Electric power carbon emission accurate measurement and optimization method based on artificial intelligence

Abstract

The invention belongs to the technical field of gas compressors, and discloses an electric power carbon emission accurate measurement and optimization method based on artificial intelligence, and the method comprises the steps: data collection and processing: collecting historical power generation, load, tidal current and carbon emission data, and carrying out the cleaning and standardization; training a dynamic carbon emission factor model, training a carbon emission factor prediction model of each thermal power generating unit by using historical data, and predicting by using the carbon emission factor prediction model to obtain a dynamic carbon emission factor; and training a GNN carbon flow model, and training the GNN model to estimate node carbon intensity by using the power grid data and the dynamic carbon emission factors. According to the invention, two core problems of inaccurate measurement and difficult optimization of the carbon emission of the electric power system are solved. Accurate carbon perception is realized through a dynamic carbon emission factor and a GNN carbon flow model, and an optimal scheduling strategy considering economy, safety and low carbon is searched in a complex power grid environment through reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of electricity carbon emission measurement technology, and in particular to an artificial intelligence-based method for accurate measurement and optimization of electricity carbon emissions. Background Technology

[0002] The World Meteorological Organization (WMO) Greenhouse Gas Report indicates that the global average atmospheric CO2 concentration in 2018 was 407.8% (PPM). The Paris Agreement, signed in 2015, is a milestone in the international community's response to climate change, clearly defining the long-term goal of global climate change governance: to limit the increase in global average temperature to well below 2°C above pre-industrial levels, and to endeavor to limit the temperature increase to 1.5°C above pre-industrial levels. However, existing fossil fuel energy infrastructure, if operated at its normal lifespan and usage patterns, cannot achieve this goal. The power industry is a significant contributor to carbon emissions and a key area for carbon reduction. Traditional methods use static average carbon emission factors for regions or countries, ignoring the significant differences in carbon emission efficiency of thermal power units under different operating conditions (such as load factor and start-up / shutdown states), thus leading to inaccurate carbon emission calculations. Summary of the Invention

[0003] To overcome the technical deficiencies of existing technologies, this invention provides an artificial intelligence-based method for accurate calculation and optimization of electricity carbon emissions. It achieves accurate "carbon sensing" through dynamic carbon emission factors and GNN carbon flow models, making the calculation results more accurate.

[0004] The technical solution adopted in this invention is: an artificial intelligence-based method for accurate calculation and optimization of carbon emissions from electricity, comprising the following steps: Step 1: Data Acquisition and Processing. Collect historical power generation, load, power flow, and carbon emission data, and clean and standardize them. Step 2: Train the dynamic carbon emission factor model. Use historical data to train the carbon emission factor prediction model for each thermal power unit, and use the carbon emission factor prediction model to predict the dynamic carbon emission factor. Step 3: Train the GNN carbon flow model. Using grid data and dynamic carbon emission factors, train the GNN model to estimate node carbon intensity. Step 4: Construct a reinforcement learning environment: Based on the power grid model, load forecasting, renewable energy forecasting, and carbon flow model, construct an RL environment; train the RL agent in the environment to learn the optimal scheduling strategy.

[0005] Preferably, the power generation data includes the power generation capacity (MW), fuel type (for thermal power), and unit efficiency of various power plants (thermal power, hydropower, wind power, photovoltaic, etc.); the power grid data includes the power grid topology, line power flow, and node injection power; and the load data includes the electricity load of each region / user.

[0006] Preferably, the method for constructing the dynamic carbon emission factor model is to use a gradient boosting decision tree (GBDT) or a neural network to fit the carbon emission rate of the generator set. Let the carbon emission rate of unit i at time t be , then we have , in It is the active power of unit i at time t. Other characteristics (such as fuel calorific value, ambient temperature, and unit health status); Use historical data (including , The model f is trained using the actual measured carbon emissions, where the loss function is the mean squared error: In the formula These are model predictions.

[0007] Preferably, the method for training the GNN carbon flow model using grid data and dynamic carbon emission factors to estimate node carbon intensity is as follows: Step 1: Construct the carbon flow equation; Step 2: Perform carbon flow calculations at the load nodes; Step 3: Carbon flow tracing based on graph neural network (GNN) to obtain the carbon intensity of each node.

