Ai-enhanced electro-carbon dynamic optimization method and system

Abstract

The application belongs to the technical field of industrial control, and particularly relates to an AI-enhanced electric carbon dynamic optimization method and system, which comprises the following steps: acquiring instantaneous active power under the running state of an air compressor and an external regional power grid instantaneous carbon emission factor; acquiring air inlet temperature and air inlet pressure at the air inlet section of the air compressor, and exhaust pressure and exhaust temperature at the exhaust section; combining the ideal isentropic compression ratio work and the actual compression ratio work calculated based on the above-mentioned physical parameters to determine the isentropic efficiency; calculating and acquiring a penalty factor by using the isentropic efficiency and a penalty amplification coefficient obtained through offline calibration; performing integral operation on each characteristic variable to obtain dynamic carbon emission, and inputting the dynamic carbon emission into a locally deployed reinforcement learning prediction model to output an execution action and send the execution action to a controller for execution; and the application establishes a mapping between internal thermodynamic state evolution and external consumption, accurately identifies the decaying equipment with increased carbon emission, and realizes automatic control.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology. More specifically, this invention relates to an AI-enhanced method and system for dynamic optimization of carbon dioxide emissions. Background Technology

[0002] In the enterprise energy management and electricity carbon emission monitoring industry, accurate calculation of carbon emissions generated during the operation of industrial equipment is an important basis for enterprises to adjust their daily production plans and meet environmental compliance requirements. In actual industrial manufacturing workshops, high-energy-consuming equipment such as air compressors and chillers operate at high frequency, and the electricity they consume is the main source of carbon emissions. In order to understand the actual carbon emissions in the workshop during production, the industry generally adopts the method of collecting electricity meter data and converting it into carbon emission values ​​for automated monitoring.

[0003] The Intergovernmental Panel on Climate Change (IPCC) published the "IPCC Guidelines for National Greenhouse Gas Inventories 2006." The guidelines established a technical methodology that calculates total greenhouse gas emissions by directly multiplying the electricity consumed by operating equipment by a pre-defined carbon emission factor. However, this methodology was originally designed to aid in macro-level annual data collection and compliance reporting. Its granularity is very coarse. In real-time control at the factory floor, the grid connection rates of clean energy sources such as wind and solar power fluctuate dramatically throughout the day. This method, relying on a fixed factor, cannot reflect the true changes in the cleanliness of electricity over different time periods, leading to discrepancies between the calculated real-time carbon emissions and the actual situation.

[0004] The "Guidelines for Accounting and Reporting Greenhouse Gas Emissions of Enterprises" and the "Guidelines for Accounting and Reporting Greenhouse Gas Emissions of Power Generation Enterprises in China (Trial)" issued by China's Ministry of Ecology and Environment provide a calculation model where emissions from purchased and used electricity equal actual electricity consumption multiplied by the average carbon emission factor of the power grid, and this factor is usually fixed at a certain constant within a specific year. However, these technical documents primarily serve long-term annual compliance verification for enterprises. When applied to dynamic monitoring at the second or minute level in workshops, the use of a static annual factor fails to capture the dynamic evolution of the power grid's carbon intensity during daily peak or off-peak electricity consumption, leaving subsequent carbon emission optimization strategies that rely on this data without an accurate real-time data foundation.

[0005] In existing technologies, although carbon emission calculations are typically performed by collecting electricity meter data from equipment and combining it with officially published factors, most of these solutions rely solely on a one-dimensional, proportional conversion of surface-level electricity consumption. Industrial equipment, such as air compressors, often experience physical degradation after long-term operation, such as filter blockage, pipe leaks, or mechanical wear, leading to a decrease in internal heat conversion efficiency. In actual production lines, even if two identical machines consume the exact same amount of electricity, their effective air volume or cooling output can differ significantly due to variations in their mechanical health. Existing calculation methods completely disregard the specific physical processes of equipment operation, focusing only on the amount of externally input electrical energy. This makes the system prone to masking ineffective high carbon emissions caused by deteriorating mechanical efficiency. Because the system cannot detect energy efficiency degradation caused by internal heat and wear, it cannot accurately identify the abnormal equipment truly causing increased carbon emissions, thus failing to provide effective production scheduling and intervention suggestions, impacting the practicality of workshop carbon emission monitoring and dynamic adjustment. Summary of the Invention

[0006] To address the technical problems of the traditional static carbon emission accounting system being detached from the actual thermodynamic and physical decay laws of equipment, resulting in distorted carbon flow monitoring and a lack of accurate data for subsequent dynamic load optimization and scheduling, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides an AI-enhanced dynamic optimization method for electrical carbon emissions, including obtaining the instantaneous active power of an air compressor under operating conditions and obtaining the instantaneous carbon emission factor of the external regional power grid. The method further includes: obtaining the intake temperature and intake pressure at the air compressor intake section; obtaining the exhaust pressure and exhaust temperature at the air compressor exhaust section; calculating the ideal isentropic compression ratio based on the intake temperature, intake pressure, and exhaust pressure; calculating the actual compression ratio based on the intake temperature and exhaust temperature; determining the isentropic efficiency by calculating the ratio of the ideal isentropic compression ratio to the actual compression ratio; calculating the penalty factor using the isentropic efficiency and a penalty amplification factor; multiplying the instantaneous active power, the instantaneous carbon emission factor of the power grid, and the penalty factor and integrating them over a set time window to obtain the dynamic carbon emissions; inputting the dynamic carbon emissions into a locally deployed reinforcement learning prediction model to output an execution action and sending it to the air compressor controller.

