A method for grading evaluation of regulation flexibility for high energy-consuming industrial user load

Abstract

A method for hierarchical evaluation of load regulation flexibility of high energy-consuming industrial users, including four steps of multi-dimensional basic data collection, multi-time scale load regulation coordination matrix construction, multi-time scale coordination optimization based on decision potential field and damping mechanism, and hierarchical evaluation of regulation flexibility, a virtual hierarchical damping mechanism is embedded in the reward function of the artificial intelligence model, and differential damping punishment is applied to the strategy changes of different time scales such as monthly, day-ahead and real-time, and the longer the time scale, the stronger the damping, so as to inhibit unnecessary strategy jump and chain oscillation across time scales. In addition, by combining global guidance of potential field with local stability of damper, the method can output smooth, stable and executable coordinated regulation strategy under the constraint of multiple strong coupling of electricity, carbon and equipment. At the same time, it also establishes a multi-dimensional dynamic evaluation framework to complete the hierarchical evaluation and optimization decision of the load regulation flexibility of high energy-consuming industrial users by means of rolling mechanism.

Description

Technical Field

[0001] This invention relates to the field of power grid control technology, and in particular to a method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads. Background Technology

[0002] The integration of a high proportion of fluctuating renewable energy sources poses unprecedented challenges to the real-time balancing and flexible adjustment capabilities of the power grid. Simultaneously, the formal operation of the carbon emissions trading market has made carbon emission costs one of the core economic factors affecting enterprise production and operation, especially for energy-intensive industrial users. Against this backdrop, deeply exploring the regulation potential of high-energy-consuming industrial loads and guiding them from traditional rigid electricity consumption patterns to adjustable and interruptible flexible resources is of great significance for improving the grid's ability to absorb renewable energy, ensuring the safe and stable operation of the system, and promoting energy conservation and carbon reduction in the industrial sector.

[0003] Currently, the industry has conducted extensive research and practice on the participation of industrial loads in system regulation, resulting in several technical approaches. Existing technologies mainly focus on using price signals or dispatch instructions to incentivize loads to change their electricity consumption behavior. A search of Chinese patent publication number CN114759597B reveals a dispatch method for high-energy-consuming loads participating in peak shaving based on variational mode decomposition. This method decomposes the system's equivalent load curve into components of different frequencies and assigns high-energy-consuming loads and thermal power units to cope with fluctuations at different frequencies, aiming to improve wind power absorption capacity. Chinese patent publication number CN118199074A proposes a pre-dispatch-re-dispatch ultra-carbon demand response method based on source-load coordinated carbon reduction. By introducing an "ultra-carbon electricity price" mechanism, it drives users to shift to low-carbon behaviors in both the pre-dispatch and re-dispatch stages, achieving a synergy between system economy and low-carbon performance.

[0004] However, existing technologies still have significant limitations and problems in practical applications, mainly due to a lack of profound understanding and effective solutions to the inherent conflicts and decision-making oscillations in the cross-timescale load regulation behavior of high-energy-consuming industries. High-energy-consuming production processes are continuous and tightly coupled, and their load regulation involves multiple timescales, including real-time control, intraday planning, day-ahead scheduling, and even monthly carbon quota planning. Existing methods typically isolate scheduling problems at different timescales or use simple sequential recursive solutions. For example, after the day-ahead plan is determined, deviation corrections during intraday or real-time phases may lead to significant deviations from the day-ahead economic or low-carbon plan in pursuit of short-term optimality, and vice versa. This strong coupling of multi-objective, multi-constraint, and multi-timescale decision-making loops easily causes optimization algorithms to fall into a state of repeated oscillations and difficulty in convergence—a state of "strategy oscillation." Summary of the Invention

[0005] The purpose of this invention is to propose a hierarchical evaluation method for the control flexibility of loads of high-energy-consuming industrial users. This method can solve the problems existing in the prior art and is a simple and easy-to-implement hierarchical evaluation method. It achieves smooth and coordinated control under multiple constraints of electricity, carbon, and equipment by constructing a decision potential energy field and embedding a virtual hierarchical damping mechanism. At the same time, it establishes a multi-dimensional dynamic evaluation framework and uses a rolling mechanism to complete the hierarchical evaluation and optimization decision of the control flexibility of loads of high-energy-consuming industrial users.

[0006] The technical solution of this invention: a method for graded evaluation of control flexibility for high-energy-consuming industrial user loads, characterized by comprising the following steps: S1. Collect multi-dimensional basic data of target users and construct a structured multi-dimensional basic dataset; The multi-dimensional basic data in step S1 includes at least monthly carbon emission quota data, electricity price fluctuation data at different time scales, and adjustable operating parameters and process constraints data of various high-energy-consuming load equipment.

[0007] Step S1 specifically includes: receiving electricity price data from the power trading platform and the power grid dispatching agency, receiving monthly / annual carbon emission quota data from the carbon quota allocation management department, as well as carbon quota spot and futures price data from the carbon trading market, historical carbon emission data from users, and load equipment data from the user-side energy management system; aligning the collected data from multiple sources and heterogeneous data, filling in missing values ​​and removing outliers, and outputting a structured multi-dimensional basic dataset.

[0008] The structured multi-dimensional basic dataset obtained in step S1 is a data table with timestamp as the primary key, containing the following dimensions: [timestamp, device ID, real-time electricity price, day-ahead electricity price forecast, carbon emission quota balance, carbon price, real-time device power, upper limit of adjustable device power, lower limit of adjustable device power, associated process constraint identifier]. It is generated by aligning multi-source heterogeneous data, filling missing values ​​and removing outliers, and then associating and regularizing the multi-source data according to a unified timestamp and device ID.

[0009] S2. Construction of Multi-Time-Scale Load Regulation Coordination Matrix: Based on the multi-dimensional basic dataset obtained in step S1, construct a multi-time-scale load regulation coordination matrix that characterizes the response characteristics and constraints of user loads participating in power grid regulation at different time scales, and output the matrix. The dimensions of the multi-timescale load regulation coordination matrix in step S2 include at least five time scales: monthly, weekly, daily, intraday, and real-time. The matrix elements include at least the adjustability potential, adjustment cost, and carbon emission intensity of a specific load device at each time scale.

