Power system carbon potential tracking and predicting method based on physical and data dual drive

Abstract

The invention provides an electric power system carbon potential tracking and predicting method based on physical and data dual drive. The method comprises the following steps: constructing a time sequence characteristic data set; constructing a source-side unit double-layer collaboration map, and generating a normalized adjacent matrix; constructing a physically guided source side graph neural network model, and outputting a predicted source side dynamic carbon emission factor and a node carbon injection amount sequence; constructing a high-dimensional input tensor through a space-time diagram attention mechanism; constructing a whole-network space-time diagram neural network model to realize dynamic tracking of carbon potential; the method comprises the following steps: introducing a Kirchhoff carbon flow conservation law as a physical regularization term in a space-time law deduction process, constructing a mixed loss function containing the physical regularization term, and executing an optimal sentinel mechanism in a back propagation process of whole-network space-time diagram neural network model training, and finally, outputting a dynamic carbon potential prediction result of the load side after physical verification. According to the invention, the precision and credibility of carbon potential prediction can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and more specifically to a method for tracking and predicting carbon potential in power systems based on a dual physical and data-driven approach. Background Technology

[0002] Accurate measurement and real-time tracking of carbon emission flows (CEF), that is, clarifying the carbon emission responsibility behind every node, every moment, and every kilowatt-hour in the power system, has become the core data foundation for realizing low-carbon dispatching of the power system, carbon market trading and settlement, and user-side carbon management.

[0003] Currently, the main methods for carbon flow tracking and prediction are quasi-steady-state calculation methods based on physical models and data-driven methods based on deep learning. However, existing technologies have the following main shortcomings: 1. The "static" nature of source-end carbon emission perception, disconnected from physical realities, fails to reflect the transient physical characteristics under changing generator operating conditions. Most existing carbon flow calculation methods typically employ static average carbon emission factors when processing source-end generator data. This leads to a severe neglect of the thermodynamic transient characteristics of generator combustion efficiency deteriorating significantly due to deviations from steady state when responding to dynamic grid topology changes or balancing renewable energy fluctuations. Existing prediction models largely lack the ability to model the nonlinear coupling relationship between "output change rate and combustion efficiency," resulting in significant physical biases in the source-end carbon emission data itself. Due to this lack of physical modeling at the source, these biases are amplified across the entire network during power flow calculations, leading to low accuracy in the final carbon emission factor prediction.

[0004] 2. Purely data-driven "black box" models lack physical consistency constraints and closed-loop corrections, resulting in low reliability of dynamic carbon potential predictions throughout the entire process. Most current AI prediction research is purely "data fitting," but the power system is a physical system that strictly follows physical laws; carbon flow transmission must strictly adhere to Kirchhoff's laws and the law of conservation of carbon mass. Existing technologies make models prone to outputting results that violate physical common sense, lacking a complete closed-loop data link from "source-side state" to "grid-side power flow" and then to "load-side carbon potential." This lack of physical consistency makes AI models unreliable in actual scheduling decisions and unable to directly replace traditional physical computing engines. Summary of the Invention

[0005] The purpose of this invention is to provide a method for tracking and predicting carbon potential in power systems based on both physical and data-driven approaches, so as to improve the accuracy and reliability of carbon potential prediction.

[0006] A method for carbon potential tracking and prediction in power systems based on both physical and data-driven approaches includes: Step S1: Collect source-side physical parameters, energy storage-side physical parameters, grid-side operating parameters, and load-side operating parameters in the power system, and preprocess the collected parameters to obtain a time-series feature dataset; Step S2: Construct a source-side unit two-layer collaborative graph based on the time-series feature dataset and generate a normalized adjacency matrix. The source-side unit two-layer collaborative graph includes physical connection edges representing electrical position coupling and logical collaborative edges representing similar units when responding to the system's automatic power generation control commands, thereby mapping the dual coupling relationship of the unit group in the physical space and logical control space. Step S3: Construct a physically guided source-side graph neural network model based on the source-side unit dual-layer collaborative graph, and output the predicted source-side dynamic carbon emission factor and node carbon injection sequence through the source-side graph neural network model; Step S4: Based on the branch power flow section data of the power grid side and the predicted source-side dynamic carbon emission factor and node carbon injection sequence, a high-dimensional input tensor representing the carbon flow diffusion law of the whole network and the source-side dynamic carbon potential prediction information and the real-time electrical operation status data of the power grid side is constructed through the spatiotemporal graph attention mechanism. Step S5: Construct a full-network spatiotemporal graph neural network model, using high-dimensional input tensors as the input source, and employing a parallel architecture of backbone extrapolation and expert correction to extract spatial and temporal characteristics, and extrapolate the spatiotemporal laws of carbon potential distribution changes with the current throughout the entire process, thereby achieving dynamic tracking of carbon potential. Step S6: In the process of spatiotemporal law deduction, Kirchhoff's carbon flow conservation law is introduced as a physical regularization term. A hybrid loss function containing the physical regularization term is constructed. During the backpropagation process of training the spatiotemporal graph neural network model, the optimal sentinel mechanism is executed, and finally, the dynamic carbon potential prediction result of the load side after physical verification is output.