[0008] This invention obtains the temperature and pressure at the air compressor's inlet and outlet sections, calculates the ideal isentropic compression work ratio and the actual compression work ratio by combining gas state parameters, and obtains the isentropic efficiency, which characterizes the actual operating energy efficiency level. The isentropic efficiency, which includes mechanical and physical attenuation, is converted into a penalty factor and integrated into the static carbon emission calculation mechanism. This allows the final integrated dynamic carbon emission to truly reflect the surge in ineffective carbon emissions caused by filter blockage or mechanical wear. It effectively overcomes the shortcomings of traditional single-factor conversion, which cannot detect the degradation of underlying hardware. This provides a data foundation with high physical discernibility for subsequent reinforcement learning prediction models, ensuring that the output execution actions can accurately suppress ineffective work under high-loss conditions, and improving the accuracy of carbon emission control for high-energy-consuming equipment in manufacturing workshops.

[0009] Preferably, the method further includes: deploying an edge computing gateway to establish a bidirectional data transmission channel with the air compressor controller; concurrently reading the underlying operating status register through a preset fixed sampling frequency; performing analog-to-digital conversion and data format alignment on discrete analog signals on various mechanical components to generate a continuous time series dataset; and adding clock synchronization timestamps to the intake temperature, the intake pressure, the exhaust pressure, the exhaust temperature, and the instantaneous active power.

[0010] Preferably, the ideal isentropic compression ratio is calculated by combining the intake air temperature, the intake air pressure, and the exhaust air pressure, including: In the formula, For ideal isentropic compression ratio work; It is the isentropic exponent; It is the gas constant; Intake air temperature; This refers to the intake pressure. This refers to the exhaust pressure.

[0011] Based on the adiabatic law of no heat exchange in the continuous flow of gas, this invention introduces measured state parameters such as intake pressure and intake temperature into the basic control volume energy equation for integral derivation, thereby constructing an ideal gas isentropic compression work model that is completely determined by objective boundary conditions. This derivation process eliminates the mechanical friction and fluid resistance losses that accompany the actual operating environment, establishes the technical work baseline required for an air compressor to complete a specific pressurization technical target under ideal conditions, and provides a physical reference anchor for quantitatively evaluating the degree of energy efficiency drop of real equipment.

[0012] Preferably, the actual compression ratio is calculated by combining the intake temperature and the exhaust temperature, including: calculating the difference between the exhaust temperature and the intake temperature to obtain the intake and exhaust temperature rise; and calculating the product of the intake and exhaust temperature rise and the specific heat capacity at constant pressure to obtain the actual compression ratio.

[0013] Preferably, the penalty factor is calculated using the isentropic efficiency and the penalty amplification factor, wherein the penalty factor is equal to , To penalize the amplification factor, It is isentropic efficiency.

[0014] This invention introduces a penalty amplification factor on top of the basic energy efficiency monitoring logic, converting the isentropic efficiency, which shows a monotonically decreasing trend, into a penalty factor that is dynamically amplified with the sub-health state of physical equipment. This mechanism substantially transforms the thermodynamic mutation characteristics into a numerical penalty control quantity at the carbon emission accounting level. This allows the system to apply an amplification effect based on this control quantity once the work performance of the underlying mechanical components deteriorates, so as to significantly distinguish between normal operating conditions and high-loss operating conditions, highlighting the core weight of equipment health in the energy management system.

[0015] Preferably, the dynamic carbon emissions are obtained by multiplying the instantaneous active power, the instantaneous carbon emission factor of the power grid, and the penalty factor, and then integrating the results over a set time window, including: In the formula, For a moment Dynamic carbon emissions; For periodic time intervals; For a moment The instantaneous active power; For a moment The instantaneous carbon emission factor of the power grid; For a moment The penalty factor.

[0016] This invention performs continuous definite integral calculations on the time axis of instantaneous active power, instantaneous carbon emission factor of the power grid, and penalty factor with physical adaptive adjustment capability within a set time window. This mechanism, which heterogeneously combines the characteristics of basic power consumption with the underlying thermodynamic evolution law, breaks through the barrier of the traditional static accounting method's insensitivity to abnormal energy consumption by capturing the dual dynamic evolution law of external energy cleanliness fluctuations and internal mechanical loss accumulation of equipment. It outputs dynamic carbon emissions that have both the fact of physical equipment degradation and the dual attributes of external energy constraints.

[0017] Preferably, the offline calibration method for the penalty amplification coefficient is as follows: extract the tensor mapping span ratio required for the input layer of the reinforcement learning prediction model; set the value of the severe working condition penalty factor equal to the product of the base working condition penalty factor value and the tensor mapping span ratio; construct a linear algebraic equation in one variable based on the equal correlation of this value; solve the algebraic equation to calculate the unique real solution of the penalty amplification coefficient.

[0018] Preferably, the method for obtaining the severe operating condition penalty factor and the benchmark operating condition penalty factor is as follows: under the benchmark operating conditions of a brand-new model, objective physical parameters are collected and a first isentropic efficiency benchmark value is calculated; the initial efficiency decay physical quantity is obtained by subtracting the first isentropic efficiency benchmark value from 1; the benchmark operating condition penalty factor is obtained by adding the initial efficiency decay physical quantity to 1; under deteriorating operating conditions, objective physical parameters are collected and a second isentropic efficiency threshold is calculated; the deteriorating efficiency decay physical quantity is obtained by subtracting the second isentropic efficiency threshold from 1; and the severe operating condition penalty factor is obtained by adding the deteriorating efficiency decay physical quantity to 1.