[0010] The S2 step specifically includes: S21. Decompose the monthly carbon emission allowance into nested carbon emission budget constraint boundary sets for weekly, daily, and intraday periods. S22. For each load device, combining electricity price fluctuation data and its adjustable operating parameters, solve for its optimal economic adjustment range at each time scale. And fit the adjustment cost function ; Step S22 specifically includes: S221. Interval Solution: With the objective of minimizing electricity costs, and constrained by the upper and lower limits of the equipment's adjustable power and the allowable fluctuation ranges of temperature and pressure, linear programming (LP) or mixed-integer linear programming (MILP) models are established and solved for the corresponding electricity price sequences at monthly, day-ahead, and real-time time scales. This yields the economically optimal value for equipment power adjustment and the allowable deviation range at each scale, i.e., the optimal economic adjustment interval. , where t is the time point; S222, Function Fitting: Collect cost change data for the equipment under different power adjustment amounts ΔP. adjust = ΔP * Electricity price for the corresponding time period, using a quadratic function on (ΔP, Cost) adjust The data points are fitted with least squares to obtain an approximate adjustment cost function for the device, which is then used to predict the adjustment cost at other times. The coefficients a, b, and c are determined through fitting, and t is a time point. For different time scales, the same quadratic function form as shown in the above formula is used, but the coefficients are updated according to the typical electricity price level of that scale.

[0011] In step S222, the approximate adjustment cost function retains its form for different time scales (monthly, day-ahead, and real-time), but the coefficients a, b, and c need to be determined separately. The method for determining these coefficients is as follows: each time scale has its corresponding typical electricity price curve or sequence; the typical electricity price data for a specific time scale is used as input, along with the power adjustment amount of the equipment. Together, cost data sets at this scale are generated. Then, the least squares method is used to refit the coefficients a, b, c at that time scale. Therefore, the coefficients are determined by fitting, and the electricity price level at different time scales drives independent fitting to obtain different sets of coefficients.

[0012] S23. Combine the carbon emission budget constraint boundary set obtained in step S21 with the optimal economic adjustable range for each load device obtained in step S22. Adjustment cost function and physical limit parameters The data is organized by time scale as rows and load equipment as columns. Each matrix element is integrated into a data tuple containing feasible domain boundaries and cost coefficients, and the multi-time scale load regulation coordination matrix is ​​output.

[0013] The multi-timescale load regulation coordination matrix in step S23 is a two-dimensional structure, i.e., a multi-timescale load regulation coordination matrix. That is, all matrix elements Collections organized by rows and columns:

[0014] Where the row index is the time scale. S is the set of decision-making levels across all time scales, i.e.: The column index is for load devices. I is the set of all target high-energy-consuming industrial user-side load devices participating in power grid regulation; matrix elements It is a structured data tuple representing time. Time scale Lower equipment The feasible domain and cost characteristics of regulation, namely: in, It is the carbon emission budget boundary, obtained from step S21, representing the time scale. At any moment The remaining carbon emissions budget cap; The economic adjustment range, obtained from step S22, represents the equipment's... At any moment The economically optimal power adjustment range that meets process constraints; These are the adjustment cost function coefficients, which are the fitting results from step S22, i.e., if the cost function is in quadratic form. Then the coefficient set is stored here. ; It is the lower limit of absolute power adjustment; It is the upper limit of absolute power adjustment; This refers to the maximum power ramp rate; the absolute power adjustable lower limit. Adjustable upper limit of absolute power Maximum power gradeability These are all physical limit parameters, which are the equipment's... The inherent constraints.

[0015] S3. Multi-timescale collaborative optimization based on decision potential energy field and damping mechanism: The multi-timescale load regulation collaborative matrix constructed in step S2 and the external power grid multi-timescale regulation demand command are used as inputs. By introducing decision potential energy field and virtual hierarchical damping mechanism, the collaborative optimization calculation is performed using artificial intelligence model. A collaborative regulation strategy that can avoid strategy oscillation and realize cross-timescale load resource collaborative optimization is output, and the system-level multi-timescale complementary regulation potential is quantified. In step S3, the decision potential field is determined by the carbon potential field U. c With economic potential field U e It is formed by weighted fusion, that is: U total= αU c +βU e Where α and β are preset weighting coefficients.

[0016] The carbon potential energy field U c Defined as the following function: Among them, C actual (t) represents the actual cumulative carbon emissions up to time t, C budget (t) represents the carbon emission budget for the corresponding time scale, k c The carbon potential coefficient is defined as follows: The economic potential field function is defined as: Cost actual (t) represents the actual total electricity cost up to time t, Cost reference (t) represents the reference optimal cost curve based on long-term electricity price forecasts, and k e This is the economic potential coefficient.

[0017] The virtual hierarchical damping mechanism in step S3 refers to the mechanism implemented through the reward function R of the artificial intelligence model. total Add a damping penalty term R to (t) damp Implementation, that is: Where λ is the damping strength coefficient.

[0018] The damping penalty term R damp The calculation formula is: Where S is the set of all time-scale decision-making layers. Its specific composition is determined based on the actual needs of power grid dispatching and user production plans. πs(t) is the strategy vector of time scale s in decision period τ, and ω s Let be the stratified damping coefficient corresponding to this time scale, and satisfy . .

[0019] In step S3, the artificial intelligence model uses a quasi-steady-state rolling coordination mechanism to perform optimization. At the beginning of each short-timescale decision cycle, this mechanism freezes the decision variables of a longer timescale to the optimal values ​​obtained in the previous round of optimization and solves the current short-cycle optimization sub-problem. After the short-cycle decision is executed, correction is triggered at a fixed period or when the deviation between the actual performance and the plan exceeds a threshold, and the deviation is fed back to the subsequent long-cycle decision plan.

[0020] The S3-level correction triggered at a fixed period or when the deviation between actual performance and plan exceeds a threshold refers to the quantitative adjustment of decision variables at longer time scales (daytime, real-time, etc.) in the quasi-steady-state rolling coordination mechanism. This adjustment is applied to the previously frozen optimal decision results at longer time scales. Its core function is to feed back short-cycle execution deviations to long-cycle decision plans within a limited range, achieving coordinated correction of decisions at different time scales. The correction amount is calculated using the following formula: in, To correct the gain coefficient, The deviation between actual performance and plan, This represents the maximum allowable adjustment range. Finally, through the above optimization calculations, the following two specific results are output: ① Coordinated regulation strategy: The strategy is a multi-dimensional instruction sequence A(t'), that is, for each future decision time t' and each load device i, a specific power adjustment instruction value ΔP is output. i (t'), then: Where each element ΔP i (t') represents the specific power adjustment value for load device i at decision time t'. The multidimensional command sequence A(t') satisfies the electrical, carbon, and equipment constraints across all time scales, and is smooth and stable across time scales under the guidance of the potential energy field and the damping mechanism; ② Quantitative results of multi-time scale complementary regulation potential, including: Monthly potential: = ∑(Monthly adjustable capacity of all devices); Current potential: = ∑(Capacity of all devices capable of rapid response at the day-ahead scale); Real-time potential: = ∑(the capacity of all devices that can be adjusted instantaneously at a real-time scale); Complementarity coefficient: This is used to measure the potential gain from cross-scale collaboration. The monthly potential V is... month Potential V day Real-time potential V real Both the complementarity coefficient α and the maximum adjustable capacity extracted by analyzing the synergistic control strategy at various time scales are calculated.

