Artificial intelligence-based energy consumption comprehensive intelligent regulation method and system

Abstract

The application discloses an energy consumption comprehensive intelligent regulation method and system based on artificial intelligence, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring and processing energy consumption regulation data, forming energy consumption comprehensive state data set, process control topology data and control constraint configuration data, identifying time delay action relationship verified by conditional independence and error improvement through historical control period and operation context, determining a prediction time domain, constructing a topology constraint time delay control action graph, an artificial intelligence time sequence energy consumption response model and a prediction error envelope; determining a candidate control action variable group based on the topology constraint time delay control action graph, forming a current benchmark prediction trajectory, generating and screening candidate regulation trajectories, and obtaining an executable energy consumption regulation instruction set; and executing the executable energy consumption regulation instruction set to form regulation execution feedback data. The application improves the prediction accuracy, safety and real energy-saving capability of energy consumption comprehensive intelligent regulation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a comprehensive intelligent control method and system for energy consumption based on artificial intelligence. Background Technology

[0002] With the continuous improvement of digitalization, networking, and automation in industrial production, energy consumption management has gradually evolved from traditional manual meter reading, quota calculation, and post-event statistics to online monitoring and intelligent control throughout the entire production process. Existing integrated intelligent energy consumption control systems typically connect to electricity meters, gas meters, steam flow meters, and equipment controllers through an energy management platform. This aggregates energy metering data, equipment operation data, production task data, and environmental data, performs unified conversions for different energy media, and analyzes energy consumption levels and trends by production unit, equipment, or process. Building upon this, some systems further integrate PLCs, DCSs, and production execution systems to adjust equipment start-up and shutdown, operating frequency, speed, valve opening, and temperature, pressure, and flow setpoints based on equipment load, production cycle time, and process setpoints. In recent years, technologies such as neural networks, time-series predictive models, and model predictive control have also been increasingly applied to energy consumption control. These technologies predict subsequent energy consumption and process status using historical operating data and select the most effective control scheme from multiple options, enabling energy consumption management to evolve from simple monitoring, statistics, and early warning to predictive analysis, proactive decision-making, and closed-loop control.

[0003] Although current integrated intelligent energy consumption control technologies can achieve multi-source data acquisition, energy consumption prediction, and control command issuance, problems still exist. First, existing methods struggle to accurately identify the impact path and lag time of control variables on energy consumption and process state by combining equipment connection relationships, operating conditions, and control response processes, leading to a mismatch between the prediction model and the dynamic changes in the actual production process. Second, when generating control schemes, existing methods typically focus on predicting energy consumption reduction, neglecting process safety, changes in energy accumulation status, and feasibility of execution, and lacking mechanisms for safety correction of candidate schemes that do not meet constraints. Therefore, current integrated intelligent energy consumption control still suffers from inaccurate identification of control effects and insufficient reliability of control schemes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent energy consumption control method and system based on artificial intelligence to solve the problems of inaccurate identification of control action and insufficient reliability of control scheme.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent energy consumption control method based on artificial intelligence, comprising:

[0008] Acquire and process energy consumption regulation data to form a comprehensive energy consumption status dataset, process control topology data, and control constraint configuration data. Identify the time delay relationship verified by conditional independence and error improvement through historical control cycles and operating contexts, determine the prediction time domain, and construct a topology constraint time delay control action diagram, an artificial intelligence time series energy consumption response model, and a prediction error envelope.

[0009] Based on the topological constraint time delay control action diagram, candidate control action variable groups are determined, and the current benchmark prediction trajectory is formed. Candidate control trajectories are generated and screened to obtain a set of executable energy consumption control instructions.

[0010] Execute a set of executable energy consumption control commands to generate control execution feedback data, confirm the actual control status based on the control execution feedback data, and generate actual control execution results;

[0011] Based on the actual control and regulation results, the multi-cycle energy saving effect and rebound energy consumption are verified, and a closed-loop control and regulation evaluation result is formed. Based on the closed-loop control and regulation evaluation result, the prediction and control basis is revised, and a comprehensive intelligent control result for energy consumption is generated.

[0012] As a preferred embodiment of the AI-based intelligent energy consumption control method of the present invention, the step of acquiring and processing energy consumption control data to form an energy consumption comprehensive status dataset, process control topology data, and control constraint configuration data refers to collecting energy metering, equipment operation, production tasks, environmental disturbances, process status, and actuator feedback data according to the control cycle, performing time alignment, anomaly marking, and energy unit conversion on various types of data, calculating the comprehensive energy consumption, and associating equipment connection, energy transmission, control logic, and process constraints to form an energy consumption comprehensive status dataset, process control topology data, and control constraint configuration data.

[0013] As a preferred embodiment of the artificial intelligence-based integrated intelligent energy consumption control method of the present invention, the construction of the topological constraint time-delay control action diagram, the artificial intelligence time-series energy consumption response model, and the prediction error envelope includes:

[0014] Based on the comprehensive energy consumption status dataset, effective historical control cycles are screened, and control variables, uncontrollable variables, equipment status variables, process status variables, and operating context variables are determined according to the data source and its corresponding control and acquisition objects.

[0015] Initial directed edges are established based on process control topology and variable categories. Directed edges are then selected by combining the effect of conditional time delay and prediction error improvement, and the lag period is determined to form a topological constraint time delay control effect graph.

[0016] Based on the topological constraint time delay control action diagram, the action path from the control variable to the comprehensive energy consumption is extracted, the energy accumulation state and the maximum historical recovery time are identified, and the prediction time domain length is determined by combining the path accumulation time delay and the maximum response time of the actuator.

[0017] Time series samples are constructed based on effective historical control cycles, variable categories, energy accumulation states, and prediction time domain lengths. An artificial intelligence time series energy consumption response model is trained, and an independent calibration sample is used to form a prediction error envelope.

[0018] As a preferred embodiment of the AI-based intelligent energy consumption control method of the present invention, the step of determining the candidate control variable group based on the topology-constrained time-delay control action diagram and forming the current baseline prediction trajectory refers to forming a combination of the current historical state sequence, the current uncontrollable variable sequence, the current operating context variable, and the current control variable based on the comprehensive energy consumption state dataset, screening the intervenable control variables acting on the comprehensive energy consumption node in the prediction time domain according to the topology-constrained time-delay control action diagram, forming a set of candidate control variable groups, and obtaining the current baseline prediction trajectory using an AI time-series energy consumption response model.

[0019] As a preferred embodiment of the artificial intelligence-based comprehensive intelligent energy consumption control method of the present invention, the step of generating and screening candidate control trajectories to obtain a set of executable energy consumption control instructions includes:

[0020] Based on the candidate control variable group, the current control variable combination, the operating context, the process control topology, and the actuator constraints, candidate control combinations are generated, linkage control is supplemented and historically restricted combinations are excluded, and a candidate control variable sequence is formed according to the execution order and response time to obtain a set of candidate control schemes.

[0021] The candidate control variable sequence, the current historical state sequence, and the current uncontrollable variable sequence are input into the artificial intelligence time series energy consumption response model. The process constraints, terminal state equivalence, and conservative energy saving judgments are made by combining the prediction error envelope and control constraints to form a set of feasible control schemes and a set of control schemes to be safely corrected.

[0022] Keeping the equipment start-stop status and operating mode in the control scheme to be corrected unchanged, generating a correction combination that rolls back to the current control value according to the control resolution, re-executing the process constraint, terminal state equivalence and conservative energy saving judgment, and selecting the combination with the smallest modification distance and all passing to form a set of safety correction control schemes;

[0023] Feasible control schemes and safety correction control schemes are used as alternative execution schemes. The target control scheme is selected according to the number of control variables, the conservative energy saving lower limit, and the control adjustment range. The set of executable energy consumption control instructions is formed by combining historical optimization records, execution order, and safety backoff rules.

[0024] As a preferred embodiment of the AI-based integrated intelligent energy consumption control method of the present invention, the step of executing the set of executable energy consumption control instructions to form control execution feedback data refers to reading the set of executable energy consumption control instructions and saving the state data before control, issuing control instructions to the corresponding controllers according to the instruction execution order, and forming control execution feedback data based on the execution results and the actual combination of control variables.

[0025] As a preferred embodiment of the intelligent energy consumption control method based on artificial intelligence described in this invention, the step of confirming the actual control state based on control execution feedback data and forming the actual control execution result refers to judging whether an effective control effect has been formed based on the control execution feedback data. When the combination of actual control variables is inconsistent with the combination of target control variables, an actual control variable sequence is formed, and the process constraints, terminal state equivalence, and conservative energy saving judgment are re-performed using an artificial intelligence time-series energy consumption response model. The stability of the actual control state that meets the judgment conditions is confirmed, and the actual control execution result is formed.