[0007] The power system carbon potential tracking and prediction method based on both physical and data-driven approaches provided by this invention has the following advantages: 1. High-fidelity modeling of transient characteristics at the source has been achieved, eliminating data bias at the source. This invention is the first to embed a differentiable physical mechanism layer for calculating basic physical prediction values ​​and a residual error correction module for nonlinear residual correction into the source-side graph neural network model. This enables explicit quantification of the additional carbon emissions of generator units during ramp-up and peak-shaving processes, filling the blind spots of traditional steady-state models under dynamic operating conditions. This significantly improves the prediction accuracy of source-side carbon emission factors and provides high-precision boundary conditions for calculating carbon flow across the entire network.

[0008] 2. A closed-loop monitoring architecture covering the entire process from source to grid to load to storage has been constructed, reducing engineering deployment costs. This invention utilizes source-side predictive data to directly drive grid-side simulations, establishing a complete data link from "source-side unit operating status" to "grid-side power flow distribution" and then to "load-side carbon potential sensing," and incorporates energy storage, accurately characterizing the time-shifting characteristics of energy storage as a "carbon flow buffer," thus improving the completeness of carbon metering in the new power system. This closed-loop design eliminates the need for expensive real-time carbon emission monitoring equipment (CEMS) across the entire grid in practical applications; high-precision carbon potential sensing can be achieved solely using existing SCADA electrical data, significantly reducing hardware investment costs for grid decarbonization management.

[0009] 3. Physical consistency constraints ensure the credibility of the results and improve robustness. By introducing Kirchhoff's carbon flow conservation law as a physical regularization term during training, and in conjunction with a sentinel mechanism, the output is forced to satisfy the conservation of energy and carbon mass. This not only prevents the black-box model from producing predictions that violate physical common sense in sparse data regions, but also gives the model a clear physical meaning, making its results more easily accepted and trusted by power dispatching and operation personnel. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the carbon potential tracking and prediction method for power systems based on both physical and data-driven approaches provided by this invention. Figure 2 Thermographic map of the spatiotemporal distribution of carbon potential for 118 bus nodes; Figure 3 Comparison of 24-hour dynamic carbon potential tracking curves for different unit groups; Figure 4 This is a box plot comparing the prediction error distribution of the method of this invention and the traditional method; Figure 5 The graph shows the convergence process of physical regularization training based on Kirchhoff's laws. Detailed Implementation

[0011] To facilitate understanding of the present invention, a more complete description will be given below with reference to various embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] Please see Figure 1 The embodiments of the present invention provide a method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches, including steps S1-S6: Step S1: Collect source-side physical parameters, energy storage-side physical parameters, grid-side operating parameters, and load-side operating parameters in the power system, and preprocess the collected parameters to obtain a time-series feature dataset.

[0014] In this embodiment, the IEEE 118-node standard test system, which has undergone a complete process modification, is selected as the simulation object. The system has a complex topology, including 54 generating units (coal-fired, gas-fired, and renewable energy), 118 bus nodes, 186 transmission lines, and newly added distributed energy storage sites. It covers complex operating conditions with baseload units and peak-shaving units coexisting and a high proportion of renewable energy access, and can well represent the power system operation characteristics of the actual regional power grid.

[0015] To construct comprehensive data on the dynamic physical characteristics of the power system and ensure its standardization and usability, this embodiment first utilizes a data acquisition module to collect and preprocess multi-dimensional data. This acquires real-time operational status data and inherent physical attribute data for the entire power system chain, from power source to grid to load to storage. Preprocessing is then performed to construct a physical data foundation that accurately reflects the spatiotemporal evolution of carbon potential throughout the power system. Specifically, step S1 includes: The power plant's DCS interface is used to collect source-side physical parameters, specifically the inherent attributes of the source-end operating status data, including: the unit's rated capacity. Minimum technical output Active power output of the unit The gradient rate, characterizing the transient adjustment intensity, was obtained by first-order difference calculation. Simultaneously, the lower heating value of the unit's fuel and the design coal consumption and gas consumption curves are read, including the ramp rate. The calculation formula is:

[0016] in, for The active power output of the time unit and The difference in active power output between the time units This represents the sampling time interval.