[0021] The artificial intelligence model in step S3 specifically refers to the use of the Actor-Critic framework; the training process is carried out in a simulation environment that includes electricity market fluctuations, carbon price changes and stochastic grid demand commands, and the training objective is to minimize the comprehensive loss function that includes the original reward, the potential energy field guidance term and the damping penalty term.

[0022] The output of the Actor network in the Actor-Critic framework is the action. The input is the system state, that is: ; The Critic network in the Actor-Critic framework evaluates state value. The training objective is set to minimize the overall loss function, i.e.: in, and For the standard policy gradient and value loss, For the decision potential field, For damping penalty term, , , The corresponding weights are used; this design enables the model to simultaneously optimize economic efficiency, carbon emission targets, and suppress policy oscillations during learning.

[0023] The input to the artificial intelligence model is S t The output of the Actor network is a multidimensional instruction sequence A(t'), the Critic network is responsible for evaluating the state value, and the comprehensive loss function includes the original reward, the potential energy field guidance term, and the damping penalty term. These are embedded into the core of the AI ​​model to jointly construct the artificial intelligence model.

[0024] S4. Stratification and Evaluation of Control Flexibility: Based on the collaborative control strategy obtained in step S3 and the quantitative results of the multi-timescale complementary control potential, the control flexibility of all user loads is stratified and graded according to the control timing, response speed, control capacity, and unit control carbon efficiency contribution of the load equipment, and a standardized flexibility grading evaluation report is generated to improve the control flexibility of high-energy-consuming industrial user loads.

[0025] Step S4 specifically includes: S41: Based on the multi-dimensional instruction sequence output from step S3, i.e., the coordinated control strategy vector. And grid carbon emission intensity data, to calculate the load equipment Unit adjustment carbon efficiency contribution value during the assessment period: in, For equipment Total carbon emissions to be avoided The total amount of electrical energy it regulates. Let be the marginal carbon emission factor of the power grid at time t; S42: Tiered and segmented distribution of all user load: Level 1 flexibility resources: If equipment The adjustment actions mainly occur in real-time or intraday time windows, and their average response delay is... The extent of the impact on key process parameters At the same time satisfy If so, it is rated as a Level 1 flexibility resource; Secondary flexibility resources: if equipment The adjustment actions are mainly arranged on a daily or weekly scale and need to be made at least in advance. The notification timeline requires coordinated adjustments to other equipment within the same process chain to simultaneously meet the following requirements. If so, it is rated as a Level 2 flexibility resource; Level 3 flexibility resources: if equipment The main basis for switching to long-term operation mode is based on monthly quota and meets the requirements. It is then assessed as a Level 3 flexibility resource, among which This is the average response delay threshold, in seconds or minutes. For the maximum allowable deviation of key process parameters such as temperature and pressure; The high carbon efficiency threshold is measured in tons of carbon dioxide per megawatt-hour; , These represent the lower and upper limits of the medium carbon efficiency contribution range, respectively. This is the threshold value for low carbon efficiency; The advance notification time threshold is in hours or days.

[0026] Advantages of the present invention: (1) This hierarchical evaluation method for the regulation flexibility of loads of high-energy-consuming industrial users maps the complex decision space of multiple objectives and constraints into a unified "decision potential energy field". In this field, states such as carbon emission exceeding limits and cost deviation are quantified as "high potential energy", guiding the optimization process to automatically slide towards the equilibrium point of "low potential energy". Secondly, a "virtual hierarchical damping mechanism" is embedded in the reward function of the artificial intelligence model. Different damping penalties are applied to policy changes at different time scales: monthly, daily, and real-time. The longer the time scale, the stronger the damping, thereby effectively suppressing unnecessary policy jumps and chain oscillations across time scales. Finally, by combining the global guidance of the potential energy field with the local stability of the damper, the optimization algorithm can output a smooth, stable, and executable coordinated regulation strategy under multiple strong coupling constraints of electricity, carbon, and equipment.

[0027] (2) This method for classifying and evaluating the load regulation flexibility of high-energy-consuming industrial users constructs a multi-dimensional dynamic evaluation framework. This framework not only quantifies the physical regulation capacity of the load, but also, more importantly, integrates its economic regulation cost and the unique "unit regulation carbon efficiency contribution" index. Based on this, according to multiple dimensions such as the response time scale, the degree of impact on the production process, and the unit regulation carbon efficiency contribution value of the load in the optimization strategy, it dynamically divides the load into "Level 1 (fast / efficient)," "Level 2 (planned / coordinated)," and "Level 3 (basic / reconfigured)" flexibility resources with clear technical and economic characteristics, and generates a standardized classification evaluation report.

[0028] (3) This method for classifying and evaluating the control flexibility of loads from high-energy-consuming industrial users designs a rolling operation mode of "freeze-solve-feedback". During online decision-making on a short time scale, the optimal decision result on a longer time scale is temporarily "frozen" as fixed background parameters, thereby significantly reducing the complexity and real-time requirements of online solution. After the short-time scale decision is executed, the deviation between the actual execution effect and the expected result is accumulated and "gently" fed back to the longer-term decision for correction in a low-frequency and limited-amplitude manner. This process innovatively establishes a buffer and decoupling mechanism between time scales. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall system architecture of a hierarchical evaluation method for the control flexibility of high-energy-consuming industrial user loads, as described in this invention.

[0030] Figure 2 This is a schematic diagram of the core steps of a multi-timescale collaborative optimization method for a graded evaluation method of control flexibility for high-energy-consuming industrial user loads, as described in this invention.

[0031] Figure 3This is a schematic diagram of the complete data processing chain from data to service for a hierarchical evaluation method for the control flexibility of loads of high-energy-consuming industrial users, as described in this invention. Detailed Implementation

[0032] Example: A method for graded evaluation of control flexibility for high-energy-consuming industrial user loads, characterized by the following steps: S1. Collect multi-dimensional basic data from the target user, including monthly carbon emission quota data, electricity price fluctuation data at different time scales, and adjustable operating parameters and process constraints data of various high-energy-consuming load equipment. Construct a structured multi-dimensional basic dataset based on this data. This dataset is a data table with timestamp as the primary key and containing the following dimensions: [Timestamp, Equipment ID, Real-time Electricity Price, Day-ahead Electricity Price Forecast, Carbon Emission Quota Balance, Carbon Price, Real-time Equipment Power, Upper Limit of Adjustable Equipment Power, Lower Limit of Adjustable Equipment Power, Related Process Constraint Identifier]. It is generated by aligning multi-source heterogeneous data, filling missing values, and removing outliers, and then associating and standardizing the multi-source data according to a unified timestamp and equipment ID.