[0026] As a preferred embodiment of the AI-based integrated intelligent energy consumption control method of the present invention, the step of verifying multi-cycle energy-saving effects and rebound energy consumption based on actual control execution results to form a closed-loop control evaluation result includes:

[0027] Based on the actual control execution results, verification data in the prediction time domain is collected for the control process that forms an effective control effect and a stable control state, forming post-control verification time domain state data, actual uncontrollable variable sequence and updated and maintained control variable sequence;

[0028] The historical state sequence before regulation, the updated and maintained control variable sequence, and the actual uncontrollable variable sequence are input into the artificial intelligence time series energy consumption response model to obtain the updated and maintained control prediction trajectory. Combined with the actual cumulative net energy saving, the cumulative energy consumption prediction error envelope, and the compliance of the actual terminal state and process state, a closed-loop regulation evaluation result is formed.

[0029] As a preferred embodiment of the AI-based integrated intelligent energy consumption control method of the present invention, the step of modifying the prediction and control basis according to the closed-loop control evaluation results and generating integrated intelligent energy consumption control results refers to selecting control processes with stable actual control states, fault-free equipment, and complete verification data based on the closed-loop control evaluation results, forming new control response samples, updating the AI ​​time-series energy consumption response model, prediction error envelope, and topological constraint time delay control action diagram, and writing effective, rebound, or unavailable control combinations into the corresponding historical record databases to generate integrated intelligent energy consumption control results.

[0030] Secondly, the present invention provides an intelligent energy consumption control system based on artificial intelligence, comprising:

[0031] The multi-source data construction module is used to collect and verify multi-source energy consumption data, convert them in a unified manner, and form comprehensive state, process topology, and control constraint data.

[0032] The time delay effect modeling module is used to classify variables, identify time delay control effects, determine the prediction time domain, and train an artificial intelligence energy consumption response model.

[0033] The intelligent control and decision-making module is used to screen controllable variables, generate candidate solutions, and complete safety, terminal, and energy-saving evaluations, as well as correction and optimization.

[0034] The instruction execution verification module is used to issue target control instructions, collect execution feedback, and verify the actual control status, execution deviation, and operational stability.

[0035] The closed-loop evaluation and update module is used to verify the actual energy-saving effect, process constraints and terminal status, and update the model, action diagram and historical control record library.

[0036] The beneficial effects of this invention are as follows: By using variable classification, operating context constraints, control action path identification, and lag period analysis techniques, it achieves an accurate characterization of the dynamic relationship between control variables, process states, and comprehensive energy consumption. Furthermore, by combining the energy accumulation state recovery process, it rationally determines the prediction time domain, improving the matching degree between energy consumption and process state predictions and actual production processes. Through candidate scheme generation, process constraint verification, terminal state equivalence judgment, conservative energy-saving evaluation, and minimum amplitude safety correction techniques, it achieves hierarchical screening and optimization of control schemes, reducing the risks of superficial energy saving, state out-of-bounds behavior, and over-adjustment, thereby improving the safety of energy consumption control, the actual energy-saving effect, and the reliability of on-site execution. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of an artificial intelligence-based integrated intelligent energy consumption control method.

[0039] Figure 2 This is a schematic diagram of an integrated intelligent energy consumption control system based on artificial intelligence.

[0040] Figure 3 A flowchart for constructing the time series model and error envelope is provided.

[0041] Figure 4 Develop a flowchart for the target control scheme. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figure 1 , Figure 3 and Figure 4 This is one embodiment of the present invention, which provides an intelligent energy consumption control method based on artificial intelligence, comprising the following steps:

[0046] S1. Acquire and process energy consumption regulation data to form a comprehensive energy consumption status dataset, process control topology data, and control constraint configuration data. Identify the time delay relationship verified by conditional independence and error improvement through historical control cycles and operating contexts, determine the prediction time domain, and construct a topology constraint time delay control action diagram, an artificial intelligence time series energy consumption response model, and a prediction error envelope.

[0047] S1.1: Collect energy metering, equipment operation, production tasks, environmental disturbances, process status and actuator feedback data according to the control cycle, perform time alignment, anomaly marking and energy unit conversion on various types of data, calculate comprehensive energy consumption, and associate with equipment connection, energy transmission, control logic and process constraints to form comprehensive energy consumption status dataset, process control topology data and control constraint configuration data.

[0048] Specifically, energy metering data, equipment operation data, production task data, environmental disturbance data, process status data, and actuator feedback data are collected according to the control cycle. This includes reading equipment connection relationships, energy transmission relationships, control logic, process interlock relationships, process transfer time, actuator control range, single allowable adjustment amount, allowable adjustment direction, control resolution, allowable response time, allowable execution error, allowable equipment state transition relationships, actuator transition response characteristics, safety backoff rules, allowable process status range, joint process constraint parameters, allowable process status stability deviation, and maximum allowable recovery time. Using the start and end times of the control cycle as a unified time boundary, all types of data are time-aligned. Data with timestamps exceeding the corresponding control cycle, cumulative metering readings going backwards, or equipment operation status inconsistent with actuator feedback data are marked as abnormal data, and the corresponding abnormality reasons are retained. From the time-aligned valid energy metering data, the cumulative metering readings of various energy media at the start and end times of the control cycle are read. The difference between the cumulative metering reading at the end time and the cumulative metering reading at the start time is determined as the cumulative metering difference for the corresponding energy media. Based on the cumulative metering difference, the system calculates the energy consumption of various types and the comprehensive energy consumption of the current control cycle. Using the control cycle as the basis for time correlation, the comprehensive energy consumption is correlated with equipment operation data, production task data, environmental disturbance data, process status data, actuator feedback data, and anomaly indicators within the same control cycle. Object matching is performed according to the production unit, equipment, sensor, or actuator to which each data belongs, forming a comprehensive energy consumption status dataset arranged by control cycle. For equipment connection relationships, energy transmission relationships, control logic, and process interlocking relationships, records are made according to the objects at both ends of the relationship, the relationship type, and the relationship direction. The corresponding process transfer time, actuator control constraints, and permissible process status range are then correlated to form process control topology data. Simultaneously, production unit identifiers, equipment operation modes, product types, production stages, and applicable environmental conditions are used as operating condition matching conditions. Constrained process status identifiers, applicable control cycles for constraints, permissible upper and lower limits of process status, joint process constraint coefficients, permissible boundaries of joint processes, permissible deviations of process status stability, and maximum permissible recovery time are matched to form control constraint configuration data.

[0049] The formula for calculating energy consumption is:

[0050] ;

[0051] In the formula, It is the first Within the first control cycle, the first Energy consumption of energy-type media This is the cumulative metering difference corresponding to the energy medium. It is the energy conversion factor for the corresponding energy medium.

[0052] The formula for calculating total energy consumption is:

[0053] ;

[0054] In the formula, It is the first The total energy consumption per control cycle It refers to the number of energy media types involved in the comprehensive energy consumption calculation.

[0055] It should be noted that energy metering data includes cumulative electricity meter readings, cumulative gas consumption, and cumulative steam consumption; equipment operation data includes equipment start-up and shutdown status, operating mode, load rate, speed, frequency, and valve opening; production task data includes target output, actual output, production cycle time, product type, and production stage; environmental disturbance data includes ambient temperature and ambient humidity; process status data includes process temperature, pressure, flow rate, liquid level, gas storage capacity, intermediate product inventory, equipment thermal status, and product quality inspection values; and actuator feedback data includes actual control values, actual operating status, instruction issuance time, actual arrival time, instruction reception status, and fault status.

[0056] S1.2: Based on the comprehensive energy consumption status dataset, select effective historical control cycles, and determine control variables, uncontrollable variables, equipment status variables, process status variables, and operating context variables according to the data source and its corresponding control and acquisition objects.

[0057] Specifically, a set of valid historical control cycles is formed by selecting historical control cycles from the comprehensive energy consumption status dataset that have valid energy metering data, no equipment malfunctions, normal production process completion, and product quality inspection values ​​meeting process requirements. From this set, equipment operation data, production task data, environmental disturbance data, process status data, and comprehensive energy consumption are read. Based on the correspondence between these data and actuators, production tasks, equipment, and sensors, control variables, uncontrollable variables, equipment status variables, and process status variables are determined. Equipment start / stop status, operating mode, frequency, speed, valve opening, and temperature that can be directly changed through PLC, DCS, or equipment controllers are also included. Setpoints, pressure setpoints, and flow setpoints are defined as control variables; production task data, environmental disturbance data, equipment availability status, and production stage that are not directly changed by the current energy consumption control command are defined as uncontrollable variables; equipment fault status and the equipment's own operating status that are not defined as control variables or process state variables are defined as equipment state variables; temperature, pressure, flow rate, liquid level, gas storage capacity, intermediate product inventory, equipment thermal status, and product quality detection values ​​collected by process sensors are defined as process state variables; product type, production stage, environmental disturbance status, and equipment operating mode are read from production task data, environmental disturbance data, and equipment operating data to form a set of operating context variables.

[0058] It should be noted that the runtime context variable set is used to control the impact of different products, production stages, environmental conditions and equipment operating modes on the identification of node relationships, without changing the original variable categories of the corresponding data.

[0059] S1.3: Based on the process control topology and variable categories, initial directed edges are established. Combined with the effect of conditional time delay and prediction error improvement, directed edges are screened and the lag period is determined to form a topological constraint time delay control effect graph.