[0008] Preferably, the method for constructing the carbon flow equation assumes that the power grid has N nodes, and the injected power of node i is... The generator injected power is The load power is Then, the carbon flux of node i is |: In the above formula, is the generator carbon emission factor of node i. For load nodes, =0; According to power flow calculations, the relationship between node-injected power and line power flow is described by the power transmission distribution factor matrix, and the power flow on line l is... It can be represented as: ; Similarly, we can define the distribution of carbon flow in the lines. The injected carbon flow at node i (i.e., the carbon emissions from generator i) will be distributed to each line according to the power distribution ratio. Therefore, the carbon flow on line l... for: .

[0009] The method for calculating the carbon flow at load nodes in step 2 is as follows: the carbon consumption of load node j is equal to the carbon flow flowing into that node. According to Kirchhoff's laws, the sum of the carbon flows flowing into node j equals the carbon flow consumed by that load. We can obtain the carbon flow of each load node by solving the carbon flow equation, defining the node carbon intensity (unit: tons of CO2 / MWh) as... Let represent the carbon emissions per unit of electricity consumed by node j. According to the carbon flow conservation law, the carbon flow balance equation for node j is: , In this way, by constructing a carbon flow model of the entire network, the carbon intensity of each node can be solved.

[0010] Preferably, the method for obtaining the carbon intensity of each node through carbon flow tracing based on graph neural networks (GNNs) involves constructing the power grid as a graph. Node V includes generating nodes and load nodes, and edge E is a transmission line. Node characteristics include node type (generating / loading), power injection / consumption, and carbon emission factor (generating node). Edge characteristics include line power flow and reactance. Construct a GNN model to simulate the flow of carbon in the power grid through a message passing mechanism, and update the representation of nodes and edges at each layer of GNN; For node i, its hidden state Updated to: , in It is a neighbor of node i. It is an edge feature. and It's a neural network, which ultimately maps the hidden states of nodes to their carbon intensity through an output layer. ; The carbon intensity result of each node calculated using the carbon flow equation is used as a supervision signal to train the GNN model, thus obtaining the GNN carbon flow model.

[0011] The beneficial effects of this invention are: 1. This invention solves the two core problems of "inaccurate measurement" and "difficult optimization" of carbon emissions in power systems. It achieves accurate "carbon sensing" through dynamic carbon emission factors and GNN carbon flow models, and then uses reinforcement learning to find the optimal scheduling strategy that balances economy, safety and low carbon emissions in complex power grid environments.

[0012] 2. This invention transforms the carbon emissions of the power system from a vague, ex-post statistical macro-indicator into a micro-controllable variable that can be measured in real time, accurately, and proactively optimized, thereby greatly improving the precision and intelligence of the low-carbon operation of the power system.

[0013] 3. This invention can predict carbon emission factors at future moments, providing crucial input for subsequent forward-looking optimized scheduling. This enables the scheduling system not only to know the current "carbon concentration" but also to predict future trends, pinpointing the responsibility for carbon emissions to specific generator units, and providing solid data support for carbon footprint tracking, carbon trading, and precise carbon reduction.

[0014] 4. This invention integrates physical mechanisms with data-driven approaches: Generative Neural Networks (GNNs) can naturally capture the physical laws governing the topology and power flow of the power grid. It treats "carbon flow" as a virtual flow dependent on "power flow," enabling the model to possess both data-driven learning capabilities and respect physical constraints, resulting in estimations that are more realistic. It can provide accurate, spatiotemporally differentiated carbon footprint calculations for the electricity consumption of each factory, building, and even charging station. Compared to "black box" models, GNNs work by learning the relationships between nodes and edges, and their output can be interpreted to some extent through the power grid topology and power flow direction, enhancing the model's credibility. Attached Figure Description

[0015] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the various embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this invention to facilitate a better understanding of this application. However, the technical solutions claimed in the claims of this application can be implemented even without these technical details and with various variations and modifications based on the following embodiments.

[0018] like Figure 1 As shown, this embodiment provides a method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence, including the following steps: Step 1: Data Acquisition and Processing. Historical power generation, load, power flow, and carbon emission data are collected, cleaned, and standardized, laying a high-quality data foundation. Cleaning and standardization ensure the consistency and reliability of subsequent AI model input data, avoiding the problem of "garbage in, garbage out," which is a prerequisite for achieving "accurate calculations."