[0017] Slope rate It is a key physical quantity characterizing whether the unit is in a rapid peak-shaving state, and is directly related to the nonlinear change of combustion efficiency. It is used to make transient corrections to steady-state carbon emissions in the subsequent step S3.

[0018] Physical parameters of the energy storage side are collected through the BMS interface of the energy storage management system, including the charging and discharging power of the entire network of energy storage power stations. Real-time state of charge and energy storage cycle efficiency .

[0019] In this embodiment, the charging state is set. When the value is negative, it is considered a carbon load, absorbing carbon emissions from the power grid; under discharge conditions... If the value is positive, it is considered a carbon flow source, releasing the stored carbon potential. This accurately characterizes the time-shifting characteristics of energy storage as a 'carbon flow buffer' for the power grid.

[0020] Collect parameters from both the grid side and the load side, including: voltage amplitude at all bus nodes of the entire network. Voltage phase angle Active load reactive load And the active power flow direction of key branch sections. With traffic data To characterize the storage and release properties of carbon streams over time.

[0021] The collected parameters then need to be cleaned, aligned, and Z-score standardized to eliminate the gradient descent difficulties caused by numerical differences, ultimately yielding a time-series feature dataset containing physical mechanism attributes and electrical operating states.

[0022] Step S2: Construct a source-side unit two-layer collaborative graph based on the time-series feature dataset and generate a normalized adjacency matrix. The source-side unit two-layer collaborative graph includes physical connection edges representing electrical position coupling and logical collaborative edges representing similar units when responding to the system's automatic power generation control commands, thereby mapping the dual coupling relationship of the unit group in the physical space and logical control space.

[0023] Specifically, step S2 includes: A two-layer collaborative graph of source-side units is constructed based on a time-series feature dataset. If two generators are connected to the same bus node of the power network, physical connection edges are established in the graph structure to represent the coupling relationship of electrical locations.

[0024] Specifically, iterate through all nodes where the generator sets are located. Determine the location of the generator unit; if the node where the generator unit is located... and the node where the unit is located If nodes are connected to the same bus or to adjacent buses that are electrically very close, then undirected edges, i.e., nodes, are created in the graph structure. and nodes Physical connection edge between If there is no physical connection, then The physical connection edge reflects the strong coupling between units that belong to the same local voltage stability constraint or reactive power support requirement.

[0025] If two generator sets have the same fuel type attribute, then logical cooperative edges are constructed in the graph structure to characterize the group cooperative response behavior of similar units in the process of system load distribution and peak shaving.

[0026] Specifically, iterate through all nodes where the generator sets are located. Perform fuel attribute clustering. If the node where the unit is located... and the node where the unit is located If the nodes belong to the same fuel type, a logical collaborative edge is established, i.e., the nodes... and nodes Logical connection edges between If there is no logical connection, then When responding to the system's automatic generation control (AGC) commands or balancing new energy fluctuations, similar types of generating units, especially gas-fired peak-shaving units, exhibit highly consistent group coordinated behavior, such as gas-fired peak-shaving units collectively increasing output or collectively shutting down.

[0027] Based on the aforementioned physical connection edges and logical coordination edges, an adjacency matrix for the power supply units is generated. And calculate the corresponding degree matrix. Furthermore, combined with the identity matrix The normalized adjacency matrix is ​​calculated. The adjacency matrix Degree matrix and identity matrix The dimension is consistent with the total number of generator units in the entire network, and is used to aggregate spatial coordination information among generator units in graph convolution.

[0028] Specifically, the normalized adjacency matrix Calculated using the following formula:

[0029]

[0030] in, For physical connection edges, For logical connection edges, The weight coefficients for logical collaboration edges are used to adjust the relative importance of physical topology and logical collaboration in the graph.

[0031] Normalized adjacency matrix This serves as the spatial structure input for subsequent graph neural networks, preventing numerical explosions in nodes with high degrees during feature aggregation.

[0032] Step S3: Construct a physically guided source-side graph neural network model based on the source-side unit dual-layer collaborative graph, and output the predicted source-side dynamic carbon emission factor and node carbon injection sequence through the source-side graph neural network model.

[0033] In this embodiment, a source-side unit dual-layer collaborative graph constructed using S2 is used. To address the shortcomings of traditional methods that only use source-side static carbon emission factors to study subsequent load-side carbon potential, a physical-guided source-side graph neural network model is proposed. This model adopts a three-stage architecture of "basic prediction + graph collaborative feature extraction + residual error correction" to generate high-precision source-side dynamic carbon emission factors.