[0033] The target user's multi-dimensional basic data specifically includes: receiving electricity price data from power trading platforms and grid dispatching agencies, receiving monthly / annual carbon emission quota data from carbon quota allocation management departments, as well as carbon quota spot and futures price data from the carbon trading market, user historical carbon emission data, and load equipment data from the user-side energy management system. The collected data is aligned with multi-source heterogeneous data, missing values ​​are filled, and outliers are removed to output a structured multi-dimensional basic dataset.

[0034] S2. Based on the multi-dimensional basic dataset obtained in step S1, construct a multi-time-scale load regulation coordination matrix that characterizes the response characteristics and constraints of user loads participating in power grid regulation at different time scales, and output the matrix. The matrix dimensions include at least five time scales: monthly, weekly, daily, intraday, and real-time. The matrix elements include at least the adjustability potential, adjustment cost, and carbon emission intensity of specific load equipment at each time scale.

[0035] Specifically, it includes: S21. Decompose the monthly carbon emission allowance into nested carbon emission budget constraint boundary sets for weekly, daily, and intraday periods. S22. For each load device, combining electricity price fluctuation data and its adjustable operating parameters, solve for its optimal economic adjustment range at each time scale. And fit the adjustment cost function ; S221. Interval Solution: With the objective of minimizing electricity costs, and constrained by the upper and lower limits of the equipment's adjustable power and the allowable fluctuation ranges of temperature and pressure, linear programming (LP) or mixed-integer linear programming (MILP) models are established and solved for the corresponding electricity price sequences at monthly, day-ahead, and real-time time scales. This yields the economically optimal value for equipment power adjustment and the allowable deviation range at each scale, i.e., the optimal economic adjustment interval. , where t is the time point; S222, Function Fitting: Collect cost change data for the equipment under different power adjustment amounts ΔP. adjust = ΔP * Electricity price for the corresponding time period, using a quadratic function on (ΔP, Cost) adjust The data points are fitted with least squares to obtain an approximate adjustment cost function for the device, which is then used to predict the adjustment cost at other times. Wherein, coefficients a, b, and c are determined through fitting, and t is a time point. For different time scales, the same quadratic function form as shown in the above formula is used, but the coefficients are updated according to the typical electricity price level of that scale. That is, each time scale has its corresponding typical electricity price curve or sequence; the typical electricity price data of a specific time scale is used as input, along with the power adjustment amount of the equipment. Together, cost data sets at this scale are generated. Then, the least squares method is used to refit the coefficients a, b, c at that time scale. Therefore, the coefficients are determined by fitting, and the electricity price level at different time scales drives independent fitting to obtain different sets of coefficients.

[0036] S23. Combine the carbon emission budget constraint boundary set obtained in step S21 with the optimal economic adjustable range for each load device obtained in step S22. Adjustment cost function and physical limit parameters The data is organized by time scale (rows) and load equipment (columns). Each matrix element is integrated into a data tuple containing the feasible domain boundary and cost coefficient. The resulting multi-time scale load control coordination matrix is ​​a two-dimensional structure. That is, all matrix elements Collections organized by rows and columns: Where the row index is the time scale. S is the set of decision-making levels across all time scales, i.e.: The column index is for load devices. I is the set of all target high-energy-consuming industrial user-side load devices participating in power grid regulation; matrix elements It is a structured data tuple representing time. Time scale Lower equipment The feasible domain and cost characteristics of regulation, namely: in, It is the carbon emission budget boundary, obtained from step S21, representing the time scale. At any moment The remaining carbon emissions budget cap; The economic adjustment range, obtained from step S22, represents the equipment's... At any moment The economically optimal power adjustment range that meets process constraints; These are the adjustment cost function coefficients, which are the fitting results from step S22, i.e., if the cost function is in quadratic form. Then the coefficient set is stored here. ; It is the lower limit of absolute power adjustment; It is the upper limit of absolute power adjustment; This refers to the maximum power ramp rate; the absolute power adjustable lower limit. Adjustable upper limit of absolute power Maximum power gradeability These are all physical limit parameters, which are the equipment's... The inherent constraints.

[0037] S3. Multi-timescale collaborative optimization based on decision potential energy field and damping mechanism: The multi-timescale load regulation collaborative matrix constructed in step S2 and the external power grid multi-timescale regulation demand command are used as inputs. By introducing decision potential energy field and virtual hierarchical damping mechanism, the collaborative optimization calculation is performed using artificial intelligence model. A collaborative regulation strategy that can avoid strategy oscillation and realize cross-timescale load resource collaborative optimization is output, and the system-level multi-timescale complementary regulation potential is quantified. Among them, the decision potential field is composed of the carbon potential field U c With economic potential field U e It is formed by weighted fusion, that is: U total= αU c +βU e Where α and β are preset weighting coefficients, C actual (t) represents the actual cumulative carbon emissions up to time t, C budget (t) represents the carbon emission budget for the corresponding time scale, k c Cost is the carbon potential energy coefficient. actual (t) represents the actual total electricity cost up to time t, Cost reference (t) represents the reference optimal cost curve based on long-term electricity price forecasts, and k e This is the economic potential coefficient.

[0038] Virtual hierarchical damping mechanism refers to the mechanism that uses the reward function R of an artificial intelligence model to dampen the reward function R. total Add a damping penalty term R to (t) damp Implementation, that is: Where λ is the damping strength coefficient, and S is the set of decision-making layers across all time scales. Its specific composition is determined based on the actual needs of power grid dispatching and user production plans. πs(t) is the strategy vector of time scale s in decision period τ, and ω s Let be the stratified damping coefficient corresponding to this time scale, and satisfy . .