[0019] Step 2: Train the dynamic carbon emission factor model. Use historical data to train a carbon emission factor prediction model for each thermal power unit, and then use this model to predict the dynamic carbon emission factor. Moving from "static coarse estimation" to "dynamic precision," traditional methods use regional or national static average carbon emission factors, ignoring the significant differences in carbon emission efficiency of thermal power units under different operating conditions (such as load factor and start-up / shutdown status). This method models each unit individually, more accurately reflecting its real-time carbon emission level. It can predict carbon emission factors for future moments, providing crucial input for subsequent forward-looking optimized scheduling. This allows the scheduling system to not only know the current "carbon concentration" but also predict future trends. By assigning responsibility for carbon emissions precisely to specific generating units, it provides solid data support for carbon footprint tracking, carbon trading, and precise carbon reduction.

[0020] Step 3: Train the GNN carbon flow model. Using grid data and dynamic carbon emission factors, train the GNN model to estimate nodal carbon intensity; from "source-end accounting" to "precise grid-load-end source tracing"; fusion of physical mechanisms and data-driven approaches: GNNs can naturally capture the topology of the power grid and the physical laws of power flow. It treats "carbon flow" as a virtual flow dependent on "power flow," enabling the model to have both data-driven learning capabilities and respect physical constraints, resulting in more realistic estimation results; generating nodal carbon intensity: This is the key output of this method. It can calculate the real-time carbon intensity (gCO2 / kWh) of each node (bus) in the power grid. This means that it can provide accurate, spatiotemporally differentiated carbon footprint accounting for the electricity consumption of each factory, each building, and even each charging pile; enhanced interpretability: Compared to "black box" models, GNNs work by learning the relationships between nodes and edges, and their output results can be interpreted to some extent through the grid topology and power flow direction, enhancing the model's credibility.

[0021] Step 4: Constructing a Reinforcement Learning Environment: Based on the power grid model, load forecasting, renewable energy forecasting, and carbon flow model, construct an RL environment; train the RL agent in this environment to learn the optimal scheduling strategy. This progresses from "single-point optimization" or "rule-based optimization" to "global adaptive optimization"; addressing high-dimensional complex problems: Power system scheduling is a high-dimensional, nonlinear, and multi-constraint complex decision-making problem. Through continuous interaction with the environment, the RL agent can learn complex strategies far exceeding the experience of human experts, effectively handling the uncertainties of renewable energy and load. Because the environmental model includes load and renewable energy forecasts as well as dynamic carbon flow information, the RL agent learns a complete set of strategies, not just a static solution. It can adaptively adjust the scheduling strategy according to different system states (such as weather and load levels) to achieve global, long-term low-carbon economic scheduling.

[0022] The power generation data includes the power generation capacity (MW), fuel type (for thermal power), and unit efficiency of various power plants (thermal power, hydropower, wind power, photovoltaic, etc.); the power grid data includes the power grid topology, line power flow, and node injection power; and the load data includes the electricity load of each region / user.

[0023] The method for constructing the dynamic carbon emission factor model is to use gradient boosting decision tree (GBDT) or neural network to fit the carbon emission rate of the generator set. Let the carbon emission rate of unit i at time t be , then we have , in It is the active power of unit i at time t. Other characteristics (such as fuel calorific value, ambient temperature, and unit health status); Use historical data (including , The model f is trained using the actual measured carbon emissions, where the loss function is the mean squared error: In the formula These are model predictions. The specific step of "using gradient boosting decision trees or neural networks to fit the carbon emission rate of generator units" has significant advantages in its construction method, being key to achieving the leap from "static averaging" to "dynamic accuracy." Ultra-high accuracy: It captures complex nonlinear relationships, automatically learning the complex, implicit mappings between these factors and carbon emission rates from the data without requiring manually pre-specified formulas. This capability is difficult to achieve with traditional linear regression or physical modeling, thus achieving unprecedented prediction accuracy.

[0024] Traditional static factors are constants, while the factors generated by this method are functions that change over time. By inputting predicted operating parameters for future moments (such as planned output and predicted ambient temperature) into a pre-trained GBDT or neural network, the dynamic carbon emission factor for future moments can be directly predicted. This makes proactive low-carbon scheduling optimization possible.

[0025] Using a data-driven and adaptive approach, it accurately reproduces the complex carbon emission behavior of each thermal power unit in the real world, transforming the carbon emission factor from a rough statistical constant into a high-precision, predictable dynamic variable, laying the most solid foundation for the accurate carbon measurement and optimization of the entire power system.