[0034] Specifically, the source-side graph neural network model includes a basic prediction module, a graph collaborative feature module, and a residual error correction module.

[0035] The basic prediction module is used to calculate basic physical prediction values ​​based on the unit's operating status and the embedded differentiable physical mechanism layer. This enables the model to learn the general physical laws governing unit emissions.

[0036] First, based on the steady-state heat rate curve, a baseline value for calculating the basic carbon emission intensity under steady-state conditions is calculated using polynomial fitting. The expression is:

[0037] in, The first transient penalty coefficient characterizing the unit's consumption characteristics. For the coefficients of the constant term, These are transient ramp penalty mechanism coefficients, which reflect the basic emission characteristics of different types of units (such as coal-fired or gas-fired) under steady-state design conditions.

[0038] Then, based on the ramp rate and the learnable second transient penalty coefficient The degradation of combustion efficiency during rapid load changes of the unit is quantified using physical formulas. The expression is:

[0039] in, The transient sensitivity coefficient is preferred. The value is between 0.05 and 0.08.

[0040] Ultimately, the fundamental physics predictions for: .

[0041] The graph collaboration feature module is used to build upon the constructed normalized adjacency matrix. By combining spatiotemporal convolutional networks with graph attention mechanisms and temporal convolutional networks, the operating status features of neighboring units are aggregated to extract the collaborative feature vector of the unit group; and fuel feature vectors are output based on the fuel attribute data of the units, after passing through the fuel feature embedding layer, to distinguish the error distribution characteristics of units with different fuels.

[0042] The residual error correction module is used for: First, the collaborative feature vectors and fuel feature vectors of the unit group are clustered and then input into an independent residual error correction network. This network freezes the basic physical parameters and specifically fits the equipment aging and environmental errors that cannot be covered by the physical formulas. It outputs nonlinear residual corrections that cannot be covered by the basic physical calculations. ; Then the basic physical prediction values With nonlinear residual correction By superimposing the results, the source-side dynamic carbon emission factor is obtained. The expression is:

[0043] Finally, based on With the active power output of the generator unit Compute node carbon injection sequence The expression is: .

[0044] Node carbon injection sequence The source-side dynamic boundary conditions for the whole network simulation are input into step S4.

[0045] Step S4: Based on the branch power flow section data of the power grid side and the predicted source-side dynamic carbon emission factor and node carbon injection sequence, a high-dimensional input tensor is constructed using a spatiotemporal graph attention mechanism to characterize the carbon flow diffusion law of the entire network and the source-side dynamic carbon potential prediction information and the real-time electrical operation status data of the power grid side.

[0046] Specifically, step S4 includes: Step S41, Source-side dynamic carbon emission factor and carbon injection sequence Mapped to the corresponding mesh node, and injected as a dynamic attribute tensor of the source node; Step S42, the charging and discharging power of the entire network's energy storage power stations. With real-time state of charge As a dynamic feature injection tensor for energy storage nodes, the model can perceive whether the energy storage is currently in a charging state of "accumulating low-carbon electricity" or a discharging state of "releasing stored carbon potential". Step S43: Merge the voltage amplitude of all bus nodes in the network. Voltage phase angle Active load reactive load And constructing a dynamic power flow adjacency matrix based on real-time power flow calculation results. This matrix reflects the on / off status of the line in real time, ensuring that the carbon flow projection path is consistent with the physical power flow path. Step S44: Obtain the overall macro-level situation characteristics of the power system through a network-wide macro-level situation broadcast. These characteristics include the total load of the power system. Total reserve ratio Total climbing requirements and the total energy storage backup capacity of the system It can utilize global context features to concatenate them into the local feature vector of each node, assisting local nodes in perceiving macroscopic operating conditions; Step S45 involves tensor fusion of the parameters from steps S41 to S44, followed by multi-channel slice stacking and dimensionality normalization verification to obtain a high-dimensional input tensor. This high-dimensional input tensor includes the source-side dynamic carbon emission factor. Carbon injection sequence The charging and discharging power of the entire network's energy storage power stations Real-time state of charge Voltage amplitude of all bus nodes in the network Voltage phase angle Active load reactive load Dynamic power flow adjacency matrix Overall macro-level characteristics of the network.

[0047] Step S5: Construct a full-network spatiotemporal graph neural network model. Use the high-dimensional input tensor as the input source and adopt a parallel architecture of backbone inference and expert correction to extract spatial and temporal characteristics. Perform spatiotemporal law inference of the carbon potential distribution change with the current throughout the entire process, thereby realizing dynamic tracking of carbon potential.