[0019] Preferably, the dynamic carbon emissions are input into a locally deployed reinforcement learning prediction model to output an action, including: extracting the equipment production scheduling instruction sequence and the historical energy consumption feature dataset cached in local memory; aligning and concatenating the equipment production scheduling instruction sequence and the historical energy consumption feature dataset according to a unified system clock step to construct a time-series feature matrix; inputting the time-series feature matrix into a long short-term memory network model for gating mechanism tensor operations to output a future hourly energy consumption prediction trend dataset; obtaining the current buffer pressure value of the gas tank; and combining the dynamic carbon emissions, the current buffer pressure value of the gas tank, the instantaneous carbon emission factor of the power grid, and the... The future hourly energy consumption prediction trend dataset is subjected to structured tensor concatenation to generate a state space vector of a Markov decision process with discrete time steps. The state space vector is then input into a reinforcement learning prediction model constructed using a deep Q-network algorithm. After parameter matrix multiplication and addition operations and nonlinear activation mapping in the deep Q-network algorithm model, the action value distribution vector corresponding to the discrete instruction action is obtained. Extreme value retrieval and optimization logic is performed on the action value distribution vector to extract the parameter term with the largest algebraic weight. The instruction hard-mapped to the parameter term with the largest algebraic weight is set as the execution action of the current period.

[0020] This invention integrates the current buffer pressure value of the gas tank and the future energy consumption prediction trend dataset into the decision input mechanism of reinforcement learning. It is then concatenated with the dynamic carbon emissions, which include the equipment's decay status, to form a high-dimensional state space vector of a Markov decision process. This enables the model to output the optimal probability distribution of specific action commands based on a comprehensive consideration of the current gas storage safety margin, future demand load, and real-time high carbon penalties. After retrieving the maximum algebraic weight, it directly issues control commands, thus constructing a deterministic control link from state perception and digital algorithm optimization to the underlying mechanical displacement, achieving dynamic scheduling of high-carbon-emission equipment.

[0021] In a second aspect, the present invention provides an AI-enhanced dynamic optimization system for electric carbon emissions, comprising a processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the aforementioned AI-enhanced dynamic optimization method for electric carbon emissions.

[0022] By adopting the above technical solution, the AI-enhanced dynamic optimization method for carbon dioxide is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, facilitating its use.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention synchronously acquires objective parameters such as intake air temperature and exhaust pressure through an edge computing gateway, substitutes them into an ideal gas isentropic compression work model to derive isentropic efficiency, and then converts this isentropic efficiency into a penalty factor and multiplies it into a reference electric carbon sequence for definite integral calculation. This mechanism converts the thermodynamic parameter mutations caused by hardware degradation into numerical penalty control quantities with real-world guidance, and finally outputs dynamic carbon emissions that incorporate the facts of equipment degradation, thereby eliminating data interference caused by ineffective operation to the load scheduling model. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the AI-enhanced dynamic optimization method for electrocarbon in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses an AI-enhanced dynamic optimization method for electrocarbon, referring to... Figure 1 This includes steps S1-S3:

[0029] S1. Establish a bidirectional data transmission channel with the air compressor controller through the edge computing gateway to synchronously obtain the real-time operating condition dataset of the air compressor.

[0030] To address the operational status monitoring needs of high-energy-consuming equipment in industrial manufacturing workshops, an edge computing gateway was deployed at the physical layer. This gateway connects to the communication interface of the air compressor controller via standard industrial communication cables. Combined with the factory workshop's local area network architecture, the edge computing gateway establishes a bidirectional data transmission channel with the air compressor controller using the configured MQTT protocol, concurrently reading the underlying operational status registers at a preset fixed sampling frequency. As a hardware-level data aggregation node in the entire data acquisition chain, the edge computing gateway performs preliminary analog-to-digital conversion and data format alignment on the discrete analog signals generated by sensors distributed across various mechanical components of the air compressor, thereby constructing a continuous time-series dataset.

[0031] During the specific data acquisition process for the air compressor, the real-time operating condition dataset acquired by the edge computing gateway includes physical and electrical parameters. These parameters specifically include: the intake temperature continuously collected by the intake port temperature sensor, the intake pressure obtained by the intake port pressure transmitter, the exhaust pressure obtained by the exhaust port pressure transmitter, the exhaust temperature collected by the exhaust port temperature sensor, and the instantaneous active power read by the metering module located in the power supply circuit. The intake temperature, intake pressure, exhaust pressure, exhaust temperature, and instantaneous active power values ​​captured by each sensor are all uniformly stamped with a high-precision clock synchronization mark inside the edge computing gateway, thereby ensuring the absolute alignment of the multi-source heterogeneous physical and electrical parameters on the same time section.

[0032] In addition to data sensing of the air compressor's own operating conditions, the edge computing gateway synchronously acquires parameters characterizing the external energy environment. Specifically, the edge computing gateway initiates data retrieval requests to the energy monitoring platform in its region at specified periodic intervals through a preset external network communication interface to obtain the instantaneous carbon emission factor of the power grid in the current region. The energy monitoring platform includes, but is not limited to, the State Grid, the Southern Power Grid, local power trading centers, and third-party carbon asset management SaaS platforms. After obtaining the instantaneous carbon emission factor of the power grid, the edge computing gateway concatenates the data with the previously timestamped intake temperature, intake pressure, exhaust pressure, exhaust temperature, and instantaneous active power in its local cache. Finally, it generates a fused data packet containing complete internal physical operating conditions and external environmental parameters and sends it to the upper-level processing architecture of the system.