[0026] The method described above for training a GNN carbon flow model, using grid data and dynamic carbon emission factors, to train a GNN model to estimate node carbon intensity, is as follows: Step 1: Construct the carbon flow equation; Step 2: Calculate the carbon flow at the load nodes; Step 3: Carbon flow tracing based on graph neural networks (GNNs) to obtain the carbon intensity of each node. Traditional "average factor method" or "current tracing method" are either too coarse or computationally complex and lack the generalization ability of machine learning models; The core inputs to this method are dynamic carbon emission factors (source information) and power grid flow data (path information).

[0027] Through learning, GNNs can accurately quantify how the carbon emissions associated with the power generated by each thermal power unit are transported to various load nodes like water through the complex power grid topology.

[0028] The method for constructing the carbon flow equation is as follows: assuming the power grid has N nodes, and the injected power at node i is... The generator injected power is The load power is Then, the carbon flux of node i is |: In the above formula, is the generator carbon emission factor of node i. For load nodes, =0; According to power flow calculations, the relationship between node-injected power and line power flow is described by the power transmission distribution factor matrix, and the power flow on line l is... It can be represented as: ; Similarly, we can define the distribution of carbon flow in the lines. The injected carbon flow at node i (i.e., the carbon emissions from generator i) will be distributed to each line according to the power distribution ratio. Therefore, the carbon flow on line l... for: .

[0029] The method for calculating the carbon flow at load nodes in step 2 is as follows: the carbon consumption of load node j is equal to the carbon flow flowing into that node. According to Kirchhoff's laws, the sum of the carbon flows flowing into node j equals the carbon flow consumed by that load. We can obtain the carbon flow of each load node by solving the carbon flow equation, defining the node carbon intensity (unit: tons of CO2 / MWh) as... Let represent the carbon emissions per unit of electricity consumed by node j. According to the carbon flow conservation law, the carbon flow balance equation for node j is: , In this way, by constructing a carbon flow model of the entire network, the carbon intensity of each node can be solved.

[0030] The method for carbon flow tracing based on graph neural networks (GNNs) to obtain the carbon intensity of each node involves constructing the power grid as a graph. Node V includes generating nodes and load nodes, and edge E is a transmission line. Node characteristics include node type (generating / loading), power injection / consumption, and carbon emission factor (generating node). Edge characteristics include line power flow and reactance. Construct a GNN model to simulate the flow of carbon in the power grid through a message passing mechanism, and update the representation of nodes and edges at each layer of GNN; For node i, its hidden state Updated to: , in It is a neighbor of node i. It is an edge feature. and It's a neural network, which ultimately maps the hidden states of nodes to their carbon intensity through an output layer. ; The carbon intensity result of each node calculated using the carbon flow equation is used as a supervision signal to train the GNN model, thus obtaining the GNN carbon flow model.

[0031] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0032] Constructing a reinforcement learning environment: Based on power grid models, load forecasting, renewable energy forecasting, and carbon flow models, an RL environment is constructed; specifically, an RL agent is trained in this environment to learn the optimal scheduling strategy. Consider a multi-objective optimization problem: minimizing total carbon emissions and total operating costs while satisfying power supply and demand balance and grid security constraints. Decision variables include the output of each generator. Load reduction (if any), etc.

[0033] Set the objective function as follows: , Total carbon emissions = Total operating cost = , It is the cost function of generator i; Set the constraint adjustment as follows: Power balance: ; Power flow constraints on the line: ; Generator output upper and lower limits: ; Optimization using reinforcement learning: Since the power system optimization problem is complex and nonlinear, we can use reinforcement learning (RL) to solve it; the state s includes: the upper and lower limits of each generator output, load demand, renewable energy forecast output, grid topology (optional) and current carbon emission factor, and action a is the output setpoint of each generator (continuous action) or discrete action (such as increasing / decreasing output).

[0034] The reward function is designed as follows: Using RL algorithms suitable for continuous action spaces, such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO), demand-side response (DR) is incorporated into the optimization model. This allows loads to adjust their electricity consumption behavior based on electricity prices or carbon signals, categorizing the load into stationary and flexible loads. The electricity consumption of flexible loads can shift over time, and their utility function is... ,in This refers to the load during time period t. In the optimization problem, the elastic load is used as a decision variable, and a user utility term is added to the objective function. Generative AI (such as VAE and GAN) is used to generate various possible load and renewable energy output scenarios to train RL agents, improving their robustness. This progresses from "single-point optimization" or "rule-based optimization" to "global adaptive optimization," addressing high-dimensional complex problems: Power system dispatching is a complex decision-making problem with high dimensions, nonlinearity, and multiple constraints. Through continuous interaction with the environment, RL agents can learn complex strategies far exceeding the experience of human experts, effectively handling the uncertainties of renewable energy and load. Because the environmental model includes load and renewable energy forecasts as well as dynamic carbon flow information, the RL agent learns a complete set of strategies, not just a static solution. It can adaptively adjust dispatch strategies according to different system states (such as weather and load levels) to achieve global, long-term low-carbon economic dispatching.