[0060] Specifically, control variable nodes, equipment status nodes, process status nodes, production task nodes, environmental disturbance nodes, and comprehensive energy consumption nodes are established according to variable categories. Equipment connection relationships, energy transmission relationships, and control logic in the process control topology data are read line by line, and the objects at both ends of each relationship are mapped to the corresponding variable nodes. For equipment connection relationships, directed edges are established from upstream equipment nodes to downstream equipment nodes. For energy transmission relationships, directed edges are established from the energy outflow object node to the energy inflow object node according to the actual energy transmission direction. For control logic, directed edges are established from control variable nodes to the corresponding controlled equipment status nodes or process status nodes. When there is a direct measurement relationship between an equipment status node or process status node and comprehensive energy consumption, a directed edge is further established from the corresponding node to the comprehensive energy consumption node, ultimately forming the initial directed edge. For the initial directed edge issued by the control variable node, the sum of the allowable response time of the corresponding actuator and the process transmission time is divided by the length of a single control cycle and rounded up to determine the maximum candidate lag period. For the initial directed edges between other nodes... Starting with a directed edge, the corresponding process transfer time is divided by the length of a single control cycle and rounded up to determine the maximum candidate lag cycle. The first control cycle to the maximum candidate lag cycle is then grouped into a candidate lag cycle set. The effective historical control cycle set is divided into interaction relationship training data and interaction relationship verification data according to time sequence. For the initial directed edge from the source node to the target node, historical data of the source node within the candidate lag cycle range is read from the interaction relationship training data. Historical data of the target node, common upstream node data, and runtime context variables are also read to calculate the conditional delay effect information of the source node's historical data relative to the target node. Keeping the target node's historical data, common upstream node data, and runtime context variables unchanged, the historical data of the source node within the candidate lag cycle range is repeatedly randomly permuted, and the conditional delay effect information is recalculated to form a permuted conditional delay effect information set. The original conditional delay effect information is compared with the permuted conditional delay effect information set. When the original conditional delay effect information is less than or equal to the value in the permuted conditional delay effect information set that corresponds to a preset significance level... corresponding When determining empirical quantiles, the source node and target node are conditionally independent, and the corresponding initial directed edge is deleted; otherwise, they are conditionally dependent, and the corresponding initial directed edge is retained as a directed edge to be confirmed. For each directed edge to be confirmed, source node lag data from the candidate lag period set is selected, and a baseline prediction relationship is established using historical data of the target node, common upstream node data, and runtime context variables. The source node data of the corresponding lag period is added to the baseline prediction relationship to form an enhanced prediction relationship. The interaction relationship verification data is divided into continuous verification data segments according to time sequence, and the average of the baseline prediction relationship and the enhanced prediction relationship on each verification data segment is calculated. The absolute prediction error is calculated, and the prediction error improvement corresponding to the candidate lag period is calculated. The lower bound of the stability error improvement is calculated based on the prediction error improvement of each verification data segment. When the lower bound of the stability error improvement is greater than 0, the corresponding candidate lag period is determined as the effective lag period. When there are two or more effective lag periods, the effective lag period with the smallest average absolute prediction error of the enhanced prediction relationship on all action relationship verification data is determined as the number of lag periods of the corresponding directed edge. When there are no effective lag periods, the corresponding directed edge to be confirmed is deleted. The finally retained directed edge, control action direction and number of lag periods are associated to form the topological constraint time delay control action graph.

[0061] The formula for calculating the effect of conditional delay is:

[0062] ;

[0063] In the formula, It is the source node Relative to the target node The conditional delay effect information is represented in bits or nanoseconds. Is the target node at the ? The value for each control cycle, The source node is in the maximum candidate lag period The historical value sequence within, It includes historical data of the target node, data of common upstream nodes other than the source node, and runtime context variables.

[0064] The formula for calculating the improvement in prediction error is:

[0065] ;

[0066] In the formula, It is the first The first verification data segment The improvement in prediction error corresponding to each candidate lag period It is the mean absolute prediction error of the baseline prediction relationship. It is the mean absolute prediction error that enhances the prediction relationship.

[0067] The formula for calculating the lower bound of the stability error improvement is:

[0068] ;

[0069] In the formula, It is the first The lower bound of the stability error improvement corresponding to each candidate lag period It is a set of verification data segments. It is at a lower quantile level The corresponding empirical quantile.

[0070] It should be noted that, By setting the error improvement amount of each continuous verification data segment using empirical quantile statistics, the typical value range is [0.10~0.30], and the preferred value is 0.20. The 20th percentile can require that the error improvement of most continuous verification data segments is positive, avoiding the retention of directed edges that are effective only in a few data segments.

[0071] Preset significance level The risk of misjudgment allowed by the conditional independence test is pre-set and remains unchanged in the same action relationship identification process. The typical value range is [0.01~0.10], and the preferred value is 0.05. A smaller significance level can improve the strictness of the directed edge retention condition. In this embodiment, it is 0.05 to balance the exclusion of accidental associations and the retention of effective action relationships.

[0072] S1.4: Extract the action path from the control variable to the comprehensive energy consumption based on the topological constraint time delay control action diagram, identify the energy accumulation state and the maximum historical recovery time, and determine the prediction time domain length by combining the path accumulation time delay and the maximum response time of the actuator.

[0073] Specifically, the directed action paths from the control variable nodes to the comprehensive energy consumption nodes are read from the topological constraint time-delay control action diagram, and process state nodes on each directed action path are extracted. Based on the physical type corresponding to the process state identifier, process state nodes with energy or material accumulation characteristics such as temperature, pressure, liquid level, gas storage, intermediate product inventory, equipment thermal state, and others are selected from the process state nodes to form energy accumulation state candidate nodes. For each energy accumulation state candidate node, the time of issuance of the control command, the allowable stable range before control, and the process state sequence after control are read from the effective historical control cycle set, and records of equipment failure, production interruption, and data anomaly are excluded. When the process state responds and continues to change across control cycles or gradually recovers to the allowable stable range before control, it is identified as an energy accumulation state node, and the control command is issued to the process state to re-enter and maintain this state continuously. The time interval between the allowable stable ranges before regulation is determined as the historical recovery time; the maximum historical recovery time corresponding to each energy accumulation state is determined as the maximum historical recovery time of the energy accumulation state; when there is no valid historical recovery record for a certain energy accumulation state, the maximum allowable recovery time corresponding to the control constraint configuration data is adopted; when no energy accumulation state node is formed, the maximum historical recovery time of the energy accumulation state is set to zero; and each directed action path from the control variable node to the comprehensive energy consumption node in the topology constraint delay control action diagram is read, the sum of the action durations corresponding to each directed edge on each directed action path is calculated, and the maximum value is determined as the maximum path cumulative action delay; the prediction duration is determined based on the maximum path cumulative action delay, the maximum allowable response time of the actuator, and the maximum historical recovery time of the energy accumulation state, and the prediction time domain length is determined based on the prediction duration and the length of a single control cycle.

[0074] The formula for calculating the maximum cumulative path delay is:

[0075] ;

[0076] In the formula, It is the maximum cumulative path delay. It is the set of directed action paths from the control variable node to the comprehensive energy consumption node. Directed edge The number of lag periods, It is the length of a single control cycle.

[0077] The predicted duration is:

[0078] ;

[0079] In the formula, It is the predicted duration. This is the maximum permissible response time for the executing agency. It is the maximum historical recovery time of the energy accumulation state.

[0080] The prediction time domain length is:

[0081] ;

[0082] In the formula, It is the prediction time domain length, representing the number of control cycles that continuously participate in the prediction of subsequent energy consumption and process status starting from the target control cycle.

[0083] It should be noted that, Based on statistical analysis of the energy metering refresh cycle, PLC / DCS sampling cycle and actuator response time, the typical value range is [5~60]s, with a preferred value of 10s. 10s can usually take into account the metering update speed of electricity, gas and steam, and retain enough status observation points before the actuator is in place, avoiding excessively short cycles that amplify communication jitter.

[0084] S1.5: Construct time series samples based on effective historical control cycles, variable categories, energy accumulation states, and prediction time domain length, train an artificial intelligence time series energy consumption response model, and use independent calibration samples to form a prediction error envelope.