[0033] S2. Based on the ideal gas isentropic compression work model, calculate the ratio of ideal isentropic compression specific work to actual compression specific work to determine the isentropic efficiency. Use the isentropic efficiency and the offline calibrated penalty amplification factor to calculate the penalty factor. Combine the instantaneous active power and the instantaneous carbon emission factor of the power grid to obtain the dynamic carbon emission through integral calculation.

[0034] Traditional methods for calculating total greenhouse gas emissions by directly multiplying the electricity consumed by equipment operation by a preset emission factor completely ignore the specific physical processes of air compressor work. This leads to the system failing to detect ineffective high carbon emissions caused by deteriorating mechanical efficiency when air compressors experience physical degradation such as filter blockage, pipe leaks, or mechanical wear after long-term operation. Consequently, subsequent reinforcement learning prediction models suffer from data interference and distorted judgments due to missed carbon emission anomalies during load scheduling. Therefore, this invention, based on the energy conservation law of continuous gas flow in classical engineering thermodynamics and the ideal gas law... The system uses the equation of state to convert the instantaneous active power, intake temperature, intake pressure, exhaust pressure, and exhaust temperature collected by the edge computing gateway into ideal isentropic compression work and actual compression work. Then, based on the ideal gas isentropic compression work model, the instantaneous isentropic efficiency is determined by calculating the ratio of instantaneous actual compression work to ideal isentropic compression work. A quantitative causal mapping between the evolution of the internal thermodynamic state of the air compressor and external electrical energy consumption is established. The efficiency decay is used to penalize the basic carbon emission calculation, reflecting the internal enthalpy difference change caused by sub-health of the equipment, so that the calculated carbon emission index reflects the energy degradation inside the equipment.

[0035] In thermodynamics, the energy efficiency of an air compressor during actual operation is measured by its isentropic efficiency, which is defined as the isentropic compression work required for a gas to achieve the same pressure increase under ideal conditions. Compared with actual compression ratio The ratio of .

[0036] Therefore, in the first stage of the specific derivation, we first define that the air compressor draws gas from the intake pressure. Compress to exhaust pressure The required ideal isentropic compression work is specifically defined as follows: the air entering the air compressor is considered an ideal gas with constant specific heat capacity, and its flow process inside the cylinder is abstracted as an adiabatic process without mechanical friction and without heat exchange with the outside, i.e., an isentropic process. At this time, the state parameters of the gas at any flow cross-section satisfy the isentropic equation: In the formula, This is instantaneous pressure, with the dimension of Pascal (Pa). Specific volume of a gas, with dimensions in cubic meters per kilogram (m³ / kg). ); The isentropic exponent is dimensionless and ranges from 1.33 to 1.4. Since this embodiment treats the air entering the air compressor as an ideal gas with constant specific heat, the isentropic exponent in this embodiment is... ; This is the isentropic process constant.

[0037] According to the fluid dynamics control volume energy equation, the total technical work done by the equipment on the gas during the flow and compression process is determined by the integral form of the pressure-driven specific volume, and its integral expression is as follows: In the formula, The work done by the ideal isentropic compression is expressed in joules per kilogram (J / kg).

[0038] By rearranging and transforming the aforementioned isentropic equation, the specific volume of a gas can be expressed as a pressure factor: Substitute this specific volume expression into the integral expression of the total technical work mentioned above, and extract the constant term. and includes instantaneous pressure of The terms perform basic calculus operations, among which, The result of the integral is ,but , This is the intake pressure, measured in Pascals (Pa). The pressure is the exhaust pressure, and its dimension is Pascal (Pa).

[0039] For the expression of the specific work of ideal isentropic compression, the isentropic process constant that cannot be directly measured in the industrial field is considered. Using the known physical boundary conditions at the air compressor inlet section, the isentropic process constant is... Expressed as a product of intake state parameters: In the formula, The intake air volume is expressed in cubic meters per kilogram. Substitute the intake relationship into the aforementioned integral result and forcibly extract the intake pressure term. Outside the parentheses, after exponential simplification and term merging, the expression for the ideal isentropic compression work is simplified as follows: .

[0040] Due to the specific volume of the intake air during actual operation in the workshop Since these are physical quantities that cannot be directly measured online, in order to convert them into parameters that can be measured by temperature sensors, the classic ideal gas law is introduced, which is expressed as follows: In the formula, The total volume of the gas is expressed in cubic meters. ); The total mass of the gas is expressed in kilograms. ), is the gas constant, with dimensions of joules per kilogram Kelvin (J / (kg·K)); Intake air temperature, measured in Kelvin (K); while intake air specific volume... It is equal to the ratio of the total volume of the gas to the total mass of the gas, that is... The intake specific volume is obtained by combining the two methods. .