[0035] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Data Acquisition and Processing. Collect historical power generation, load, power flow, and carbon emission data, and clean and standardize them. Step 2: Train the dynamic carbon emission factor model. Use historical data to train the carbon emission factor prediction model for each thermal power unit, and use the carbon emission factor prediction model to predict the dynamic carbon emission factor. Step 3: Train the GNN carbon flow model. Using grid data and dynamic carbon emission factors, train the GNN model to estimate node carbon intensity. Step 4: Construct a reinforcement learning environment: Based on the power grid model, load forecasting, renewable energy forecasting, and carbon flow model, construct an RL environment; train the RL agent in the environment to learn the optimal scheduling strategy.

2. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 1, characterized in that: The power generation data includes the power generation capacity (MW), fuel type (for thermal power), and unit efficiency of various power plants (thermal power, hydropower, wind power, photovoltaic, etc.); the power grid data includes the power grid topology, line power flow, and node injection power; and the load data includes the electricity load of each region / user.

3. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 1, characterized in that: The method for constructing the dynamic carbon emission factor model is to use gradient boosting decision tree (GBDT) or neural network to fit the carbon emission rate of the generator set. Let the carbon emission rate of unit i at time t be , then we have , in It is the active power of unit i at time t. Other characteristics (such as fuel calorific value, ambient temperature, and unit health status); Use historical data (including , The model f is trained using the actual measured carbon emissions, where the loss function is the mean squared error: In the formula These are model predictions.

4. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 1, characterized in that: The method described above for training a GNN carbon flow model, using grid data and dynamic carbon emission factors, to train a GNN model to estimate node carbon intensity, is as follows: Step 1: Construct the carbon flow equation; Step 2: Calculate the carbon flow at the load nodes; Step 3: Carbon flow tracing based on graph neural network (GNN) to obtain the carbon intensity of each node.

5. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 4, characterized in that: The method for constructing the carbon flow equation is as follows: assuming the power grid has N nodes, and the injected power at node i is... The generator injected power is The load power is Then, the carbon flux of node i is |: In the above formula, is the generator carbon emission factor of node i. For load nodes, =0; According to power flow calculations, the relationship between node-injected power and line power flow is described by the power transmission distribution factor matrix, and the power flow on line l is... It can be represented as: ; Similarly, we can define the distribution of carbon flow in the lines. The injected carbon flow at node i (i.e., the carbon emissions from generator i) will be distributed to each line according to the power distribution ratio. Therefore, the carbon flow on line l... for: .

6. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 4, characterized in that: The method for calculating the carbon flow at load nodes in step 2 is as follows: the carbon consumption of load node j is equal to the carbon flow flowing into that node. According to Kirchhoff's laws, the sum of the carbon flows flowing into node j equals the carbon flow consumed by that load. We can obtain the carbon flow of each load node by solving the carbon flow equation, defining the node carbon intensity (unit: tons of CO2 / MWh) as... Let represent the carbon emissions per unit of electricity consumed by node j. According to the carbon flow conservation law, the carbon flow balance equation for node j is: ， In this way, by constructing a carbon flow model of the entire network, the carbon intensity of each node can be solved.

7. The method for accurate calculation and optimization of electricity carbon emissions based on artificial intelligence according to claim 1, characterized in that: The method for carbon flow tracing based on graph neural networks (GNNs) to obtain the carbon intensity of each node involves constructing the power grid as a graph. Node V includes generating nodes and load nodes, and edge E is a transmission line. Node characteristics include node type (generating / loading), power injection / consumption, and carbon emission factor (generating node). Edge characteristics include line power flow and reactance. Construct a GNN model to simulate the flow of carbon in the power grid through a message passing mechanism, and update the representation of nodes and edges at each layer of GNN; For node i, its hidden state Updated to: ， in It is a neighbor of node i. It is an edge feature. and It's a neural network, which ultimately maps the hidden states of nodes to their carbon intensity through an output layer. ; The carbon intensity result of each node calculated using the carbon flow equation is used as a supervision signal to train the GNN model, thus obtaining the GNN carbon flow model.

Images