[0048] Specifically, step S5 includes: Step S51: Input the high-dimensional input tensor into the backbone inference pathway of the whole-network spatiotemporal graph neural network model. Spatial extrapolation is performed to extract universal carbon flow diffusion patterns across the entire network. Specifically, a deep residual graph convolutional network is used to extract the spatial topological feature matrix of carbon flows across the entire network, expressed as:

[0049] in, and They are respectively Time of the first and the The input tensor of the layer network, For activation function, This is the learnable parameter matrix of the spatial convolutional layer; Step S52: The LSTM unit takes the entire network carbon flow spatial topology feature matrix as input and outputs the global overall prediction result of the load-side carbon potential. ; Specifically, for energy storage nodes, LSTM cells implicitly achieve time-shifted memory of carbon flow through their gating mechanism: that is, during the charging period... Record the integral of the mixed carbon potential of the power grid during the discharge period. This allows for the accurate simulation of energy storage as a dual "energy-carbon" buffer.

[0050] Step S53: In parallel, start the expert correction path in the whole-network spatiotemporal graph neural network model. To obtain the local node prediction results of the carbon potential on the load side. ; Specifically, expert-modified pathways This is used for feature processing of difficult nodes. If the high-dimensional input tensor output in step S45 is diverted to the expert correction path... Through expert-corrected pathways Capture local high-frequency fluctuations at nodes; for challenging nodes with persistently large prediction errors, typically hub nodes or connecting line nodes where multiple lines intersect, expert correction pathways are implemented. It focuses on extracting local high-frequency fluctuation characteristics of specific nodes, and is not diluted by the average pattern of the entire network.

[0051] Step S54, using the overall prediction results of the load-side carbon potential. Using the baseline value, the predicted results of local nodes of carbon potential on the load side are used. As a correction factor, the dynamic carbon potential on the output load side. This enables dynamic tracking of carbon potential. The expression is:

[0052] in, For adaptive gating coefficients, This indicates element-wise multiplication.

[0053] Specifically, adaptive gating coefficient Through learnable gating mechanisms Adjustments are made. Gating mechanism. The expression is: .

[0054] When the node is in a stable operating condition Approaching 1, the model primarily employs a backbone network; when nodes experience drastic fluctuations or topological abrupt changes, When the value approaches 0, the model automatically switches to the correction result from the expert network.

[0055] Step S6: In the process of spatiotemporal law deduction, Kirchhoff's carbon flow conservation law is introduced as a physical regularization term. A hybrid loss function containing the physical regularization term is constructed. During the backpropagation process of training the spatiotemporal graph neural network model, the optimal sentinel mechanism is executed, and finally, the dynamic carbon potential prediction result of the load side after physical verification is output.

[0056] To ensure the physical reliability of the prediction results, the dynamic carbon potential on the load side output in step S54 is used in the training and deduction process of the whole-network spatiotemporal graph neural network model. We introduce a strong physical constraint framework based on Kirchhoff's carbon flow conservation and construct a hybrid loss function and an optimal sentinel mechanism.

[0057] First, a hybrid loss function including a physical regularization term is constructed. Specifically, based on the proportional sharing principle of carbon flow calculation, and according to the current power flow distribution and switching status of the power grid, a theoretical matrix equation for the carbon flow distribution of the entire network is established as a verification benchmark and a physical benchmark for subsequent verification of whether the predicted values ​​conform to Kirchhoff's laws. The expression of the theoretical matrix equation is as follows:

[0058] in, This represents the carbon potential vector of all nodes in the network. Inject a diagonal matrix into all nodes. This is the branch power flow distribution matrix. Indicates transpose. Inject vector for source carbon; Then the load-side dynamic carbon potential output in step S54 is... As the test object, using the theoretical matrix equation as the physical benchmark, the physical loss term is calculated to test whether the node where the predicted value is located satisfies the balance between the inflow and outflow of carbon flux. The expression is:

[0059] in, For physical regularization, it is constructed based on the generalized Kirchhoff carbon flux conservation law and represents the degree of violation of the Kirchhoff carbon flux conservation law; These are the weighting coefficients for physical constraints. To test the time-series index, To verify the node index; Carbon flux from upstream lines includes carbon flux flowing in from upstream lines, carbon injection from local generating units, and carbon flux released by energy storage discharge. The carbon flux of downstream lines includes the outflow carbon flux to downstream lines, the carbon flux consumed by local loads, and the carbon flux absorbed by energy storage charging. By minimizing the physical regularization term, predictions that violate physical conservation are penalized, forcing the solution space of the neural network to converge within the physically feasible region.