[0041] The intake specific volume Substituting the calculation formula into the simplified results above, the expression for the ideal isentropic compression work, determined by the three measurable physical parameters of intake temperature, intake pressure, and exhaust pressure, is finally derived:

[0042]

[0043] In the formula, The work done by ideal isentropic compression is expressed in joules per kilogram (J / kg). It is the isentropic exponent, dimensionless; is the gas constant, with dimensions of joules per kilogram Kelvin (J / (kg·K)); The intake air temperature is expressed in Kelvin (K). This is the intake pressure, measured in Pascals (Pa). The pressure is the exhaust pressure, and its dimension is Pascal (Pa).

[0044] In the second stage of the detailed derivation, an energy conversion model is further established for the actual operation of the air compressor. This model is used to calculate the isentropic efficiency, which reflects the true energy efficiency level of the equipment. Specifically, in real-world industrial scenarios, due to unavoidable internal leakage, mechanical friction, and resistance losses caused by airflow disturbances, the actual compression process is a polytropic flow process accompanied by entropy increase. According to the first law of thermodynamics, the actual work consumed by an open system in adiabatic flow is equal to the actual enthalpy difference between the gas inlet and outlet. Under the assumption of constant specific heat capacity of an ideal gas, the actual compression specific work is directly determined by the temperature rise of the inlet and outlet and the specific heat properties of the gas. Its quantitative expression is as follows:

[0045]

[0046] In the formula, The actual compression specific work is expressed in joules per kilogram (J / kg). It is the specific heat capacity at constant pressure. The specific heat capacity at constant pressure is the amount of heat required to raise the temperature of a unit mass of air by 1 K under constant pressure. It is usually 1005 and its dimension is joules per kilogram Kelvin (J / (kg·K)). The exhaust temperature is expressed in Kelvin (K).

[0047] In summary, based on the thermodynamic physical definition of compressor efficiency, the real-time isentropic efficiency of an air compressor... Defined as the ratio of the ideal isentropic compression specific work to the actual compression specific work when achieving the same boost technology goal, derived from the aforementioned two stages. Expressions and Dividing the expressions yields the isentropic efficiency. .

[0048] In the third stage of the specific derivation, the isentropic efficiency, which reflects the physical degradation of the equipment, needs to be embedded into the traditional carbon accounting architecture of pure electricity to complete the final reconstruction of the electric-carbon coupling metric: the traditional baseline carbon emission calculation is determined only by the product of the input power and the static factor, expressed as follows: , For a moment The instantaneous active power, in units of kilowatts (kW); For a moment The instantaneous carbon emission factor of the power grid is expressed in kilograms per kilowatt-hour (kg / (kW·h)). In terms of carbon emission calculation, due to energy efficiency degradation caused by equipment filter blockage, valve wear, or cooler scaling, the efficiency of converting electrical energy into effective gas pressure decreases, and excess energy is converted into waste heat, leading to increased exhaust temperature. An abnormal increase occurs when the actual compression specific power... Increased, leading to isentropic efficiency The isentropic efficiency decreases, meaning that the isentropic efficiency decreases as the performance of the air compressor degrades. Monotonically decreasing, thus causing efficiency to decay. The trend is upward; in order to transform this thermodynamic abrupt change characteristic into a penalty control quantity at the carbon emission level, a penalty amplification factor is introduced. A penalty factor with physical adaptive adjustment capability was constructed. Finally, this penalty factor is multiplied as a nonlinear correction term into the aforementioned baseline carbon emission calculation formula, and applied over a continuous time window. Performing a definite integral operation on this complex physical flow, the specific formula for calculating dynamic carbon emissions is as follows:

[0049]

[0050] In the formula, For a moment The dynamic carbon emissions, expressed in kilograms (kg); For periodic time intervals; For a moment The instantaneous active power, in units of kilowatts (kW); For a moment The instantaneous carbon emission factor of the power grid is expressed in kilograms per kilowatt-hour (kg / (kW·h)). The penalty amplification factor is dimensionless; For a moment Its isentropic efficiency is dimensionless.

[0051] Among them, time The expression for isentropic efficiency is:

[0052]

[0053] In the formula, It is the isentropic exponent, dimensionless; is the gas constant, with dimensions of joules per kilogram Kelvin (J / (kg·K)); For a moment The intake air temperature, in Kelvin (K); For a moment The intake pressure is expressed in Pascals (Pa). For a moment The exhaust pressure, in Pascals (Pa); It is the specific heat capacity at constant pressure, with dimensions of joules per kilogram Kelvin (J / (kg·K)). For a moment The exhaust temperature is expressed in Kelvin (K).

[0054] Among them, the edge computing gateway performs definite integral operations on the dynamic sequence of electric carbon after nonlinear correction by the penalty factor within a time window. This mechanism, which heterogeneously combines the thermodynamic evolution law with the electric carbon mapping, enables the determination of the marginal carbon increment of system energy consumption when the operating point of the device shifts to the high loss region due to energy efficiency deterioration. Ultimately, the dynamic carbon emissions containing the physical facts of device degradation are used as a highly distinguishable state feature.

[0055] Among them, the penalty amplification factor The determination was performed using an offline calibration procedure based on physical benchmark comparison and data fitting.

[0056] (1) In the first stage of the offline calibration program, the air compressor is placed under the benchmark operating condition of having completed mechanical maintenance and with brand new physical consumables and is running at full load. The intake temperature, intake pressure, exhaust pressure and exhaust temperature are collected synchronously and concurrently through the edge computing gateway. The above objectively collected values ​​are substituted into the isentropic efficiency derivation model to obtain the first isentropic efficiency benchmark value of the air compressor under the current ideal mechanical state. Then, by subtracting the first isentropic efficiency benchmark value from 1, the initial efficiency decay physical quantity of the system under the benchmark operating condition is calculated. Finally, by adding the initial efficiency decay physical quantity to 1, the penalty factor of the system under the benchmark operating condition is calculated.