[0039] The artificial intelligence model employs a quasi-steady-state rolling coordination mechanism for optimization. At the start of each short-term decision cycle, this mechanism freezes the decision variables of a longer timescale to the optimal values ​​obtained in the previous optimization round and solves the current short-term optimization sub-problem. After the short-term decision is executed, correction is triggered at fixed intervals or when the deviation between actual performance and the plan exceeds a threshold, feeding the deviation back to subsequent long-term decision plans. Triggering correction when the deviation exceeds the threshold means that in the quasi-steady-state rolling coordination mechanism, when a deviation occurs between actual performance and the plan after the execution of a short-term (intraday, real-time) decision, a quantitative adjustment can be made to the decision variables of subsequent longer timescales (daily, monthly, etc.). The adjustment target is the previously frozen optimal decision result of the longer timescale. The core function is to feed back the short-term execution deviation to the long-term decision plan with a limited magnitude, achieving coordinated correction of decisions at different timescales. The formula for calculating the correction amount is: in, To correct the gain coefficient, The deviation between actual performance and plan, This represents the maximum allowable adjustment range. Finally, through the above optimization calculations, the following two specific results are output: ① Coordinated regulation strategy: The strategy is a multi-dimensional instruction sequence A(t'), that is, for each future decision time t' and each load device i, a specific power adjustment instruction value ΔP is output. i (t'), then: Where each element ΔP i (t') represents the specific power adjustment value for load device i at decision time t'. The multidimensional command sequence A(t') satisfies the electrical, carbon, and equipment constraints across all time scales, and is smooth and stable across time scales under the guidance of the potential energy field and the damping mechanism; ② Quantitative results of multi-time scale complementary regulation potential, including: Monthly potential: = ∑(Monthly adjustable capacity of all devices); Current potential: = ∑(Capacity of all devices capable of rapid response at the day-ahead scale); Real-time potential: = ∑(the capacity of all devices that can be adjusted instantaneously at a real-time scale); Complementarity coefficient: This is used to measure the potential gain from cross-scale collaboration. The monthly potential V is... month Potential V day Real-time potential V real Both the complementarity coefficient α and the maximum adjustable capacity extracted by analyzing the synergistic control strategy at various time scales are calculated.

[0040] The artificial intelligence model adopts the Actor-Critic framework, and its input is S. t The Actor network outputs a multidimensional instruction sequence A(t'), the Critic network evaluates the state value, and the comprehensive loss function includes the original reward, the potential energy field guidance term, and the damping penalty term. These are embedded into the core of the AI ​​model to jointly construct the artificial intelligence model. The training process is conducted in a simulated environment including electricity market fluctuations, carbon price changes, and stochastic grid demand instructions. The training objective is to minimize the comprehensive loss function, which includes the original reward, the potential energy field guidance term, and the damping penalty term. The output of the Actor network in the Actor-Critic framework is the action. The input is the system state, that is: ; Critic network evaluates state value The training objective is set to minimize the overall loss function, i.e.: in, and For the standard policy gradient and value loss, For the decision potential field, For damping penalty term, , , The corresponding weights are used; this design enables the model to simultaneously optimize economic efficiency, carbon emission targets, and suppress policy oscillations during learning.

[0041] S4. Stratification and Evaluation of Control Flexibility: Based on the collaborative control strategy obtained in step S3 and the quantitative results of the multi-timescale complementary control potential, the control flexibility of all user loads is stratified and graded according to the control timing, response speed, control capacity, and unit control carbon efficiency contribution of the load equipment, and a standardized flexibility grading evaluation report is generated to improve the control flexibility of high-energy-consuming industrial user loads.

[0042] S41: Based on the multi-dimensional instruction sequence output from step S3, i.e., the coordinated control strategy vector. And grid carbon emission intensity data, to calculate the load equipment Unit adjustment carbon efficiency contribution value during the assessment period: in, For equipment Total carbon emissions to be avoided The total amount of electrical energy it regulates. Let be the marginal carbon emission factor of the power grid at time t; S42: Tiered and segmented distribution of all user load: Level 1 flexibility resources: If equipment The adjustment actions mainly occur in real-time or intraday time windows, and their average response delay is... The extent of the impact on key process parameters At the same time satisfy If so, it is rated as a Level 1 flexibility resource; Secondary flexibility resources: if equipment The adjustment actions are mainly arranged on a daily or weekly scale and need to be made at least in advance. The notification timeline requires coordinated adjustments to other equipment within the same process chain to simultaneously meet the following requirements. If so, it is rated as a Level 2 flexibility resource; Level 3 flexibility resources: if equipment The main basis for switching to long-term operation mode is based on monthly quota and meets the requirements. It is then assessed as a Level 3 flexibility resource, among which This is the average response delay threshold, in seconds or minutes. For the maximum allowable deviation of key process parameters such as temperature and pressure; The high carbon efficiency threshold is measured in tons of carbon dioxide per megawatt-hour; , These represent the lower and upper limits of the medium carbon efficiency contribution range, respectively. This is the threshold value for low carbon efficiency; The advance notification time threshold is in hours or days.

[0043] To better explain the technical solutions, 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 embodiments of the present invention, and not all embodiments. 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.

[0044] Example: A hierarchical evaluation method for the control flexibility of loads from high-energy-consuming industrial users, such as... Figure 1 and Figure 3 As shown, it includes the following steps: S1. Multi-dimensional basic data acquisition step: Receive and collect multi-dimensional basic data of the target high-energy-consuming industrial user; the multi-dimensional basic data includes at least: the user's monthly carbon emission quota data in the electricity market and carbon trading market, electricity price fluctuation data at different time scales, and adjustable operating parameters and process constraints data of each high-energy-consuming load equipment within the user; the output of this step is a structured multi-dimensional basic dataset. S2. Multi-timescale load regulation coordination matrix construction step: Receive the multi-dimensional basic dataset from step S1, and construct a multi-timescale load regulation coordination matrix based on it to characterize the response characteristics and constraints of user loads participating in grid regulation at different time scales; the dimensions of the coordination matrix include at least five time scales: monthly, weekly, day-ahead, intraday, and real-time; the matrix elements include at least the adjustability potential, adjustment cost, carbon emission intensity, and coupling relationship with grid dispatch instructions for specific load equipment at each time scale; the output of this step is the multi-timescale load regulation coordination matrix. S3. Multi-timescale collaborative optimization steps based on decision potential field and damping mechanism: Receive the multi-timescale load regulation collaborative matrix from step S2, and the multi-timescale regulation demand command from the external power grid. The specific processing procedure is as follows: S31. First, based on the carbon emissions, economic costs, and equipment constraints in the multi-timescale load regulation coordination matrix, a decision potential energy field spanning multiple timescales is constructed for each core optimization objective; wherein, when the system state deviates from the optimal equilibrium point in the decision potential energy field, "potential energy" will be generated. S32. Then, an artificial intelligence model is used for optimization calculation. The training and inference process of this model not only aims to maximize long-term returns, but also introduces a virtual hierarchical damping mechanism. That is, in the reward function or loss function of the model, a damping term is added to suppress the unnecessary drastic changes in the strategy in adjacent periods for the interaction between decision-making layers at different time scales, and a stronger damping coefficient is configured for decision variables with longer periods to stabilize the decision background. S33. Next, in the online application or iterative training of the artificial intelligence model, quasi-steady-state rolling coordination is implemented: when solving the optimization sub-problems at the intraday and real-time time scales, the decision variables at the previous day and monthly time scales are temporarily frozen to the current optimal values; after the short-time scale decision is executed, the results are fed back at a low frequency to gently correct the subsequent longer-term decision plans, thereby decoupling the strong real-time coupling between time scales. S34. Finally, through the comprehensive processing of S31 to S33 above, the artificial intelligence model outputs a collaborative regulation strategy. Under the guidance of the decision potential field and the stabilizing effect of the virtual hierarchical damping mechanism, this strategy can effectively avoid strategy oscillation, realize the collaborative optimization of load resources across monthly, day-ahead, intraday and real-time scales, and quantify the system-level multi-time-scale complementary regulation potential. S4. Stratification and Evaluation of Regulation Flexibility: Receive the quantification results of the coordinated regulation strategy and the multi-timescale complementary regulation potential from step S3; classify the regulation flexibility of all user loads according to the regulation timing, response speed, regulation capacity, and contribution to overall energy efficiency and carbon efficiency exhibited by the load equipment in the coordinated regulation strategy; generate a standardized flexibility grading evaluation report including flexibility level, effective regulation time window, maximum adjustable capacity, and unit regulation carbon efficiency contribution.

