... The learning vector satisfies: ; Define the causal completion score function as follows: ； In the formula, For causal completion scoring function, For head entity Embedded vector, Tail entity Embedded vector, , For relationship Learnable parameters, The function is a sigmoid function, and its output is the probability of the existence of causal edges. This is the matrix transpose symbol; Based on the probability of the existence of causal edges output by the causal completion score function, construct the completed adjacency matrix: ； In the formula, To complete the adjacency matrix, This is the original adjacency matrix. To set a threshold, For logical OR.

7. The method for metering carbon electricity based on causal interpretable graph networks according to claim 1, characterized in that, Explicitly quantify the marginal carbon responsibility of different nodes at different times, including: Map the causally aligned power-meteorological-carbon sequence to a high-dimensional potential causal representation; Based on the high-dimensional potential causal representation, and using a dynamic adjacency matrix, the instantaneous and delayed electrocarbon coupling relationships between nodes are captured. Based on the instantaneous and delayed electrocarbon coupling relationship between nodes, the front-door criterion of causal recommendation is adopted to identify causal effects in the presence of confounding factors. For the identified causal effects, attention weights are used to backtrack the critical path, outputting the real-time carbon factor and visual attribution map for each node; and based on the real-time carbon factor of each node, energy-saving and energy storage optimization schemes are generated and pushed.

8. The method for metering carbon based on causal interpretable graph networks according to claim 7, characterized in that, Obtaining a high-dimensional latent causal representation involves the following steps: By compressing causally aligned power-meteorological-carbon sequences into a unified-dimensional potential causal characterization, differentiable, homoscaled inputs are provided to downstream applications. ； ； In the formula, , These are the trainable weights and biases of the first layer, respectively. For Swish activation function, , These represent the trainable weights and biases of the second layer, respectively. For the first layer of potential representation, For the final potential causal characterization; Specifically, attention embedding is used for the final potential causal representation. The adjacency matrix is ​​explicitly embedded into a multi-head attention mechanism, allowing information to be passed only between nodes with causal relationships, blocking non-causal paths, thereby enhancing the accuracy and interpretability of causal reasoning. ； ； In the formula, , , , These are the trainable projection matrices for the query, key, and value in the attention process, respectively. For element-wise multiplication, For a moment Causal mask of time For a moment Time node , Causal mask between them To complete the adjacency matrix, The dimension of the query and key vector.

9. The method for measuring carbon electricity based on causal interpretable graph networks according to claim 7, characterized in that, Based on a dynamic adjacency matrix, the instantaneous and delayed electrocarbon coupling relationships between nodes are captured, including: Based on the adjacency matrix, we first use neighbor aggregation: Then, using spacetime gate fusion, the instantaneous and delayed electrocarbon coupling relationships between nodes are obtained: In the formula, The elements in the adjacency matrix are represented as nodes. , The per-unit electrical admittance between them , The first The layers can be trained with weights; the former processes neighbor information, while the latter processes its own history. To represent the representation of the same node at the previous time step, it is obtained from the output buffer of the previous layer. The GRU is a gated recurrent unit, and its output is... This represents the instantaneous and delayed electrocarbon coupling relationship between nodes. For the first Layer, moment Middle node The neighbor aggregation message vector, , For each node , At any moment The representation of time.

10. The method for measuring carbon electricity based on causal interpretable graph networks according to claim 7, characterized in that, Using the front-door criterion for causal recommendation, causal effects are identified in the presence of confounding factors, including: Counterfactual construction: ; Carbon emissions readout: ; Display the calculated marginal responsibility factor: ; In the formula, For nodes exist Photovoltaic irradiance at any given time The proportion of human-induced disturbances. This is represented as a value range of {-0.3, -0.15, 0, 0.15, 0.3}, used only for a single forward computation, and does not retain gradients. The photovoltaic irradiance after applying a virtual photovoltaic perturbation, To refeed the photovoltaic irradiance after applying a virtual photovoltaic perturbation into the same convolutional network, the resulting counterfactual representation is... This represents the counterfactual carbon emissions after photovoltaic perturbation. This is a real-time regional carbon emission factor. For activation function, For linear readout weights, This is the matrix transpose symbol. The sampling interval is a fixed constant. This serves as the baseline zero value for all subsequent perturbation comparisons. This can be expressed as replacing the average difference with a differential approximation, thus obtaining the instantaneous marginal carbon responsibility. The marginal responsibility coefficient characterizes the identification of causal effects in the presence of confounding factors.

11. The method for metering carbon based on causal interpretable graph networks according to claim 7, characterized in that, Generate and push energy-saving and energy storage optimization solutions through the following steps: The instantaneous and delayed electrocarbon coupling relationships between nodes are used to generate variable-attribution weights for readability and visualization: In the formula, For variable-attribution weight, It is an exponential function. This is a trainable projection matrix, where each row corresponds to one original variable channel. The instantaneous and delayed electrocarbon coupling relationship between nodes; Based on marginal responsibility coefficients and variable-attribution weights, a three-dimensional action of "load shifting - energy storage - photovoltaic" is generated online to achieve energy saving and quantifiable carbon emission reduction. The continuous actions are: ； Objective function: ； In the formula, In order to increase the proportion of photovoltaic capacity, The input features are the marginal responsibility coefficient and the variable-attribution weight concatenation. , These represent the weights and biases of the last layer of the policy network. For element-wise tangent, The three rows are independently mapped to load translation π1, energy storage charging and discharging π2, and photovoltaic capacity expansion π3, generating three-dimensional actions simultaneously in one forward pass; Represented as three hyperparameters, Let be the objective function. This is the marginal responsibility coefficient. Indicates the load shift ratio. Indicates the ratio of energy storage charge / discharge power. This indicates the proportion of photovoltaic capacity that can be increased.

12. A system for an electrocarbon metering method based on a causal interpretable graph network according to any one of claims 1-11, characterized in that, include: The data acquisition layer is used to collect power data, meteorological data, and topology data, and to encapsulate these three types of data in a unified manner. The data processing layer is used to perform missing data repair, spatiotemporal alignment, and causal variable construction on the received unified encapsulated multi-source data to generate aligned power-meteorological-carbon sequences. The causal relationship inference model layer is used to explicitly quantify the marginal carbon responsibility of different nodes at different time periods based on the power-meteorological-carbon sequence. Collect power data, meteorological data, and topology data, and encapsulate these three types of data in a unified manner; The received, uniformly packaged multi-source data is subjected to missing data repair, spatiotemporal alignment, and causal variable construction to generate aligned power-meteorological-carbon sequences. Based on the power-meteorological-carbon sequence, the marginal carbon responsibility of different nodes at different time periods is explicitly quantified.

Images