[0057] (2) In the second stage of the offline calibration program, the real-time physical parameters of the air compressor are continuously monitored until the equipment shows, for example, a monotonous increase in exhaust temperature and approaches the preset thermodynamic safety alarm boundary. The system defines this cross-sectional state as a deteriorating operating condition. At this time, the edge computing gateway once again synchronously collects the current temperature and pressure parameters and calculates the second isentropic efficiency threshold that causes the air compressor to drop sharply under this deteriorating operating condition. Similarly, by subtracting the second isentropic efficiency threshold from 1, the physical quantity of deteriorated efficiency attenuation caused by substantial deterioration of mechanical performance of the air compressor is calculated. At this time, the physical quantity of deteriorated efficiency attenuation shows a significant numerical expansion compared with the initial efficiency attenuation physical quantity in the first stage. This data expansion phenomenon reflects the fact that the abnormal enthalpy difference work loss caused by the increased fluid resistance and mechanical friction inside the equipment. Finally, by adding the physical quantity of deteriorated efficiency attenuation to 1, the penalty factor of the system under severe operating conditions is calculated.

[0058] (3) In the third stage of the offline calibration process, the system performs the inverse algebraic equation for the penalty amplification coefficient based on the feature discrimination and normalization boundary requirements of the input state space tensor of the backend reinforcement learning prediction model. Specifically, in order to ensure that the reinforcement learning prediction model can capture the physical degradation of the device with sufficient numerical span when performing matrix operations and activation function mapping, thereby effectively changing the probability numerical distribution for the discrete action space, the tensor mapping span ratio required by the input layer of the reinforcement learning prediction model is pre-extracted. Accordingly, the penalty factor of the system under severe operating conditions is made to reach the penalty factor of the system under the reference operating conditions. This allows us to construct a linear algebraic equation in one variable that includes physical extremum constraints: The unique real solution to the penalty amplification factor is obtained by directly solving the algebraic equation; tensor mapping span ratio. In this embodiment, a tensor mapping span ratio greater than 1 is set to be required by the reinforcement learning prediction model for the state input. It is 1.3.

[0059] For example, when the first isentropic efficiency baseline value is 0.85, the initial efficiency decay physical quantity is: The penalty factor of the system under the baseline operating condition is When the second isentropic efficiency threshold is 0.65, the physical quantity that causes efficiency degradation is: The system's penalty factor under harsh operating conditions ; thereby constructing a linear equation in one variable: Solve the algebraic equation to calculate the penalty amplification factor. .

[0060] S3. Use a long short-term memory network model to extrapolate the future hourly energy consumption prediction trend dataset, combine it with dynamic carbon emissions to construct a state space vector, and input it into the reinforcement learning prediction model to output the execution action and send it to the air compressor controller.

[0061] The edge computing gateway periodically reads the workshop manufacturing execution system through a preset external communication interface to extract the equipment scheduling instruction sequence within a future set time window. Simultaneously, the edge computing gateway extracts historical energy consumption feature datasets of corresponding high-energy-consuming equipment cached in local memory. The system deploys a Long Short-Term Memory (LSTM) network model inside the edge computing gateway. As a specific parameter configuration embodiment, the LTM network model's network topology sequentially includes a feature dimension input layer, two LTM hidden layers each configured with 64 neurons, and a fully connected output layer. During the offline training phase, an adaptive moment estimation optimization algorithm is used, with the initial learning rate set to 0. .001; In the prediction calculation phase, the system aligns and concatenates the equipment scheduling instruction sequence and historical energy consumption feature dataset according to a unified system clock step, constructs a time-series feature matrix, and inputs it into the Long Short-Term Memory (LSTM) network model. The LSM network model uses its internal forget gate, input gate, and output gate to perform tensor operations on the time-series feature matrix in the hidden layer through a gating mechanism to forget the historical state and update the current state. Finally, it outputs a sequence containing multiple consecutive discrete time point estimated energy consumption scalars through a fully connected output layer. This sequence is defined by the system and written into the local memory, thus constituting the future hourly energy consumption prediction trend dataset.

[0062] After acquiring dynamic carbon emissions containing the actual thermodynamic decay characteristics of the devices, the edge computing gateway uses this as the basic input data stream to feed into the locally deployed reinforcement learning prediction model for matrix operation processing. To construct complete operating environment variables, the edge computing gateway obtains the current buffer pressure value of the gas tank in real time from the pressure transmitter installed on the gas tank via industrial communication cables. Simultaneously, the system combines the instantaneous carbon emission factor of the power grid obtained in the previous steps with the future hourly energy consumption prediction trend dataset generated and stored in the local memory, and performs structured tensor concatenation on the above multidimensional heterogeneous data. The concatenated data tensor is defined as a tensor with discrete time as the step size. The state space vector of the Markov decision process is defined as follows: Specifically, the Markov decision process is constructed by constraints from the state space, action space, reward function, and discount factor. The state space vector contains four dimensions of physical and predictive features: the current scalar value of the dynamic carbon emissions, the current scalar value of the gas tank buffer pressure, the current scalar value of the instantaneous carbon emission factor of the power grid, and a prediction vector composed of the future hourly energy consumption prediction trend dataset. This state space vector performs periodic numerical acquisition and register overwrite updates within each set system clock step, thereby establishing the data basis for the subsequent generation of scheduling instructions.