[0085] Specifically, it reads the set of effective historical control cycles, variable categories, set of energy accumulation state nodes, and prediction time domain length. The comprehensive energy consumption and process state variables are determined as the target variables for prediction, and the prediction time domain length is used as the basis for prediction. The data in the effective historical control cycle set are organized temporally; and according to the chronological order of the control cycles, the control variables, uncontrollable variables, process state variables, and comprehensive energy consumption of consecutive historical control cycles are combined into a historical state sequence; the historical state sequence and the data from the start of the target control cycle are then combined. The control variable sequence and uncontrollable variable sequence of each control cycle are combined to form the model input, and the same... The actual comprehensive energy consumption and process status data within each control cycle are used as supervision labels to form time-series samples. Equipment start-up and shutdown states are binary encoded, and operating modes are categorized. Continuous numerical variables are normalized according to the effective value range in the effective historical control cycle set. Time-series samples are mutually exclusively divided into model training samples, model validation samples, and independent calibration samples according to time order. An artificial intelligence time-series energy consumption response model is constructed using LSTM (Long Short-Term Memory network). The model training samples are used to train the artificial intelligence time-series energy consumption response model, and the model validation samples are used to compare the multi-cycle energy consumption prediction errors and process status prediction errors corresponding to different model parameters. The model parameters with the smallest prediction error are retained to obtain the artificial intelligence time-series energy consumption response model. The continuous numerical prediction results output by the model are inversely normalized to the original physical units of the corresponding prediction targets. Independent calibration samples are input into the artificial intelligence time-series energy consumption response model. For the actual and predicted values ​​of the prediction targets in the prediction control cycle in the independent calibration samples, the absolute prediction residuals are calculated. The absolute prediction residuals corresponding to the same prediction targets and the same prediction control cycle are sorted from smallest to largest, and the coverage probability is... The residual value corresponding to the sorting position is determined as the prediction error envelope width; the artificial intelligence time-series energy consumption response model, prediction error envelope, topological constraint delay control action diagram, and prediction time domain length are then used. This correlation serves as the basis for subsequent trajectory prediction, safety assessment of regulation, and verification of actual energy-saving effects.

[0086] The formula for calculating the prediction error envelope width is:

[0087] ;

[0088] ;

[0089] In the formula, It is the first The first independent calibration sample The predicted target is in the first The absolute prediction residual for each prediction control period and These correspond to the actual value and the predicted value, respectively. It is the prediction error envelope width. It is the number of independent calibration samples. It is the coverage probability. The corresponding empirical quantile.

[0090] It should be noted that the artificial intelligence time-series energy consumption response model includes a time-series feature extraction layer, a prediction condition fusion layer, a multi-step energy consumption prediction output layer, and a process state multi-step prediction output layer. The time-series feature extraction layer uses LSTM to extract the time-series change features of control variables, uncontrollable variables, process state variables, and comprehensive energy consumption in the historical state sequence. The prediction condition fusion layer combines the time-series change features with the continuous energy consumption after the start of the target control cycle. The control variable sequence and uncontrollable variable sequence of each control cycle are fused; the energy consumption multi-step prediction output layer outputs subsequent values ​​based on the fusion result. The predicted comprehensive energy consumption for each control cycle, and the multi-step prediction output layer of the process state outputs subsequent values ​​based on the fusion results. Predictive process status data for each control cycle;

[0091] The model parameters include the LSTM hidden state dimension, learning rate, and batch size, and the model parameters are determined by using the multi-cycle energy consumption prediction error and process state prediction error corresponding to the model validation samples.

[0092] The coverage probability is uniformly determined according to the safety level of the controlled process, with a typical value range of [0.90~0.99], preferably 0.95, and 95% can balance error coverage capability and candidate scheme feasibility; the same independent calibration sample does not participate in the model parameter training to avoid the prediction error envelope being too narrow due to the small residual of the training sample.

[0093] S2. Based on the topological constraint time delay control action diagram, determine the candidate control action variable group, form the current benchmark prediction trajectory, generate and screen candidate control trajectories, and obtain a set of executable energy consumption control instructions;

[0094] S2.1: Based on the comprehensive energy consumption state dataset, the current historical state sequence, the current uncontrollable variable sequence, the current operating context variable, and the current control variable combination are formed. According to the topological constraint time delay control action diagram, the intervenable control variables acting on the comprehensive energy consumption node in the prediction time domain are screened to form a set of candidate control action variable groups. The current baseline prediction trajectory is obtained by using the artificial intelligence time series energy consumption response model.

[0095] Specifically, the system reads the most recent consecutive effective control cycle data from the comprehensive energy consumption status dataset to form the current historical state sequence. When the number of recent consecutive effective control cycles is less than the historical input length specified by the artificial intelligence time-series energy consumption response model, an insufficient current state data flag is generated, and the candidate control scheme for this control cycle is not generated. The system also reads the prediction time domain length. The production task data, environmental disturbance data, equipment availability status, and production stage within the control period are correlated to form a sequence of currently uncontrollable variables. The current equipment operation mode, product type, production stage, and environmental disturbance conditions are read from the current historical state sequence and the current uncontrollable variable sequence to form the current operating context variables. The actual feedback values ​​of the actuators and the actual operating status of the equipment are read to form the current combination of control variables. The directed action paths between control variable nodes and comprehensive energy consumption nodes are retrieved from the topology constraint time-delay control action diagram. Control variables whose control action can be transmitted to the comprehensive energy consumption node within the prediction time domain and whose corresponding actuators are in an adjustable state are identified as interventionist control variables. Based on the shared relationships of actuators, upstream and downstream supply and demand relationships, and process interlocking relationships in the process control topology data, the interventionist control variables are combined. Control variables requiring linkage adjustment are grouped into the same candidate control action variable group, while control variables without control coupling relationships are formed into independent candidate control action variable groups, resulting in a set of candidate control variable groups. The current control variable combination is continuously applied within the prediction time domain. Each control cycle remains unchanged, forming a current hold control variable sequence. The current historical state sequence, the current hold control variable sequence, and the current uncontrollable variable sequence are input into the artificial intelligence time series energy consumption response model to obtain the current baseline prediction trajectory. When there are no controllable variables that can be intervened or no candidate control variable group has been formed, a no reliable intervention variable identifier is formed.

[0096] It should be noted that the current baseline prediction trajectory includes the trajectory of changes in the predicted energy consumption, predicted process state, and energy accumulation state for each control cycle within the prediction time domain. When environmental prediction data within the prediction time domain is not available, the environmental disturbance data is determined based on the trend of environmental disturbance changes in the current historical state sequence. If the trend cannot be determined, the most recent valid environmental disturbance data remains unchanged. When production task data, equipment availability status, or production stage data are missing, the most recent valid configuration data that is still within its applicable period continues to be used. If it cannot be confirmed that the configuration is still valid, an incomplete prediction condition flag is generated, and the candidate control scheme for this control cycle is not generated.

[0097] The historical input length specified for the artificial intelligence time-series energy response model is set using rolling time series cross-validation, with a typical value range of [value missing]. One control cycle, preferably one control cycle Each control cycle can cover the main operating states and energy accumulation processes before the prediction time domain, while avoiding the introduction of irrelevant data due to excessively long historical sequences.

[0098] S2.2: Generate candidate control combinations based on candidate control variable groups, current control variable combinations, operating context, process control topology, and actuator constraints. Supplement linkage control and exclude historically restricted combinations. Form a candidate control variable sequence according to execution order and response time to obtain a set of candidate control schemes.

[0099] Specifically, when no reliable intervention variable identifier is formed, the system reads records from the candidate control variable set, the current control variable combination, the current operating context variables, process control topology data, actuator control constraints, and the historical restricted control variable combination library. For each candidate control variable set in the candidate control variable set, candidate control values ​​are generated based on the corresponding actuator's allowable control range, single allowable adjustment amount, allowable adjustment direction, and control resolution. For equipment start-up / shutdown states and operating modes, candidate control states are generated based on the equipment's allowable state transition relationships. Furthermore, based on process interlocking relationships and upstream / downstream supply and demand relationships, necessary synchronized control values ​​or synchronized control states are supplemented to form an initial candidate control variable combination. The initial candidate control variable combination is then compared with historical restricted control variable combinations. The control variable combination records are matched. When the controlled object involved in the initial candidate control variable combination is consistent with the controlled object in a historical restricted control variable combination record, the corresponding candidate control value falls within the restricted value range of the record, or the corresponding candidate control state is consistent with the restricted control state of the record, and the current equipment operation mode, product type, production stage, and environmental disturbance conditions all fall within the applicable operating conditions range of the record, the corresponding initial candidate control variable combination is deleted. For the remaining initial candidate control variable combinations, the execution order of each control command is determined according to the process interlock relationship, and the expected arrival control cycle of the control variable is calculated according to the execution order and the allowable response time of the actuator. When the expected arrival control cycle of any control variable that must complete the linkage adjustment is greater than the predicted time domain length, the control variable combination is deleted. When the time limit is reached, the corresponding initial candidate control variable combination is deleted; for the remaining initial candidate control variable combinations, the current control value, candidate control value, or candidate control state is mapped to a continuous variable in the prediction time domain. The system operates on a control cycle basis. Before the expected control cycle of the actuator, the current control value or current control state is used. From the expected control cycle onwards, candidate control values ​​or candidate control states are used and maintained until the end of the prediction time domain, forming a sequence of candidate control variables. The candidate control variable group identifier, candidate control variable combination, candidate control variable sequence, instruction execution order, and expected control cycle of the actuator are associated to form a candidate control scheme. All candidate control schemes are then collected to form a set of candidate control schemes.

[0100] The formula for calculating the expected arrival control period is:

[0101] ;

[0102] In the formula, It is the first In the initial candidate control variable combination, the first The expected control period for each control variable. It is the expected start delay determined based on the instruction execution order. It is zero for parallel execution and determined based on the allowed response time of the preceding instruction for sequential execution. It is the first The allowable response time of the implementing agency. It is the length of a single control cycle.