[0060] Finally, the hybrid loss function The expression is:

[0061] in, Huber loss, a data-driven term, measures the prediction error between the predicted and labeled values ​​of carbon emission factors.

[0062] By minimizing this hybrid loss, the model is forced to approximate the real data while its solution space strictly converges within the physically feasible region defined by Kirchhoff's carbon flow conservation law. This step minimizes the hybrid loss function through strong physical constraints, driving the model parameters in the solution space to satisfy both data accuracy and physical conservation laws, providing data reference for the subsequent training and monitoring of the optimal sentinel mechanism.

[0063] Furthermore, in this embodiment, an optimal sentinel mechanism is executed during the backpropagation process of model training.

[0064] Specifically, an independent monitoring process is established during training to track the physical consistency index and prediction accuracy of the validation set in real time. When the overall index of the current round is found to be better than the historical best record, a snapshot action is immediately triggered, saving a copy of the current model parameters as the "optimal sentinel model". After training, this optimal copy is automatically loaded as the final system kernel, effectively eliminating performance oscillations of the deep learning model in the later stages of convergence and ensuring the robustness of the system output; high-precision dynamic carbon potential prediction results are output for all load-side nodes in the network.

[0065] After the model training is completed, the system automatically loads the optimal sentinel mechanism as the final system kernel, and finally outputs the dynamic carbon potential prediction results of the load side after physical verification.

[0066] The following is a simulation test of this embodiment. Figure 2 A thermal map of the spatiotemporal distribution of carbon potential at 118 bus nodes, from Figure 2 As can be seen, this invention can intuitively present the spatiotemporal distribution of carbon potential across the entire network. Figure 2The red areas clearly mark the high-carbon bottlenecks where thermal power plants are concentrated, while the green and blue areas mark the low-carbon zones dominated by new energy and energy storage, providing an intuitive basis for dispatching decisions.

[0067] Figure 3 A comparison chart of 24-hour dynamic carbon potential tracking curves for different unit groups. Figure 3 The upper section of the curve represents the average carbon emission intensity of a coal-fired power plant group, while the lower section represents the average carbon emission intensity of a gas-fired peak-shaving power plant group. The red and blue shaded areas indicate the fluctuation range within the corresponding power plant group. The results show that this invention can accurately capture the time-varying characteristics of the unit carbon emission factors: the emission factors of coal-fired units fluctuate with output in the range of 1050-1150 g / kWh, rather than being a fixed constant as in traditional methods; although gas-fired units have lower emissions in the range of 600-650 g / kWh, their fluctuation trend with peak-shaving commands is also accurately characterized. This dynamic tracking capability verifies the effectiveness of the source-side physical guidance layer in fitting the nonlinear changes in combustion efficiency, providing high-fidelity boundary conditions for subsequent full-network simulations.

[0068] Figure 4 This is a box plot comparing the prediction error distribution of the method of this invention and the traditional method. The results show that the traditional method includes pure data-driven methods (LSTM / GCN) and traditional physical computation (quasi-steady-state model). Figure 4 As can be seen, the median error of the method of this invention is significantly lower than that of the pure data-driven method and the traditional physical calculation method in terms of the mean absolute percentage error (MAPE) of the entire network. Moreover, the error distribution box is shorter and the fluctuation is smaller, which proves the high accuracy and strong robustness of the invention.

[0069] Figure 5 The graph shows the convergence process of physical regularization training based on Kirchhoff's laws. Figure 5 It can be seen that the physical conservation residual decreases rapidly with the number of training rounds, especially around the 200th training round. With the locking of the optimal sentinel mechanism and the intervention of the expert network, the residual decreases twice and eventually converges to an extremely low level, proving that the present invention strictly follows the physical conservation law.

[0070] In summary, the above-described method for carbon potential tracking and prediction in power systems based on both physical and data-driven approaches offers the following advantages: 1. High-fidelity modeling of transient characteristics at the source has been achieved, eliminating data bias at the source. This invention is the first to embed a differentiable physical mechanism layer for calculating basic physical prediction values ​​and a residual error correction module for nonlinear residual correction into the source-side graph neural network model. This enables explicit quantification of the additional carbon emissions of generator units during ramp-up and peak-shaving processes, filling the blind spots of traditional steady-state models under dynamic operating conditions. This significantly improves the prediction accuracy of source-side carbon emission factors and provides high-precision boundary conditions for calculating carbon flow across the entire network.

