Fine energy consumption modeling and real-time optimization control method based on process parameters

Abstract

The application proposes a fine energy consumption modeling and real-time optimization control method based on process parameters. The method classifies the field points into controllable process parameters, necessary state parameters, energy consumption data and constraint indicators, organizes a sample set according to the process time slice and records the corresponding relationship of the time slice; trains a controllable energy consumption response artificial intelligence model containing a double-input channel, a state modulation layer and a double-output branch; inputs the current working condition into the model for response detection, generates a candidate parameter adjustment action set and a predicted energy-saving response; evaluates the candidate action through virtual execution, and sequentially filters through hard constraints, energy-saving effectiveness, boundary margin and low disturbance to determine the target control action and issue it. The application effectively distinguishes controllable energy consumption response and state energy consumption background, combines field engineering constraints to ensure that the control instruction is safe and executable, and realizes safe and reliable conversion of fine energy consumption modeling to real-time optimization control.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and intelligent manufacturing technology, and in particular relates to a method for refined energy consumption modeling and real-time optimization control based on process parameters. Background Technology

[0002] In industrial production scenarios such as process manufacturing, discrete manufacturing, and batch processing, energy consumption is typically closely related to controllable process parameters such as temperature setpoints, pressure setpoints, pump frequency setpoints, fan frequency setpoints, valve opening commands, and heating power setpoints. It is also influenced by actual operating conditions such as actual temperature, actual pressure, actual flow rate, motor current, equipment operating status, process cycle time, quality boundaries, and equipment safety boundaries. To reduce energy consumption per unit product, existing technologies usually collect production process data from systems such as energy management systems, PLCs, DCSs, SCADA, and MES, and use statistical analysis, empirical rules, regression models, neural network models, or digital twin models to establish energy consumption prediction relationships. Based on the prediction results, energy-saving suggestions or control setpoint adjustment schemes are then provided. While these methods can reflect energy consumption trends in the production process to some extent, they still have significant shortcomings in practical applications. On the one hand, data collected in industrial settings has distinct roles. Some data points represent setpoints or execution quantities that can be written into the control system, while others are readable sensor feedback values. Still others represent energy consumption measurement results or process constraint boundaries. Existing methods often directly input this data into the model as ordinary features, making it easy for the model to learn the correlation between total energy consumption and field data, without clearly distinguishing which energy consumption changes are caused by controllable process parameters and which are caused by the current operating state or equipment load background. On the other hand, energy consumption data is usually aggregated and measured by equipment, process segment, or production line, while adjustable objects are scattered in different control loops. Furthermore, the energy consumption response after changes in controllable parameters often has a time lag. For example, after changes in heating power, fan frequency, or pump frequency, the stable energy consumption result only becomes apparent in subsequent time slices. Existing modeling methods that focus on the same moment cannot accurately learn the response relationship between "parameter action time slices and subsequent stable energy consumption results." Furthermore, existing energy consumption optimization methods typically provide control suggestions directly based on model predictions or optimization results, without fully considering engineering conditions such as the set granularity of the field control system, maximum single adjustment range, writing cycle, parameter upper and lower limits, quality qualification boundaries, equipment safety boundaries, and production cycle lower limits. This can lead to suggestions that may not be written into the field control system, or, while theoretically achieving energy-saving effects, may result in insufficient quality margin, close equipment boundaries, reduced cycle time, or excessive control disturbances. Therefore, the main deficiency of existing technologies is not merely insufficient energy consumption prediction accuracy, but rather the lack of a modeling and control method that integrates point roles, time-slice response relationships, controllable energy consumption response modeling, candidate action generation, and virtual execution screening, centered on controllable process parameters. This would enable refined energy consumption modeling results to be truly transformed into executable real-time optimized control actions. Summary of the Invention

[0003] This invention discloses a refined energy consumption modeling and real-time optimization control method based on process parameters to solve the above-mentioned problems in the prior art.

[0004] To achieve the above objectives, this invention provides a method for refined energy consumption modeling and real-time optimization control based on process parameters, the method comprising: Acquire field location data, categorize the field location data into controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators, and organize them into a process parameter sample set according to process time slices. The process parameter sample set records the correspondence between the parameter action time slice of the controllable process parameter and the subsequent stable energy consumption result time slice. A controllable energy consumption response artificial intelligence model is trained based on the process parameter sample set. The controllable energy consumption response artificial intelligence model includes a controllable process parameter input channel, a necessary state parameter input channel, a state modulation layer, a controllable energy consumption response branch, and a state energy consumption background branch. The current operating condition data is input into the controllable energy consumption response artificial intelligence model for response detection, and a candidate parameter tuning action set is generated. The candidate parameter tuning action set includes at least one candidate parameter tuning action and the predicted energy-saving response of the at least one candidate parameter tuning action. A virtual execution evaluation is performed on at least one candidate parameter tuning action in the candidate parameter tuning action set to obtain an evaluation result; based on the evaluation result, the at least one candidate parameter tuning action is screened to determine the target control action, and the target control action is sent to the field control system.

[0005] Furthermore, the classification of the field location data into controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators includes: The set value or execution instruction with write permission to the control system is marked as the controllable process parameter; The sensor feedback or device status word that is in the same device or the same control loop as the controllable process parameter is marked as the necessary status parameter. The energy metering results of the equipment whose metering range covers the controllable process parameters are marked as the energy consumption data; The process configuration values ​​or limit table data used to limit the controllable process parameters or the target control actions are marked as the constraint indicators.