[0063] In terms of network structure construction and training for the reinforcement learning prediction model, this embodiment uses a Deep Q-Network algorithm model (DQN) as the reinforcement learning prediction model. The Deep Q-Network algorithm model includes an evaluation network and a target network with identical structures. As a specific parameter configuration embodiment, both network topologies sequentially include a four-dimensional feature input layer, three fully connected hidden layers configured with 128, 64, and 32 neurons respectively and incorporating a linear rectified activation function, and an output layer with a two-dimensional output channel. During the model training phase, the system constructs an experience replay pool with a maximum storage capacity of 10,000 transition records to store the state space vector collected by the system within historical system clock steps, executed action instructions, scalar feedback values ​​calculated according to the reward function mechanism, and the state space vector at the next time step. In each training cycle, a batch size of 64 data samples is randomly drawn from the experience replay pool. The evaluation network calculates the value of the current action based on the current state space vector, and the target network calculates the value of the target action based on the state space vector of the next time step. The system uses the mean squared error function to calculate the temporal difference error between the current action value and the target action value as the loss function. The future reward discount factor of the Markov decision process is set to 0.95, and the initial learning rate for updating the evaluation network parameters is set to 0.0005. The adaptive moment estimation optimization algorithm is used to calculate the gradient based on the loss function and backpropagate to update the weight parameters of the evaluation network. At the same time, the system hard synchronizes the weight parameters of the evaluation network to the target network every 100 training steps until the loss function converges, thus obtaining the trained reinforcement learning prediction model.

[0064] In the output architecture configuration of the reinforcement learning prediction model, the system directly maps the digital algorithm output to the physical action execution layer of the underlying mechanical equipment. This configuration establishes a discrete action space composed of loading and unloading commands issued from the edge computing gateway to the air compressor controller. In the real-time control inference stage, the trained reinforcement learning prediction model, based on the state space vector input at the current time step, performs parameter matrix multiplication and addition operations and nonlinear activation mapping in the internal evaluation network to obtain the numerical distribution of each discrete command action at the output layer, i.e., the action value distribution vector. Subsequently, the system performs extreme value retrieval optimization logic in the numerical distribution to extract the parameter term with the largest algebraic weight, and uses the command hard-mapped by the largest parameter term as the execution action of the current cycle.

[0065] For example, if the output action value values ​​for the two instructions of loading and unloading are 0.85 and 0.15 respectively, since the value for the loading action is the largest, the system determines that the execution action of the current cycle is loading.

[0066] The execution action is sent directly to the air compressor controller via the physical communication port of the edge computing gateway: when the air compressor controller receives a loading command, it triggers the closure of the corresponding internal electrical control circuit to drive the mechanical components of the air compressor intake valve to open; when it receives an unloading command, it triggers the disconnection of the electrical control circuit to drive the mechanical components of the intake valve to close; thus, a deterministic unidirectional signal transmission link is constructed from the calculation results of the digital model to the physical mechanical displacement.

[0067] Simultaneously, the system is configured with a reward function mechanism based on multivariate combination rule calculation, used to calculate the scalar feedback value required for the aforementioned experience replay pool during the training phase. This mechanism performs algebraic operations by reading and merging the current state space dataset, outputting a comprehensive reward scalar for algorithm internal parameter iteration. During specific rule execution, the system hardware unit continuously compares the current buffer pressure value of the gas tank with the upper and lower limits of the safe physical boundary set in memory. Simultaneously, the system determines whether the instantaneous carbon emission factor of the power grid acquired at the current moment is at a low point within the predetermined monitoring period. When the above data matching conditions are met and the instruction output port issues a loading action, the reward function generates a corresponding positive scalar value. Specifically, when the dynamic carbon emission amount contained in the state space vector experiences a significant numerical expansion due to equipment mechanical performance degradation and the nonlinear multiplication effect of the penalty amplification coefficient, the reward function... The negative algebraic penalty term in the numerical mechanism is triggered and generates a rapidly decaying negative scalar feedback value. This negative scalar feedback value, in the temporal difference update logic of the reinforcement learning prediction model, forces the target action value corresponding to the loading command to drop sharply. As a result, when the model faces the same state input in the future, the output distribution of the corresponding unloading command quickly takes the extreme value as the dominant value. The forced intervention trigger logic when the physical equipment deteriorates is accurately reproduced in the online decision matrix. In addition, when the future hourly energy consumption prediction trend contained in the state space vector indicates that a high value carbon emission condition is about to appear, the system changes the combination weight of relevant variables to adjust the final output result of the scalar feedback value. The reinforcement learning prediction model relies on the above reward function mechanism based on objective data judgment to perform parameter iteration and regular calculation, and continuously outputs a sequence of equipment state adjustment commands with system clock timestamps for the underlying air compressor controller to execute sequentially.

[0068] This invention also discloses an AI-enhanced dynamic optimization system for electric carbon, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the AI-enhanced dynamic optimization method for electric carbon according to this invention.