[0045] The specific processing flow of the multi-dimensional basic data collection steps is as follows: Day-ahead market electricity price data and real-time market electricity price data from the power trading platform, as well as long-term electricity price forecast data from the power grid dispatching agency, are received as electricity price fluctuation data; monthly and annual carbon emission quota documents from the carbon quota allocation management department, as well as carbon quota spot and futures price data and user historical carbon emission data from the carbon trading market, are received as carbon trading market data; real-time operating parameters, historical operating logs, equipment nameplate parameters, and production process flow documents from the user-side energy management system are received as load equipment data. The processing aligns the above multi-source heterogeneous data according to a unified timestamp and equipment ID, fills missing values ​​with the historical average under similar operating conditions for the equipment, and applies the Raida criterion to outliers (…). The threshold is used to remove items, and the final output is a structured multi-dimensional basic dataset. This dataset is a data table with timestamp as the primary key and includes fields such as timestamp, device ID, real-time electricity price, day-ahead electricity price forecast, carbon emission quota balance, carbon price, real-time power of the device, upper limit of adjustable power of the device, lower limit of adjustable power of the device, and associated process constraint identifier.

[0046] like Figure 2 As shown, the specific processing flow of the multi-timescale load regulation coordination matrix construction step is as follows: First, receive the structured multi-dimensional basic dataset; second, based on the monthly carbon emission quota data, decompose it step by step into nested carbon emission budget constraint boundary sets for weekly, daily, and intraday periods; third, for each load device, combining electricity price fluctuation data and its adjustable operating parameters and process constraints, with the goal of minimizing electricity costs, establish a linear programming or mixed-integer linear programming model for solution, to obtain its optimal economic adjustment range at each time scale. And collect different power adjustment amounts Based on the cost data, a piecewise linear or quadratic function is used for least squares fitting to obtain an approximate adjustment cost function for the equipment. Finally, the carbon emission budget constraint boundary, the economic adjustment range and adjustment cost function coefficients of all load equipment, and the physical limit parameters of the equipment (including the lower limit of absolute power adjustment) are included. Adjustable upper limit of absolute power Maximum power gradeability Organized by time scale as rows and load equipment as columns, each matrix element Integrate into a single data tuple containing feasible domain boundaries and cost coefficients. Thus, the multi-timescale load regulation coordination matrix is ​​output. .

[0047] The specific construction process of the decision potential energy field is as follows: receiving carbon emission budget constraint boundary data and benchmark electricity cost data at each time scale from the multi-time-scale load regulation coordination matrix; the decision potential energy field is composed of a carbon potential energy field. With economic potential field It is formed by linear weighted fusion, that is ,in and The preset weighting coefficients; the carbon potential energy field function is defined as follows: ,in Deadline The actual cumulative carbon emissions, For the carbon emissions budget at the corresponding time scale, The carbon potential coefficient; the economic potential field function is defined as follows: ,in Deadline The actual total cost of electricity, This is a reference optimal cost curve based on long-term electricity price forecasts. The economic potential coefficient is used; this step outputs a continuous potential field function defined on the multidimensional decision state space. .

[0048] The specific implementation process of the virtual hierarchical damping mechanism is as follows: receiving policy change data from the historical decision sequences of the artificial intelligence model; the damping mechanism uses the model's reward function... Add a damping penalty term to the middle To achieve this, that is ,in The damping strength coefficient; the damping penalty term The calculation formula is ,in This represents the set of decision-making levels across all time scales. This represents the policy vector at time scale s within decision period t. This refers to the layered damping coefficient corresponding to this time scale; the layered damping coefficient The configuration rule is: the longer the time scale, the more... The larger the value, the more specific the relationship is. The mechanism outputs a reinforcement learning reward function that has been adjusted for stability. .

[0049] The specific execution process of the quasi-steady-state rolling coordination is as follows: receiving the optimal long-term decision variable value at the current moment and the short-term control demand command from the power grid; when solving the intraday or real-time scale optimization sub-problem, the day-ahead planned load curve, weekly carbon emission budget allocation value, and monthly total quota are regarded as fixed parameters, and their values ​​are frozen; based on the frozen long-term parameters and real-time input, the optimal short-term load control command set is obtained by using the artificial intelligence model; after a short-term decision execution cycle ends, the actual performance is compared with the plan, and the deviation is calculated. ; with a fixed period or when Exceeding the threshold At that time, a correction process for long-term decision variables is triggered, and the correction amount is... and deviation Proportional but limited to the maximum allowable adjustment range Within, that is ,in To correct the gain coefficient, the process ultimately outputs a time-scale decoupled, co-evolving rolling decision sequence.

[0050] The specific determination process for the hierarchical and evaluation steps of the regulation flexibility is as follows: receiving the timing sequence of regulation actions, instantaneous regulation rate data, cumulative regulation capacity data, and corresponding unit regulation carbon efficiency contribution value of each load device in the coordinated regulation strategy; calculating the regulation carbon efficiency contribution value of each load device. Unit adjustment carbon efficiency contribution value during the assessment period ,in For equipment Total carbon emissions to be avoided The total amount of electrical energy it regulates. Let be the marginal carbon emission factor of the power grid at time t.