[0103] S2.3: Input the candidate control variable sequence, the current historical state sequence, and the current uncontrollable variable sequence into the artificial intelligence time series energy consumption response model, and combine the prediction error envelope and control constraints to make process constraints, terminal state equivalence, and conservative energy saving judgments to form a set of feasible control schemes and a set of control schemes to be safely corrected.

[0104] Specifically, the process involves reading the candidate control scheme set, the AI ​​time-series energy consumption response model, the current historical state sequence, the current uncontrollable variable sequence, the current operating context variables, the current baseline prediction trajectory, the prediction error envelope, and the control constraint configuration data. For each candidate control scheme in the candidate control scheme set, the candidate control variable sequence, the current uncontrollable variable sequence, and the current historical state sequence are input into the AI ​​time-series energy consumption response model to obtain the candidate predicted energy consumption trajectory and the candidate predicted process state trajectory. Based on the candidate control variable group identifier corresponding to the candidate control scheme, the production unit and process state affected by the candidate regulation are determined. Combining the current equipment operating mode, product type, production stage, and environmental disturbance conditions, the process state constraint parameters and terminal state constraint parameters applicable to the current candidate control scheme are read from the control constraint configuration data to form an applicable control constraint configuration. When the necessary process state constraint parameters corresponding to the current candidate control scheme are missing, or the energy accumulation state participating in the terminal state equivalence judgment lacks a stable allowable deviation, an incomplete constraint configuration identifier is formed, and the corresponding candidate control scheme is deleted. For candidate control schemes with complete applicable control constraint configurations, robust process constraint judgments are performed. The evaluation process involves several steps: 1) Terminal state equivalence judgment and conservative energy saving judgment; 2) Associating candidate predicted energy consumption trajectories, candidate predicted process state trajectories, applicable control constraint configurations, robust process constraint judgment results, terminal state equivalence judgment results, conservative energy saving judgment results, lower limit of conservative energy saving in the predicted time domain, and information on corresponding failed items with the original candidate control schemes to form candidate control scheme evaluation results; 3) When the candidate control scheme evaluation results include robust process constraint passing results, terminal state equivalence passing results, and conservative energy saving passing results, the corresponding candidate control scheme is included in the feasible control scheme set; 4) When the candidate control scheme evaluation results include conservative energy saving passing results, but at least one robust process constraint judgment or terminal state equivalence judgment fails, and at least one numerical candidate control value in the corresponding candidate control variable combination still has adjustable space compared to the current control value, the corresponding candidate control scheme and its evaluation results are included in the set of control schemes to be safely corrected; 5) When the candidate control scheme evaluation results include conservative energy saving failing results, the corresponding candidate control scheme is deleted; 6) When robust process constraints or terminal state equivalence fail, and there is no adjustable numerical control value in the corresponding candidate control variable combination, the corresponding candidate control scheme is also deleted.

[0105] It should be noted that the process state constraint parameters include the constrained process state identifier, the applicable predictive control period, the upper and lower limits of the individual process state, the joint process constraint coefficient, and the joint process allowable boundary; the terminal state constraint parameters include the energy accumulation state identifier and the corresponding stability allowable deviation for participating in the terminal state equivalence judgment; the joint process constraint coefficient and the joint process allowable boundary are derived from the process card, equipment safety rules, product quality requirements, and upstream and downstream supply and demand rules, and are not calculated from the prediction error envelope.

[0106] Robust process constraint judgment:

[0107] For candidate control schemes, in the prediction time domain, the first... In the first control cycle, the predicted values ​​of each constrained process state specified by the applicable control constraint configuration are read from the candidate predicted process state trajectory to form a process state prediction vector; and the predicted values ​​of each constrained process state in the first control cycle are read from the prediction error envelope. The prediction error envelope width of the first control cycle is used to construct the error envelope matrix; the upper and lower limits of the process state, the joint process constraint coefficients, and the joint process allowable boundaries in the applicable control constraint configuration are used to construct the first control cycle error envelope matrix. The process constraint matrix and process allowable boundary vector are corresponding to each control cycle; for the upper limit constraint of the process state, the coefficient of the corresponding process state is set to 1, and the upper limit is used as the allowable boundary of the corresponding process; for the lower limit constraint of the process state, the coefficient of the corresponding process state is set to -1, and the negative of the lower limit is used as the allowable boundary of the corresponding process; for the joint process constraint, the coefficients of each corresponding process state are written according to the joint process constraint coefficients, and the joint process allowable boundary is used as the allowable boundary of the corresponding process; robust process constraint judgment is performed for each row of the process constraint matrix.

[0108] The error envelope matrix is:

[0109] ;

[0110] In the formula, It is the error envelope matrix. It is the first The constrained process state in the 1st... The prediction error envelope width for each predictive control cycle. It is the total number of constrained process states.

[0111] The constraints for robust processes are:

[0112] ;

[0113] In the formula, It is a process constraint matrix The OK, It is the transpose of a matrix or vector. It is a process state prediction vector. It is a process-allowed boundary vector The first in There are several boundary values, and each element in the vector represents the allowable boundary of the corresponding process. It is the most unfavorable expansion of the prediction error in the corresponding constraint direction. It is a process constraint matrix, where each row represents a single process state constraint or a joint process constraint.

[0114] Terminal state equivalence judgment:

[0115] When there is no energy accumulation state node in the action path corresponding to the current candidate control scheme, the terminal state equivalence pass result is directly formed; when there is an energy accumulation state node, the terminal state equivalence judgment is performed; according to the energy accumulation state identifier in the applicable control constraint configuration, the candidate terminal prediction value and the hold control terminal prediction value corresponding to the energy accumulation state at the end of the prediction time domain are read from the candidate prediction process state trajectory and the current baseline prediction trajectory, respectively; for the energy accumulation state, the terminal state equivalence judgment is performed according to the following formula:

[0116] ;

[0117] In the formula, It is the first The corresponding candidate control scheme is the first The final predicted value of the energy accumulation state. It is the terminal prediction value corresponding to the energy accumulation state in the current baseline prediction trajectory. It corresponds to the energy accumulation state in the th... The prediction error envelope width for each predictive control cycle. It is the allowable deviation for stability specified by the process control system.

[0118] When all energy accumulation states in the applicable control constraint configuration meet the terminal state equivalence conditions, a terminal state equivalence pass result is generated; when any energy accumulation state does not meet the terminal state equivalence conditions, a terminal state equivalence fail result is generated, and the corresponding energy accumulation state identifier and terminal state deviation are recorded.

[0119] Conservative energy conservation judgment:

[0120] The predicted energy consumption of the hold-up control for each control cycle within the prediction time domain is read from the current baseline prediction trajectory. The predicted energy consumption of the candidate control for the corresponding control cycle is read from the candidate predicted energy consumption trajectory. The prediction error envelope width of the comprehensive energy consumption in each prediction control cycle is read from the prediction error envelope. The conservative lower limit of the energy saving in the prediction time domain of the candidate control scheme is calculated.

[0121] ;

[0122] In the formula, It is the first The candidate control schemes will be discussed in the subsequent... The lower bound of the conservative time-domain energy limit for prediction within each control cycle. It is the [number]th [item] in the current baseline predicted trajectory. Energy consumption for maintaining control prediction for each control cycle This is the predicted total energy consumption for the corresponding candidate control scheme. The comprehensive energy consumption is in the first place The prediction error envelope width for each predictive control cycle.

[0123] When the lower limit of the conservative energy saving in the predicted time domain is greater than 0, the conservative energy saving result is considered passed; when the lower limit of the conservative energy saving in the predicted time domain is less than or equal to 0, the conservative energy saving result is considered failed.

[0124] S2.4: Keep the equipment start-up and shutdown status and operating mode in the control scheme to be corrected unchanged, generate a correction combination that rolls back to the current control value according to the control resolution, re-execute the process constraint, terminal state equivalence and conservative energy saving judgment, and select the combination with the smallest modification distance and all passes to form a set of safety correction control schemes.

[0125] Specifically, the system reads the set of control schemes to be corrected for safety. For each scheme, the system maintains the equipment start / stop status and operating mode unchanged. Based on the current control value, original candidate control value, and control resolution of each numerical control variable, the system determines the permissible level for each control variable to revert from the original candidate control value to the current control value. Correction control variable combinations are generated layer by layer according to the normalized modification distance, from smallest to largest. At each layer, combinations that do not meet the actuator's permissible range, permissible adjustment direction, single permissible adjustment amount, process interlocking relationship, upstream / downstream supply and demand relationship, or fall within the applicable range of the historically restricted control variable combination library are deleted. If a correction control variable combination in the current distance layer satisfies the subsequent three criteria, no further correction combinations with larger modification distances are generated. Following the candidate control variable sequence formation rules, the correction control variable combinations are converted into correction control variable sequences. The correction control variable sequences, the current uncontrollable variable sequences, and the current historical state sequences are then combined. Input the AI ​​time-series energy consumption response model to obtain the corrected predicted energy consumption trajectory and the corrected predicted process state trajectory. Re-evaluate these based on robust process constraint judgment, terminal state equivalence judgment, and conservative energy saving judgment. The combination of corrected control variables with the smallest normalized modification distance that simultaneously satisfies robust process constraints, terminal state equivalence conditions, and a lower limit of conservative energy saving in the predicted time domain greater than zero is determined as the safe corrected control variable combination. When there are two or more corrected combinations satisfying the conditions at the same normalized modification distance, the corrected combination with the larger lower limit of conservative energy saving in the predicted time domain is selected. The safe corrected control variable combination, the corrected control variable sequence, the corrected predicted energy consumption trajectory, the corrected predicted process state trajectory, and the corresponding lower limit of conservative energy saving in the predicted time domain are correlated to form a safe corrected control scheme. All safe corrected control schemes are then aggregated to form a safe corrected control scheme set. When none of the allowed corrected combinations satisfy the three judgment conditions, the corresponding control scheme to be safely corrected is deleted.