[0069] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. An AI enhanced electric-carbon dynamic optimization method, comprising obtaining an instantaneous active power under an operating state of an air compressor, and obtaining an instantaneous carbon emission factor of a power grid, characterized in that, The method further includes: Obtain the intake temperature and intake pressure at the intake section of the air compressor; obtain the exhaust pressure and exhaust temperature at the exhaust section of the air compressor; The ideal isentropic compression ratio work is calculated by combining the intake temperature, the intake pressure and the exhaust pressure. The actual compression ratio work is calculated by combining the intake air temperature and the exhaust air temperature; The isentropic efficiency is determined by calculating the ratio of the ideal isentropic compression ratio work to the actual compression ratio work. The penalty factor is calculated using the isentropic efficiency and the penalty amplification factor. The dynamic carbon emissions are obtained by multiplying the instantaneous active power, the instantaneous carbon emission factor of the power grid, and the penalty factor, and then integrating them over a set time window. The dynamic carbon emissions are input into a locally deployed reinforcement learning prediction model to output execution actions, including: extracting the equipment production scheduling instruction sequence and the historical energy consumption feature dataset cached in local memory; aligning and concatenating the equipment production scheduling instruction sequence and the historical energy consumption feature dataset according to a unified system clock step to construct a time-series feature matrix; inputting the time-series feature matrix into a long short-term memory network model for gating mechanism tensor operations to output a future hourly energy consumption prediction trend dataset; obtaining the current buffer pressure value of the gas tank; and combining the dynamic carbon emissions, the current buffer pressure value of the gas tank, the instantaneous carbon emission factor of the power grid, and the future hourly energy consumption... The predicted trend dataset is concatenated into structured tensors to generate a state space vector of a Markov decision process with discrete time steps. This state space vector is then input into a reinforcement learning prediction model constructed using a deep Q-network algorithm. Through parameter matrix multiplication and addition operations and nonlinear activation mapping in the deep Q-network algorithm, the action value distribution vector corresponding to the discrete command action is obtained. An extreme value retrieval and optimization logic is executed in the action value distribution vector to extract the parameter term with the largest algebraic weight. The command hard-mapped to the parameter term with the largest algebraic weight is set as the execution action for the current cycle and sent to the air compressor controller.

2. The Al-augmented electric carbon dynamic optimization method of claim 1, wherein, The method further includes: deploying an edge computing gateway to establish a bidirectional data transmission channel with the air compressor controller; concurrently reading the underlying operating status register through a preset fixed sampling frequency; performing analog-to-digital conversion and data format alignment on discrete analog signals on various mechanical components to generate a continuous time series dataset; and adding clock synchronization timestamps to the intake temperature, intake pressure, exhaust pressure, exhaust temperature, and instantaneous active power.

3. The Al-augmented electric carbon dynamic optimization method of claim 1, wherein, The ideal isentropic compression ratio work is calculated by combining the intake temperature, the intake pressure, and the exhaust pressure, including: ； wherein is the ideal isentropic compression ratio work; is the isentropic exponent; is the gas constant; is the intake air temperature; is the intake air pressure; is the exhaust air pressure.

4. The Al-augmented electro-carbon dynamic optimization method of claim 1, wherein, The actual compression specific work is calculated by combining the intake temperature and the exhaust temperature, including: calculating the difference between the exhaust temperature and the intake temperature to obtain the intake and exhaust temperature rise; and calculating the product of the intake and exhaust temperature rise and the specific heat capacity at constant pressure to obtain the actual compression specific work.

5. The AI-enhanced dynamic optimization method for electrocarbon as described in claim 1, characterized in that, The penalty factor is calculated using the isentropic efficiency and the penalty amplification factor. , To penalize the amplification factor, It is isentropic efficiency.

6. The AI-enhanced dynamic optimization method for electrocarbon as described in claim 1, characterized in that, The dynamic carbon emissions are obtained by multiplying the instantaneous active power, the instantaneous carbon emission factor of the power grid, and the penalty factor, and then integrating the results over a set time window. This includes: ； In the formula, For a moment Dynamic carbon emissions; For periodic time intervals; For a moment The instantaneous active power; For a moment The instantaneous carbon emission factor of the power grid; For a moment The penalty factor.

7. The AI-enhanced dynamic optimization method for electrocarbon as described in claim 1, characterized in that, The offline calibration method for the penalty amplification factor is as follows: Extract the tensor mapping span ratio required for the input layer of the reinforcement learning prediction model; set the severe working condition penalty factor value equal to the product of the baseline working condition penalty factor value and the tensor mapping span ratio; Based on the numerical correlation of the penalty factor under severe working conditions, a linear algebraic equation in one variable is constructed; the unique real solution of the penalty amplification coefficient is calculated by solving the algebraic equation.

8. The AI-enhanced dynamic optimization method for electrocarbon as described in claim 7, characterized in that, The methods for obtaining the severe working condition penalty factor value and the baseline working condition penalty factor are as follows: Under the benchmark operating conditions of the new model, objective physical parameters are collected and the first isentropic efficiency benchmark value is calculated; the initial efficiency decay physical quantity is obtained by subtracting the first isentropic efficiency benchmark value from 1. The baseline operating condition penalty factor is obtained by adding the initial efficiency decay physical quantity to 1; the objective physical parameters are collected and the second isentropic efficiency threshold is calculated under deteriorating operating conditions; the deteriorating efficiency decay physical quantity is obtained by subtracting the second isentropic efficiency threshold from 1; and the severe operating condition penalty factor is obtained by adding the deteriorating efficiency decay physical quantity to 1.

9. An AI-enhanced dynamic optimization system for carbon dioxide, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the AI-enhanced dynamic optimization method for electric carbon according to any one of claims 1-8.

Images