[0071] 2. A closed-loop monitoring architecture covering the entire process from source to grid to load to storage has been constructed, reducing engineering deployment costs. This invention utilizes source-side predictive data to directly drive grid-side simulations, establishing a complete data link from "source-side unit operating status" to "grid-side power flow distribution" and then to "load-side carbon potential sensing," and incorporates energy storage, accurately characterizing the time-shifting characteristics of energy storage as a "carbon flow buffer," thus improving the completeness of carbon metering in the new power system. This closed-loop design eliminates the need for expensive real-time carbon emission monitoring equipment (CEMS) across the entire grid in practical applications; high-precision carbon potential sensing can be achieved solely using existing SCADA electrical data, significantly reducing hardware investment costs for grid decarbonization management.

[0072] 3. Physical consistency constraints ensure the credibility of the results and improve robustness. By introducing Kirchhoff's carbon flow conservation law as a physical regularization term during training, and in conjunction with a sentinel mechanism, the output is forced to satisfy the conservation of energy and carbon mass. This not only prevents the black-box model from producing predictions that violate physical common sense in sparse data regions, but also gives the model a clear physical meaning, making its results more easily accepted and trusted by power dispatching and operation personnel.

Claims

1. A method for carbon potential tracking and prediction in power systems based on a dual-drive approach of physics and data, characterized in that, include: Step S1: Collect source-side physical parameters, energy storage-side physical parameters, grid-side operating parameters, and load-side operating parameters in the power system, and preprocess the collected parameters to obtain a time-series feature dataset; Step S2: Construct a source-side unit two-layer collaborative graph based on the time-series feature dataset and generate a normalized adjacency matrix. The source-side unit two-layer collaborative graph includes physical connection edges representing electrical position coupling and logical collaborative edges representing similar units when responding to the system's automatic power generation control commands, thereby mapping the dual coupling relationship of the unit group in the physical space and logical control space. Step S3: Construct a physically guided source-side graph neural network model based on the source-side unit dual-layer collaborative graph, and output the predicted source-side dynamic carbon emission factor and node carbon injection sequence through the source-side graph neural network model; Step S4: Based on the branch power flow section data of the power grid side and the predicted source-side dynamic carbon emission factor and node carbon injection sequence, a high-dimensional input tensor representing the carbon flow diffusion law of the whole network and the source-side dynamic carbon potential prediction information and the real-time electrical operation status data of the power grid side is constructed through the spatiotemporal graph attention mechanism. Step S5: Construct a full-network spatiotemporal graph neural network model, using high-dimensional input tensors as the input source, and employing a parallel architecture of backbone extrapolation and expert correction to extract spatial and temporal characteristics, and extrapolate the spatiotemporal laws of carbon potential distribution changes with the current throughout the entire process, thereby achieving dynamic tracking of carbon potential. Step S6: In the process of spatiotemporal law deduction, Kirchhoff's carbon flow conservation law is introduced as a physical regularization term. A hybrid loss function containing the physical regularization term is constructed. During the backpropagation process of training the spatiotemporal graph neural network model, the optimal sentinel mechanism is executed, and finally, the dynamic carbon potential prediction result of the load side after physical verification is output.

2. The method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches as described in claim 1, characterized in that, Step S1 specifically includes: Collect source-side physical parameters, including: the unit's rated capacity. Minimum technical output Active power output of the unit The gradient rate, characterizing the transient adjustment intensity, was obtained by first-order difference calculation. Simultaneously, the lower heating value of the unit's fuel and the design coal consumption and gas consumption curves are read, including the ramp rate. The calculation formula is: in, for The active power output of the time unit and The difference in active power output between the time units The sampling time interval; Collect physical parameters from the energy storage side, including: the charging and discharging power of all energy storage power stations in the network. Real-time state of charge and energy storage cycle efficiency ; Collect parameters from both the grid side and the load side, including: voltage amplitude at all bus nodes of the entire network. Voltage phase angle Active load reactive load And the active power flow direction of key branch sections. With traffic data ; The collected parameters were cleaned, aligned, and Z-score standardized to obtain a time-series feature dataset containing physical mechanism attributes and electrical operating states.

3. The power system carbon potential tracking and prediction method based on both physical and data-driven approaches according to claim 2, characterized in that, Step S2 specifically includes: A two-layer collaborative graph of source-side units is constructed based on time-series feature datasets. If two generators are connected to the same bus node of the power network, physical connection edges are established in the graph structure to represent the coupling relationship of electrical locations. If two generator sets have the same fuel type attribute, then logical cooperative edges are constructed in the graph structure to characterize the group cooperative response behavior of similar units in the process of system load distribution and peak shaving. Based on the aforementioned physical connection edges and logical coordination edges, an adjacency matrix for the power supply units is generated. And calculate the corresponding degree matrix. Furthermore, combined with the identity matrix The normalized adjacency matrix is ​​calculated. .