[0126] Original candidate control direction and corrected control value:

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, It is the original candidate control direction. It is a correction control value. It is the first The current control value of the control variable. This corresponds to the control resolution of the actuator. It is the control resolution level that moves along the original candidate control direction.

[0131] The formula for calculating the normalized modified distance is:

[0132] ;

[0133] In the formula, It is the normalized modification distance of the corrected combination of control variables relative to the original candidate combination of control variables. It is the set of numerical control variables in the corresponding candidate control variable group. and They are the first The upper and lower limits of the allowable control of a control variable.

[0134] S2.5: Select the feasible control scheme and the safety correction control scheme as alternative execution schemes, select the target control scheme according to the number of control variables, the conservative energy saving lower limit and the control adjustment range, and form an executable energy consumption control instruction set by combining historical optimization records, execution sequence and safety backoff rules.

[0135] Specifically, the system reads records and safety backoff rules from the feasible control scheme set, the safety correction control scheme set, and the historical preferred control variable combination library, and uses the control schemes from both sets as candidate execution control schemes. For feasible control schemes, the candidate control variable combinations, candidate control variable sequences, candidate predicted energy consumption trajectories, and candidate predicted process state trajectories are respectively used as the final control variable combinations, final control variable sequences, final predicted energy consumption trajectories, and final predicted process state trajectories. For safety correction control schemes, the safety correction control variable combinations, correction control variable sequences, correction predicted energy consumption trajectories, and correction predicted process state trajectories are respectively used as the corresponding final results. For each candidate execution control scheme, the system identifies the control variables that have actually changed relative to the current control variable combinations in the final control variable combination and counts the number of control variables participating in the adjustment. When a correction control value equals the current control value, the corresponding control variable is not counted in the number of control variables participating in the adjustment. The system then adjusts the control variables according to their participation in the adjustment process. The candidate execution control schemes are selected sequentially in the following order: from fewest to most control variables, from largest to smallest conservative energy saving limit in the prediction time domain, and from smallest to largest normalized control adjustment range. When all three selection criteria are the same, the candidate execution control schemes and current operating context variables are matched with records in the historical preferred control variable combination library, prioritizing candidate execution control schemes that match the control object, control direction, and applicable operating conditions. The finally selected candidate execution control scheme is determined as the target control scheme, and the final combination of control variables in the target control scheme is determined as the target control variable combination. According to the safety backoff rule, a safety backoff value is determined for each control variable in the target control variable combination. The control object identifier, target control value or target control state, instruction execution order, allowable response time of the actuator, safety backoff value, and final control variable sequence are associated to form an executable energy consumption control instruction set. When both the feasible control scheme set and the safety correction control scheme set are empty, an "no executable instruction" identifier is generated.

[0136] The formula for calculating the normalized control adjustment range is:

[0137] ;

[0138] In the formula, It is the normalized control adjustment magnitude of the alternative control scheme relative to the current combination of control variables. It corresponds to the final control value or final control state. It is a set of discrete control variables. It is a state change indicator; it is set to 1 if the condition is met, and 0 otherwise.

[0139] It should be noted that the safety backoff value should be the safety value configured in the process system. If no safety value is configured, the actual control value or actual control state before the control command is issued should be used.

[0140] S3. Execute the set of executable energy consumption control commands to form control execution feedback data, confirm the actual control status based on the control execution feedback data, and form the actual control execution result;

[0141] S3.1: Read the set of executable energy consumption control instructions and save the state data before control. Send the control instructions to the corresponding controllers according to the order of instruction execution. Based on the execution results and the actual combination of control variables, form control execution feedback data.

[0142] Specifically, when an executable energy consumption control instruction set exists, the current historical state sequence, current control variable combination, current equipment operating status, and target control variable combination are read and saved before the control instructions are issued, forming pre-control state data. According to the instruction execution order in the executable energy consumption control instruction set, each control instruction is issued to the corresponding PLC, DCS, frequency converter, valve controller, temperature controller, or equipment controller. For control instructions constrained by upstream and downstream supply and demand relationships, start-stop interlocking relationships, or other sequential dependencies, subsequent instructions are issued after the preceding instructions have been executed. When a sequential instruction fails to execute, the issuance of associated subsequent instructions is stopped and a safety rollback is implemented for already executed instructions. Control instructions without control coupling relationships are issued in parallel. After the control instructions are issued, the instruction reception status, actual feedback value, actual operating status, and fault status returned by each actuator are read. For numerical control instructions, the absolute deviation of control execution is calculated. When a control instruction has been received and the absolute deviation of the numerical control instruction is less than or equal to the allowable execution error of the corresponding actuator, the corresponding numerical control instruction is marked as completed. For equipment start / stop status or operating mode switching... When the actual feedback state matches the target control state, the corresponding control instruction is marked as executed successfully. If the control instruction is not received, the actuator fails to reach the target control value or target control state within the allowed response time, equipment malfunctions, or a safety interlock is triggered, the corresponding control instruction is marked as executed unsuccessfully, and a safety backoff is triggered according to the safety backoff value in the set of executable energy consumption control instructions. The actual values ​​or actual states of all control variables are determined based on the execution results of each control instruction. For control variables that have been executed successfully, the actual feedback value or actual feedback state of the actuator is used. For control variables that did not participate in this control, the actual control value or actual operating state before control is used. For control variables that have failed to execute and have completed a safety backoff, the actual feedback value or actual feedback state after the backoff is used. All control variables are arranged according to the order specified by the artificial intelligence time-series energy consumption response model to form a combination of actual executed control variables. The combination of actual executed control variables, the issuance time of each control instruction, the actual arrival time, the execution completion flag or execution failure flag, the actual feedback value of the actuator, the equipment malfunction state, and the safety backoff state are associated to form control execution feedback data.

[0143] The formula for calculating the absolute deviation of control execution is:

[0144] ;

[0145] In the formula, It is the first The absolute deviation of control execution for each numerical control command This is the actual feedback value from the implementing agency. It is the target control value.

[0146] S3.2: Based on the feedback data of regulation execution, determine whether an effective control effect has been formed. When the combination of actual execution control variables is inconsistent with the combination of target control variables, form a sequence of actual execution control variables, and use the artificial intelligence time series energy consumption response model to re-determine process constraints, terminal state equivalence and conservative energy saving. Confirm the stability of the actual control state that meets the judgment conditions, and form the actual regulation execution result.

[0147] Specifically, the system reads the control execution feedback data to obtain the actual combination of control variables, the completion or failure flags of each control command, the equipment fault status, and the safety rollback status. When all control commands fail to execute, or all controlled variables have undergone safety rollback, the system determines that the control action has not achieved effective control and records the corresponding reason for failure. When at least one control command completes execution, the system reads the current process status data after the command execution. If any process status exceeds the corresponding process allowable range, product quality requirements, or equipment safe operating range, the system marks the control action as failed and triggers safety rollback. When the current process status meets the corresponding requirements, the system combines the actual combination of control variables with the target combination of control variables. For numerical control variables, when the deviation between the actual feedback value and the target control value is no greater than the allowable execution error, and the actual state of the non-numerical control variables is consistent with the target control state, the stability of the actual execution control variable combination is directly confirmed. When the actual execution control variable combination is inconsistent with the target control variable combination due to the failure of some control commands, safety rollback, or execution deviation, the actual execution control variable sequence is constructed according to the command issuance time, actual arrival time, actual feedback value, and safety rollback status in the control execution feedback data, following the control variable sequence formation rules. Simultaneously, the latest production task data, environmental disturbance data, equipment availability status, and production stage are read to update the current uncontrollable variable sequence. The execution control variable sequence, the updated current uncontrollable variable sequence, and the current historical state sequence are input into the artificial intelligence time-series energy consumption response model to obtain the actual execution predicted energy consumption trajectory and the actual execution predicted process state trajectory. These are then re-evaluated according to robust process constraint judgment, terminal state equivalence judgment, and conservative energy-saving judgment. If the actual execution control variable sequence simultaneously meets all three judgment conditions, stability confirmation continues; otherwise, the current control is marked as an execution failure and a safety backoff is triggered. After the allowable response time of the actuator ends, the control variable feedback value, equipment operating status, and fault status are continuously read according to the stability confirmation duration specified in the equipment control program. When the numerical control variable remains within the corresponding allowable execution error range, the non-numerical control variable... When the actual state of the control variables remains unchanged and the equipment does not malfunction or trigger a safety interlock, the actual combination of executed control variables is determined as the actual control state for this regulation, and the time of completion of stability confirmation is determined as the start time of regulation effect verification. When the actual combination of executed control variables exceeds the allowable execution error during the stability confirmation period, the actual control state changes, or the equipment malfunctions or triggers a safety interlock, this regulation is marked as unstable and a safety backoff is triggered. The formation of effective control action, the sequence of actual executed control variables, the actual control state, the start time of regulation effect verification, the execution failure flag, the execution instability flag, the reason for execution failure, the equipment malfunction state, and the safety backoff state are correlated to form the actual regulation execution result.