[0051] The specific method for stratifying and classifying the entire user load is as follows: If the equipment The adjustment actions mainly occur in real-time or intraday time windows, and their average response delay is... The extent of the impact on key process parameters At the same time satisfy If so, it is rated as a Level 1 flexibility resource; If the device The adjustment actions are mainly arranged on a daily or weekly scale and need to be made at least in advance. The notification timeline requires coordinated adjustments to other equipment within the same process chain to simultaneously meet the following requirements. If so, it is rated as a Level 2 flexibility resource; If the device The main basis for switching to long-term operation mode is based on monthly quota and meets the requirements. If so, it is rated as a Level 3 flexibility resource.

[0052] in , , , , , and All are preset thresholds.

[0053] The structure and training process of the artificial intelligence model are as follows: The model adopts an Actor-Critic framework, where the Actor network is responsible for outputting the control strategy of load equipment, and the Critic network is responsible for evaluating the state value. The input feature vector of the model is composed of three parts: the first part is the flattened encoding of the multi-timescale load control coordination matrix, the second part is the encoding of the multi-timescale control demand command of the power grid, and the third part is the current system state including cumulative carbon emissions and costs. The output of the Actor network is a multi-dimensional continuous action space, where each dimension corresponds to the power adjustment amount of a load device in the next decision cycle. The output of the Critic network is the value estimate of the current state. The model training process is carried out in a simulated environment containing electricity market fluctuations, carbon price changes, and stochastic power grid demand commands, and the network parameters are updated using a near-end policy optimization algorithm. The training objective is to minimize the comprehensive loss function containing the original reward, potential field guidance term, and damping penalty term. The training process continues until the model's policy performance in the test environment converges and the variance of the fluctuation is lower than a preset threshold. After training is completed, the network parameters are fixed for online deployment. The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for hierarchical evaluation of the control flexibility of loads from high-energy-consuming industrial users, characterized in that... It includes the following steps: S1. Collect multi-dimensional basic data of target users and construct a structured multi-dimensional basic dataset; S2. Construction of Multi-Time-Scale Load Regulation Coordination Matrix: Based on the multi-dimensional basic dataset obtained in step S1, construct a multi-time-scale load regulation coordination matrix that characterizes the response characteristics and constraints of user loads participating in power grid regulation at different time scales, and output the matrix. S3. Multi-timescale collaborative optimization based on decision potential energy field and damping mechanism: The multi-timescale load regulation collaborative matrix constructed in step S2 and the external power grid multi-timescale regulation demand command are used as inputs. By introducing decision potential energy field and virtual hierarchical damping mechanism, the collaborative optimization calculation is performed using artificial intelligence model. A collaborative regulation strategy that can avoid strategy oscillation and realize cross-timescale load resource collaborative optimization is output, and the system-level multi-timescale complementary regulation potential is quantified. S4. Stratification and Evaluation of Control Flexibility: Based on the collaborative control strategy obtained in step S3 and the quantitative results of the multi-timescale complementary control potential, the control flexibility of all user loads is stratified and graded according to the control timing, response speed, control capacity, and unit control carbon efficiency contribution of the load equipment, and a standardized flexibility grading evaluation report is generated to improve the control flexibility of high-energy-consuming industrial user loads.

2. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 1, characterized in that... The multi-dimensional basic data in step S1 includes at least monthly carbon emission quota data, electricity price fluctuation data at different time scales, and adjustable operating parameters and process constraints data of each high-energy-consuming load equipment. Step S1 specifically includes: receiving electricity price data from the power trading platform and the power grid dispatching agency, receiving monthly / annual carbon emission quota data from the carbon quota allocation management department, as well as carbon quota spot and futures price data from the carbon trading market, historical carbon emission data from users, and load equipment data from the user-side energy management system; aligning the collected data from multiple sources and heterogeneous data, filling missing values ​​and removing outliers, and outputting a structured multi-dimensional basic dataset. The structured multi-dimensional basic dataset obtained in step S1 is a data table with timestamp as the primary key, containing the following dimensions: [timestamp, device ID, real-time electricity price, day-ahead electricity price forecast, carbon emission quota balance, carbon price, real-time device power, upper limit of adjustable device power, lower limit of adjustable device power, associated process constraint identifier]. It is generated by aligning multi-source heterogeneous data, filling missing values ​​and removing outliers, and then associating and regularizing the multi-source data according to a unified timestamp and device ID.

3. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 1, characterized in that... The dimensions of the multi-timescale load regulation coordination matrix in step S2 include at least five time scales: monthly, weekly, daily, intraday, and real-time. The matrix elements include at least the adjustability potential, adjustment cost, and carbon emission intensity of a specific load device at each time scale.

4. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 1, characterized in that... Step S2 specifically includes: S21. Decompose the monthly carbon emission allowance into nested carbon emission budget constraint boundary sets for weekly, daily, and intraday periods. S22. For each load device, combining electricity price fluctuation data and its adjustable operating parameters, solve for its optimal economic adjustment range at each time scale. And fit the adjustment cost function ; S23. Combine the carbon emission budget constraint boundary set obtained in step S21 with the optimal economic adjustable range of each load device obtained in step S22. Adjustment cost function and physical limit parameters The data is organized by time scale as rows and load equipment as columns. Each matrix element is integrated into a data tuple containing feasible domain boundaries and cost coefficients, and the multi-time scale load regulation coordination matrix is ​​output.

5. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 4, characterized in that... Step S22 specifically includes: S221. Interval Solution: With the objective of minimizing electricity costs, and constrained by the upper and lower limits of the equipment's adjustable power and the allowable fluctuation ranges of temperature and pressure, linear programming (LP) or mixed-integer linear programming (MILP) models are established and solved for the corresponding electricity price sequences at monthly, day-ahead, and real-time time scales. This yields the economically optimal value for equipment power adjustment and the allowable deviation range at each scale, i.e., the optimal economic adjustment interval. , where t is the time point; S222, Function Fitting: Collect cost change data for the equipment under different power adjustment amounts ΔP. adjust = ΔP * Electricity price for the corresponding time period, using a quadratic function on (ΔP, Cost) adjust The data points are fitted with least squares to obtain an approximate adjustment cost function for the device, which is then used to predict the adjustment cost at other times. The coefficients a, b, and c are determined by fitting, and t is a time point. For different time scales, the same quadratic function form as shown in the above formula is used, but the coefficients are updated according to the typical electricity price level of that scale. The multi-timescale load regulation coordination matrix in step S23 is a two-dimensional structure, i.e., a multi-timescale load regulation coordination matrix. That is, all matrix elements Collections organized by rows and columns: Where the row index is the time scale. S is the set of decision-making levels across all time scales, i.e.: The column index is for load devices. I is the set of all target high-energy-consuming industrial user-side load devices participating in power grid regulation; matrix elements It is a structured data tuple representing time. Time scale Lower equipment The feasible domain and cost characteristics of regulation, namely: in, It is the carbon emission budget boundary, obtained from step S21, representing the time scale. At any moment The remaining carbon emissions budget cap; The economic adjustment range, obtained from step S22, represents the equipment's... At any moment The economically optimal power adjustment range that meets process constraints; These are the adjustment cost function coefficients, which are the fitting results from step S22, i.e., if the cost function is in quadratic form. Then the coefficient set is stored here. ; It is the lower limit of absolute power adjustment; It is the upper limit of absolute power adjustment; This refers to the maximum power ramp rate; the absolute power adjustable lower limit. Adjustable upper limit of absolute power Maximum power gradeability These are all physical limit parameters, which are the equipment's... The inherent constraints.

6. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 5, characterized in that... In step S222, the approximate adjustment cost function retains its form for different time scales (monthly, day-ahead, and real-time), but the coefficients a, b, and c need to be determined separately. The method for determining these coefficients is as follows: each time scale has its corresponding typical electricity price curve or sequence; the typical electricity price data for a specific time scale is used as input, along with the power adjustment amount of the equipment. Together, cost data sets at this scale are generated. Then, the least squares method is used to refit the coefficients a, b, c at that time scale. Therefore, the coefficients are determined by fitting, and the electricity price level at different time scales drives independent fitting to obtain different sets of coefficients.

7. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 1, characterized in that... In step S3, the decision potential field is determined by the carbon potential field U. c With economic potential field U e It is formed by weighted fusion, that is: U total= aU c +βU e Where α and β are preset weighting coefficients; The carbon potential energy field U c Defined as the following function: Among them, C actual (t) represents the actual cumulative carbon emissions up to time t, C budget (t) represents the carbon emission budget for the corresponding time scale, k c The carbon potential coefficient is defined as follows: The economic potential field function is defined as: Cost actual (t) represents the actual total electricity cost up to time t, Cost reference (t) represents the reference optimal cost curve based on long-term electricity price forecasts, and k e This represents the economic potential coefficient.

8. The method for graded evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 7, characterized in that... The virtual hierarchical damping mechanism in step S3 refers to the mechanism implemented through the reward function R of the artificial intelligence model. total Add a damping penalty term R to (t) damp Implementation, that is: Where λ is the damping strength coefficient; The damping penalty term R damp The calculation formula is: Where S is the set of all time-scale decision-making layers. Its specific composition is determined based on the actual needs of power grid dispatching and user production plans. πs(t) is the strategy vector of time scale s in decision period τ, and ω s Let be the stratified damping coefficient corresponding to this time scale, and satisfy: The formula for calculating the correction amount in step S3, which is triggered at a fixed period or when the deviation between actual performance and the plan exceeds a threshold, is as follows: in, To correct the gain coefficient, The deviation between actual performance and plan, This represents the maximum allowable adjustment range. Finally, through computation, the following two specific results are output: ① Coordinated regulation strategy: The strategy is a multi-dimensional instruction sequence A(t'), that is, for each future decision time t' and each load device i, a specific power adjustment instruction value ΔP is output. i (t'), then: Where each element ΔP i (t') represents the specific power adjustment value of load device i at decision time t'. The multidimensional instruction sequence A(t') satisfies the electrical, carbon, and equipment constraints at all time scales, and is smooth and stable across time scales under the guidance of the potential energy field and the damping mechanism. ②Quantitative results of complementary regulation potential across multiple time scales, including: Monthly potential: = ∑(Monthly adjustable capacity of all devices); Current potential: = ∑(Capacity of all devices capable of rapid response at the day-ahead scale); Real-time potential: = ∑(the capacity of all devices that can be adjusted instantaneously at a real-time scale); Complementarity coefficient: This is used to measure the potential gain from cross-scale collaboration. The monthly potential V is... month Potential V day Real-time potential V real Both the complementarity coefficient α and the maximum adjustable capacity extracted by analyzing the synergistic control strategy at various time scales are calculated.

9. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 7, characterized in that... The artificial intelligence model in step S3 specifically refers to the use of the Actor-Critic framework; the training process is carried out in a simulation environment that includes electricity market fluctuations, carbon price changes and stochastic grid demand commands, and the training objective is to minimize the comprehensive loss function that includes the original reward, the potential energy field guidance term and the damping penalty term. The output of the Actor network in the Actor-Critic framework is the action. The input is the system state, that is: ； The Critic network in the Actor-Critic framework evaluates state value. The training objective is set to minimize the overall loss function, i.e.: in, and For the standard policy gradient and value loss, For the decision potential field, For damping penalty term, , , The corresponding weights are used; this design enables the model to simultaneously optimize economic efficiency, carbon emission targets, and suppress policy oscillations during learning.

10. The method for hierarchical evaluation of control flexibility for high-energy-consuming industrial user loads according to claim 1, characterized in that... Step S4 specifically includes: S41: Based on the multi-dimensional instruction sequence output from step S3, i.e., the coordinated control strategy vector. And grid carbon emission intensity data, calculate the load equipment Unit adjustment carbon efficiency contribution value during the assessment period: in, For equipment Total carbon emissions to be avoided The total amount of electrical energy it regulates. Let be the marginal carbon emission factor of the power grid at time t; S42: Tiered and segmented distribution of all user load: Level 1 flexibility resources: If equipment The adjustment actions mainly occur in real-time or intraday time windows, and their average response delay is... The extent of the impact on key process parameters At the same time satisfy If so, it is rated as a Level 1 flexibility resource; Level 2 flexibility resources: if equipment The adjustment actions are mainly arranged on a daily or weekly scale and need to be made at least in advance. The notification timeline requires coordinated adjustments to other equipment within the same process chain to simultaneously meet the following requirements. If so, it is rated as a Level 2 flexibility resource; Level 3 flexibility resources: if equipment The main basis for switching to long-term operation mode is based on monthly quota and meets the requirements. It is then assessed as a Level 3 flexibility resource, among which This is the average response delay threshold, in seconds or minutes. For the maximum allowable deviation of key process parameters such as temperature and pressure; The high carbon efficiency threshold is measured in tons of carbon dioxide per megawatt-hour; , These represent the lower and upper limits of the medium carbon efficiency contribution range, respectively. This is the threshold value for low carbon efficiency; The advance notification time threshold is in hours or days.

Images