4. The power system carbon potential tracking and prediction method based on both physical and data-driven approaches according to claim 3, characterized in that, Normalized adjacency matrix Calculated using the following formula: in, For physical connection edges, For logical connection edges, The weight coefficients of the logical collaborative edges.

5. The power system carbon potential tracking and prediction method based on both physical and data-driven approaches according to claim 4, characterized in that, In step S3, the source-side graph neural network model includes a basic prediction module, a graph collaborative feature module, and a residual error correction module; The basic prediction module is used to calculate basic physical prediction values ​​based on the unit's operating status and the embedded differentiable physical mechanism layer. The expression is: in, As the baseline value, This represents the value that degrades combustion efficiency. The first transient penalty coefficient characterizing the unit's consumption characteristics. For the coefficients of the constant term, This represents the coefficient for the transient ramp penalty mechanism. The learnable second transient penalty coefficient, This is the transient sensitivity coefficient; The graph collaboration feature module is used to build upon the constructed normalized adjacency matrix. By combining spatiotemporal convolutional networks with graph attention mechanisms and temporal convolutional networks, the operating status features of neighboring units are aggregated to extract unit group collaborative feature vectors; and fuel feature vectors are output based on unit fuel attribute data. The residual error correction module is used for: First, cluster the unit group collaborative feature vector and fuel feature vector, then input them into an independent residual error correction network to output a nonlinear residual correction amount. ; Then the basic physical prediction values With nonlinear residual correction By superimposing the results, the source-side dynamic carbon emission factor is obtained. The expression is: Finally, based on With the active power output of the generator unit Compute node carbon injection sequence The expression is: 。 6. The method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches as described in claim 5, characterized in that, Step S4 specifically includes: Step S41, Source-side dynamic carbon emission factor and carbon injection sequence Mapped to the corresponding mesh node, and injected as a dynamic attribute tensor of the source node; Step S42, the charging and discharging power of the entire network's energy storage power stations. With real-time state of charge As a dynamic feature injection tensor for energy storage nodes; Step S43: Merge the voltage amplitude of all bus nodes in the network. Voltage phase angle Active load reactive load And constructing a dynamic power flow adjacency matrix based on real-time power flow calculation results. ; Step S44: Obtain the overall macro-level situation characteristics of the power system through a network-wide macro-level situation broadcast. These characteristics include the total load of the power system. Total reserve ratio Total climbing requirements and the total energy storage backup capacity of the system ; Step S45 involves tensor fusion of the parameters from steps S41 to S44, followed by multi-channel slice stacking and dimensionality normalization verification to obtain a high-dimensional input tensor. This high-dimensional input tensor includes the source-side dynamic carbon emission factor. Carbon injection sequence The charging and discharging power of the entire network's energy storage power stations Real-time state of charge Voltage amplitude of all bus nodes in the network Voltage phase angle Active load reactive load Dynamic power flow adjacency matrix Overall macro-level characteristics of the network.

7. The method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches as described in claim 6, characterized in that, Step S5 specifically includes: Step S51: Input the high-dimensional input tensor into the backbone inference pathway of the whole-network spatiotemporal graph neural network model. Spatial extrapolation is performed, and a deep residual graph convolutional network is used to extract the spatial topological feature matrix of the entire network carbon flow. The expression is as follows: in, and They are respectively Time of the first and the The input tensor of the layer network, For activation function, This is the learnable parameter matrix of the spatial convolutional layer; Step S52: The LSTM unit takes the entire network carbon flow spatial topology feature matrix as input and outputs the global overall prediction result of the load-side carbon potential. ; Step S53: In parallel, start the expert correction path in the whole-network spatiotemporal graph neural network model. To obtain the local node prediction results of the carbon potential on the load side. ; Step S54, using the overall prediction results of the load-side carbon potential. Using the baseline value, the predicted results of local nodes of carbon potential on the load side are used. As a correction factor, the dynamic carbon potential on the output load side. This enables dynamic tracking of carbon potential. The expression is: in, For adaptive gating coefficients, This indicates element-wise multiplication.

8. The method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches according to claim 7, characterized in that, In step S54, the adaptive gating coefficient Through learnable gating mechanisms Adjustment, gating mechanism The expression is: 。 9. The method for carbon potential tracking and prediction of power systems based on both physical and data-driven approaches as described in claim 7, characterized in that, In step S6, the hybrid loss function The expression is: in, Huber loss for data-driven items, For physical regularization, These are the weighting coefficients for physical constraints. To test the time-series index, To verify the node index, For the carbon flux of the upstream line, This refers to the carbon flux of the downstream circuit.

Images