[0148] It should be noted that the stable confirmation time specified in the equipment control program is set by historical statistics of feedback fluctuations and misjudgments under different confirmation cycles. The typical value range is [2~5] control cycles, with 3 control cycles being preferred. Three consecutive cycles can eliminate accidental arrival and feedback jitter in a single cycle, and will not significantly delay the effect verification.

[0149] S4. Based on the actual control execution results, verify the multi-cycle energy saving effect and rebound energy consumption, form a closed-loop control evaluation result, revise the prediction and control basis according to the closed-loop control evaluation result, and generate a comprehensive intelligent energy consumption control result.

[0150] S4.1: Based on the actual control execution results, collect verification data in the prediction time domain for the control process that forms an effective control effect and a stable control state, and form post-control verification time domain state data, actual uncontrollable variable sequence and updated and maintained control variable sequence.

[0151] Specifically, the actual control execution results are read. If the control fails to achieve effective control, fails, or is unstable, energy-saving effect verification is not performed. Instead, the corresponding reasons for execution failure, equipment fault status, and safety fallback status are correlated to form a closed-loop control evaluation result. If the control achieves effective control and the actual control status is stable, data is continuously collected in the prediction time domain starting from the start time of control effect verification. The data collected includes energy metering data, equipment operation data, production task data, environmental disturbance data, process status data, and actuator feedback data for each control cycle. During the verification period, no new energy-saving controls are implemented on the control variables involved in the actual control state or the energy accumulation state on the corresponding action path. If new controls must be implemented due to production or safety requirements, the verification is terminated, the collected data is marked as incomplete verification data, the actual cumulative net energy saving and the equivalent results of the actual terminal state are not calculated, and the reason for verification termination, the number of completed verification cycles, and the safety rollback status are correlated to form an incomplete closed-loop control assessment result. This verification process concludes. Data from the entire verification period undergoes time alignment, anomaly identification, and energy unit conversion. The actual comprehensive energy consumption, actual equipment operation data, actual production task data, actual environmental disturbance data, actual process status data, and actual control variable combinations for each control cycle are correlated chronologically to form post-control verification time-domain state data. The actual production task data, actual environmental disturbance data, actual equipment availability, and actual production stage from the post-control verification time-domain state data are read and correlated according to the control cycle to form a sequence of actual uncontrollable variables. The pre-control control variable combinations are continuously linked within the verification time domain. Each control cycle remains unchanged, forming an updated sequence of control variables.

[0152] S4.2: Input the historical state sequence before regulation, the updated and maintained control variable sequence, and the actual uncontrollable variable sequence into the artificial intelligence time series energy consumption response model to obtain the updated and maintained control prediction trajectory. Combine the actual cumulative net energy saving, the cumulative energy consumption prediction error envelope, and the compliance of the actual terminal state and process state to form a closed-loop regulation evaluation result.

[0153] Specifically, the historical state sequence before regulation, the updated and maintained control variable sequence, and the actual uncontrollable variable sequence are input into the artificial intelligence time-series energy consumption response model to obtain the updated and maintained control prediction trajectory. Based on the updated and maintained control prediction trajectory and the post-regulation verification time-domain state data, the actual cumulative net energy saving is calculated, and the energy consumption prediction error envelope width of each control cycle is summed to obtain the cumulative energy consumption prediction error envelope. For each energy accumulation state, the actual terminal value is compared with the terminal prediction value in the updated and maintained control prediction trajectory. During the verification period, the actual process state of each control cycle is compared with the corresponding applicable control constraints. When any actual process state does not meet the corresponding process constraint, this regulation is marked as a process constraint failure, and the actual control variable combination is marked as an unusable control variable combination. When the process state continuously meets the requirements and all energy accumulation states meet the requirements... When the actual terminal state is equivalent to the condition and the actual cumulative net energy saving is greater than the cumulative energy consumption prediction error envelope, the current regulation is marked as a true and effective energy-saving regulation. When the process state continuously meets the requirements, but any energy accumulation state does not meet the actual terminal state equivalence condition, the current regulation is marked as energy consumption transfer or rebound compensation, and the actual control variable combination is marked as a rebound control variable combination. When both the process state and the actual terminal state meet the requirements, but the actual cumulative net energy saving is not greater than the cumulative energy consumption prediction error envelope, the current regulation is marked as energy saving effect not achieved. The verification integrity, actual cumulative net energy saving, cumulative energy consumption prediction error envelope, actual terminal state equivalence result, process state compliance, regulation assessment label, rebound control variable combination, unavailable control variable combination, and safe retreat state are correlated to form a closed-loop regulation assessment result.

[0154] The formula for calculating actual cumulative net energy savings is:

[0155] ;

[0156] In the formula, This regulation is in Actual cumulative net energy savings within each control cycle The prediction was obtained using actual production task data and actual environmental disturbance data. Each control cycle maintains the predicted energy consumption for control. This is the [number]th ...event] after the implementation of this round of regulation. Actual comprehensive energy consumption per control cycle.

[0157] The formula for calculating the cumulative energy consumption prediction error envelope is:

[0158] ;

[0159] In the formula, yes The cumulative energy consumption prediction error envelope for each control cycle The comprehensive energy consumption is in the first place The prediction error envelope width for each predictive control cycle.

[0160] The actual terminal state equivalence condition is:

[0161] ;

[0162] In the formula, It is the first The actual terminal value of the energy accumulation state, It is the terminal forecast value when the control variables are maintained before the adjustment. This corresponds to the width of the prediction error envelope. It is the allowable deviation for stability specified by the process control system.

[0163] S4.3: Based on the closed-loop control evaluation results, select control processes with stable actual control states, no equipment faults, and complete verification data to form new control response samples, update the artificial intelligence time-series energy consumption response model, prediction error envelope, and topology constraint time delay control action diagram, and write effective, rebound, or unavailable control combinations into the corresponding historical record database to generate comprehensive intelligent energy consumption control results.

[0164] Specifically, the closed-loop control evaluation results are read, and control processes that demonstrate effective control, stable actual control state, valid energy metering data, no equipment failures, and complete post-control verification time-domain state data are selected. The pre-control historical state sequence, actual executed control variable sequence, and actual uncontrollable variable sequence corresponding to the selected control process are used as model inputs. The actual comprehensive energy consumption and actual process state data from the post-control verification time-domain state data are used as supervision labels to form new control response samples. According to the sample formation time and a preset fixed ratio, the new control response samples are mutually exclusively divided into model update samples and independent calibration samples. When the number of new control response samples is insufficient to simultaneously form model update samples and independent calibration samples, only the new control response samples are accumulated, and the current model and prediction error envelope update are not performed. The model update samples are used to incrementally train the artificial intelligence time-series energy consumption response model to obtain candidate update models. The prediction errors of the candidate update models and the current model are calculated using model verification samples. When the prediction error of the candidate update model is not greater than that of the current model, the candidate update model is adopted; otherwise, the current model is retained. The finally adopted artificial intelligence time-series energy consumption response model is used to process the independent calibration. The sample is used to recalculate the absolute prediction residuals of each prediction target in each prediction control cycle, and update the corresponding prediction error envelope width according to the original coverage probability. Based on the actual response records of each downstream process state after the actual execution control variable changes, the topology constraint time delay control action diagram is corrected. The actual execution control variable combination corresponding to the real and effective energy-saving regulation, as well as the applicable equipment operation mode, product type, production stage and environmental disturbance conditions, are written into the historical preferred control variable combination library. The rebound control variable combination, unavailable control variable combination and applicable conditions are written into the historical restricted control variable combination library, which are used for sorting and excluding subsequent candidate control variable combinations. For control processes that form no reliable intervention variable identifier, no executable instruction identifier, no effective control action, or incomplete verification data, they are not used for updating the artificial intelligence time series energy consumption response model, prediction error envelope and topology constraint time delay control action diagram. Only the corresponding reasons for non-regulation, execution failure or verification termination are recorded. The target control scheme, actual regulation execution results, closed-loop regulation evaluation results and the updated topology constraint time delay control action diagram, artificial intelligence time series energy consumption response model and prediction error envelope are correlated to generate comprehensive intelligent energy consumption regulation results.

[0165] The modified topology constraint delay control action diagram is as follows:

[0166] When the actual response direction and response sequence conform to the equipment connection relationship, energy transmission direction and original control action relationship, the corresponding actual response time is added to the directed edge response record, the median value of the cumulative effective response time is calculated, and the median value is divided by the length of a single control cycle and rounded up to obtain the updated number of lag cycles; when a directed edge does not generate a corresponding downstream state response during the cumulative effective control process, and adding the corresponding source node lag data cannot reduce the target node prediction error, the corresponding directed edge is deleted.

[0167] Reference Figure 2 This embodiment also provides an artificial intelligence-based integrated intelligent energy consumption control system, including:

[0168] The multi-source data construction module is used to collect and verify multi-source energy consumption data, convert them in a unified manner, and form comprehensive state, process topology, and control constraint data.

[0169] The time delay effect modeling module is used to classify variables, identify time delay control effects, determine the prediction time domain, and train an artificial intelligence energy consumption response model.

[0170] The intelligent control and decision-making module is used to screen controllable variables, generate candidate solutions, and complete safety, terminal, and energy-saving evaluations, as well as correction and optimization.

[0171] The instruction execution verification module is used to issue target control instructions, collect execution feedback, and verify the actual control status, execution deviation, and operational stability.

[0172] The closed-loop evaluation and update module is used to verify the actual energy-saving effect, process constraints and terminal status, and update the model, action diagram and historical control record library.

[0173] In summary, this invention establishes a reliable control data foundation through: multi-source energy consumption data alignment, anomaly screening, and unified conversion; constructs a time-series energy consumption response model and prediction error envelope through conditional time delay effect identification, energy accumulation state analysis, and prediction time domain determination, thereby improving the accuracy of energy consumption and process state prediction; reduces the risks of state out-of-bounds and superficial energy saving through robust constraints on candidate schemes, terminal state equivalence, conservative energy-saving judgment, and minimum amplitude safety correction; and improves the safety, true energy saving, and continuous adaptability of the control scheme through execution feedback, multi-cycle verification, and model closed-loop updates.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A comprehensive intelligent energy consumption control method based on artificial intelligence, characterized in that, include: Acquire and process energy consumption regulation data to form a comprehensive energy consumption status dataset, process control topology data, and control constraint configuration data. Identify the time delay relationship verified by conditional independence and error improvement through historical control cycles and operating contexts, determine the prediction time domain, and construct a topology constraint time delay control action diagram, an artificial intelligence time series energy consumption response model, and a prediction error envelope. Based on the topological constraint time delay control action diagram, candidate control action variable groups are determined, and the current benchmark prediction trajectory is formed. Candidate control trajectories are generated and screened to obtain a set of executable energy consumption control instructions. Execute a set of executable energy consumption control commands to generate control execution feedback data, confirm the actual control status based on the control execution feedback data, and generate actual control execution results; Based on the actual control and regulation results, the multi-cycle energy saving effect and rebound energy consumption are verified, and a closed-loop control and regulation evaluation result is formed. Based on the closed-loop control and regulation evaluation result, the prediction and control basis is revised, and a comprehensive intelligent control result for energy consumption is generated.

2. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 1, characterized in that, The process of acquiring and processing energy consumption control data to form a comprehensive energy consumption status dataset, process control topology data, and control constraint configuration data refers to collecting energy metering, equipment operation, production tasks, environmental disturbances, process status, and actuator feedback data according to the control cycle, performing time alignment, anomaly marking, and energy unit conversion on various types of data, calculating the comprehensive energy consumption, and associating it with equipment connections, energy transmission, control logic, and process constraints to form a comprehensive energy consumption status dataset, process control topology data, and control constraint configuration data.

3. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 2, characterized in that, The construction of the topological constraint delay control action diagram, the artificial intelligence time-series energy consumption response model, and the prediction error envelope includes: Based on the comprehensive energy consumption status dataset, effective historical control cycles are screened, and control variables, uncontrollable variables, equipment status variables, process status variables, and operating context variables are determined according to the data source and its corresponding control and acquisition objects. Initial directed edges are established based on process control topology and variable categories. Directed edges are then selected by combining the effect of conditional time delay and prediction error improvement, and the lag period is determined to form a topological constraint time delay control effect graph. Based on the topological constraint time delay control action diagram, the action path from the control variable to the comprehensive energy consumption is extracted, the energy accumulation state and the maximum historical recovery time are identified, and the prediction time domain length is determined by combining the path accumulation time delay and the maximum response time of the actuator. Time series samples are constructed based on effective historical control cycles, variable categories, energy accumulation states, and prediction time domain lengths. An artificial intelligence time series energy consumption response model is trained, and an independent calibration sample is used to form a prediction error envelope.

4. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 3, characterized in that, The process of determining candidate control variable groups based on topological constraint time-delay control action diagram and forming the current baseline prediction trajectory refers to forming a combination of current historical state sequence, current uncontrollable variable sequence, current operating context variable and current control variable based on the comprehensive energy consumption state dataset, screening the intervenable control variables acting on the comprehensive energy consumption node in the prediction time domain according to the topological constraint time-delay control action diagram, forming a set of candidate control variable groups, and obtaining the current baseline prediction trajectory using an artificial intelligence time-series energy consumption response model.

5. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 4, characterized in that, The process of generating and filtering candidate control trajectories to obtain a set of executable energy consumption control instructions includes: Based on the candidate control variable group, the current control variable combination, the operating context, the process control topology, and the actuator constraints, candidate control combinations are generated, linkage control is supplemented and historically restricted combinations are excluded, and a candidate control variable sequence is formed according to the execution order and response time to obtain a set of candidate control schemes. The candidate control variable sequence, the current historical state sequence, and the current uncontrollable variable sequence are input into the artificial intelligence time series energy consumption response model. The process constraints, terminal state equivalence, and conservative energy saving judgments are made by combining the prediction error envelope and control constraints to form a set of feasible control schemes and a set of control schemes to be safely corrected. Keeping the equipment start-stop status and operating mode in the control scheme to be corrected unchanged, generating a correction combination that rolls back to the current control value according to the control resolution, re-executing the process constraint, terminal state equivalence and conservative energy saving judgment, and selecting the combination with the smallest modification distance and all passing to form a set of safety correction control schemes; Feasible control schemes and safety correction control schemes are used as alternative execution schemes. The target control scheme is selected according to the number of control variables, the conservative energy saving lower limit, and the control adjustment range. The set of executable energy consumption control instructions is formed by combining historical optimization records, execution order, and safety backoff rules.

6. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 5, characterized in that, The process of executing a set of executable energy consumption control instructions to form control execution feedback data involves reading the set of executable energy consumption control instructions and saving the state data before control, issuing control instructions to the corresponding controllers according to the instruction execution order, and forming control execution feedback data based on the execution results and the actual combination of control variables.

7. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 6, characterized in that, The process of confirming the actual control state based on the feedback data of the control execution and forming the actual control execution result refers to judging whether an effective control effect has been formed based on the feedback data of the control execution. When the combination of actual control variables is inconsistent with the combination of target control variables, an actual control variable sequence is formed, and the process constraints, terminal state equivalence and conservative energy saving judgments are re-examined using an artificial intelligence time series energy consumption response model. The stability of the actual control state that meets the judgment conditions is confirmed, and the actual control execution result is formed.

8. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 7, characterized in that, The verification of multi-cycle energy-saving effects and rebound energy consumption based on actual control execution results forms a closed-loop control evaluation result, including: Based on the actual control execution results, verification data in the prediction time domain is collected for the control process that forms an effective control effect and a stable control state, forming post-control verification time domain state data, actual uncontrollable variable sequence and updated and maintained control variable sequence; The historical state sequence before regulation, the updated and maintained control variable sequence, and the actual uncontrollable variable sequence are input into the artificial intelligence time series energy consumption response model to obtain the updated and maintained control prediction trajectory. Combined with the actual cumulative net energy saving, the cumulative energy consumption prediction error envelope, and the compliance of the actual terminal state and process state, a closed-loop regulation evaluation result is formed.

9. The comprehensive intelligent energy consumption control method based on artificial intelligence as described in claim 8, characterized in that, The process of revising prediction and control criteria based on closed-loop control evaluation results and generating comprehensive intelligent energy consumption control results refers to selecting control processes with stable actual control states, fault-free equipment, and complete verification data based on closed-loop control evaluation results, forming new control response samples, updating the artificial intelligence time-series energy consumption response model, prediction error envelope, and topological constraint time delay control action diagram, and writing effective, rebound, or unavailable control combinations into the corresponding historical record database to generate comprehensive intelligent energy consumption control results.

10. An intelligent energy consumption control system based on artificial intelligence, based on the intelligent energy consumption control method based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, include: The multi-source data construction module is used to collect and verify multi-source energy consumption data, convert them in a unified manner, and form comprehensive state, process topology, and control constraint data. The time delay effect modeling module is used to classify variables, identify time delay control effects, determine the prediction time domain, and train an artificial intelligence energy consumption response model. The intelligent control and decision-making module is used to screen controllable variables, generate candidate solutions, and complete safety, terminal, and energy-saving evaluations, as well as correction and optimization. The instruction execution verification module is used to issue target control instructions, collect execution feedback, and verify the actual control status, execution deviation, and operational stability. The closed-loop evaluation and update module is used to verify the actual energy-saving effect, process constraints and terminal status, and update the model, action diagram and historical control record library.

Images