[0006] Furthermore, the process parameter sample set records the correspondence between the time slices of the controllable process parameters and the time slices of the subsequent stable energy consumption results, including: When the controllable process parameter changes stably within the time slice of the parameter action, the energy consumption result in the subsequent time slice is selected as the energy consumption response result corresponding to the subsequent stable energy consumption result time slice according to the process stabilization cycle. When energy consumption response occurs in subsequent time slices after the same controllable process parameter changes, the average energy consumption of the first time slice after entering the stable range or the stable time slice is selected as the energy consumption response result.

[0007] Furthermore, the controllable energy consumption response artificial intelligence model also includes a linear mapping layer and a time-series modeling layer; The controllable process parameter input channel and the necessary state parameter input channel are respectively connected to the linear mapping layer and the time series modeling layer to output controllable features and state features respectively; The state modulation layer inputs the state features into the fully connected layer and processes them through an activation function to obtain modulation coefficients. The modulation coefficients are then applied element by element to the controllable features to obtain the state-modulated controllable response features. The controllable energy consumption response branch receives the controllable response features and outputs the controllable energy consumption response results, while the state energy consumption background branch receives the state features and outputs the energy consumption background results.

[0008] Furthermore, the joint loss used to train the controllable energy consumption response artificial intelligence model includes a synthetic energy consumption error term, a controllable response consistency term, and a static stability suppression term; The synthesized energy consumption error term is the error between the sum of the controllable energy consumption response result and the energy consumption background result and the energy consumption data; The controllable response consistency term is the error between the output difference of the controllable energy consumption response branch before and after parameter change and the actual energy consumption change result. The static stability suppression term is a penalty term that constrains the fluctuation of the controllable energy consumption response branch when the controllable process parameters remain stable.

[0009] Furthermore, the step of inputting current operating condition data into the controllable energy consumption response artificial intelligence model for response detection and generating a candidate parameter tuning action set includes: Based on the write granularity and parameter upper and lower limits allowed by the control system, candidate target values ​​are generated for the controllable process parameters, and the candidate parameter tuning action is constructed, which includes the action object, the current value, and the candidate target value. The current operating condition data is input into the controllable energy consumption response artificial intelligence model to obtain the baseline controllable energy consumption response result; Replace the values ​​of the controllable process parameters in the current operating condition data with the values ​​of the candidate targets, while keeping the necessary state parameters in the current operating condition data unchanged, and input them into the controllable energy consumption response artificial intelligence model to obtain the controllable energy consumption response results after the action; The predicted energy-saving response is calculated based on the difference between the baseline controllable energy consumption response and the controllable energy consumption response after the action. Based on the constraint index, the candidate parameter tuning actions are pruned to retain those that meet the writing conditions, forming the candidate parameter tuning action set.

[0010] Further, the virtual execution evaluation of at least one candidate parameter tuning action in the candidate parameter tuning action set includes: For each candidate parameter tuning action in the candidate parameter tuning action set, replace the value of the corresponding controllable process parameter in the current operating condition data with the target value of the candidate parameter tuning action to construct a virtual operating condition. The virtual operating condition is input into the controllable energy consumption response artificial intelligence model to obtain the controllable energy consumption response result after the action of the candidate parameter tuning action and the state energy consumption background result after the action. Based on the controllable energy consumption response results after the action and the state energy consumption background results after the action, the effectiveness of the candidate parameter tuning action in the current operating context is evaluated.

[0011] Further, based on the evaluation results, the step of screening the at least one candidate parameter tuning action to determine the target control action includes: Hard constraint screening is performed on the at least one candidate parameter tuning action to eliminate candidate parameter tuning actions that do not meet the upper and lower limits of parameters or correlation constraint indicators; The candidate parameter tuning actions retained after hard constraint screening are screened for energy-saving effectiveness, and candidate parameter tuning actions whose predicted energy-saving response is lower than the minimum effective predicted energy-saving response are eliminated. After screening for energy-saving effectiveness, the remaining candidate parameter tuning actions are screened for boundary margins, and candidate parameter tuning actions with mass margin, safety margin, or cycle time margin lower than the corresponding minimum margin requirement are eliminated. After screening the boundary margin, the candidate parameter tuning actions are selected with low disturbance, and the candidate parameter tuning action with the smallest adjustment amplitude or whose action direction is consistent with the previous control cycle is determined as the target control action.

[0012] Furthermore, the step of sending the target control action to the field control system includes: The target control action is converted into a control instruction, and the control instruction includes the action object, target value, write address and write time of the target control action; Write the control command into the corresponding register or tag number of the field control system; Read the actuator feedback value. When the actuator feedback value reaches the target value or enters the allowable deviation range within the allowable confirmation time, confirm that the target control action has been executed. When writing fails or the actuator feedback value deviates from the target value, the execution record of the target control action is canceled and the control settings before writing are restored.

[0013] Furthermore, each sample record in the process parameter sample set contains the controllable process parameters, the necessary state parameters, the energy consumption data, and the constraint indicators within the same process time slice; When the controllable process parameter changes during the process time slice, the control system confirms the effect and maintains a stable set value as the controllable process parameter in the sample record. When there is a response lag in the process, the sample record saves the lag window and the source of values ​​for the time slice of the controllable process parameters and the time slice of the subsequent stable energy consumption results.

[0014] The beneficial technical effects of the present invention are at least as follows: To address the aforementioned issues, this invention provides a refined energy consumption modeling and real-time optimized control method based on process parameters. By classifying field points by role and saving the correspondence between parameter action time slices and subsequent stable energy consumption results, the interference of read-only state variables on energy consumption modeling is effectively eliminated, enabling the model to accurately learn the real energy consumption response after changes in controllable parameters. Secondly, the constructed dual-channel timing model and state modulation layer can clearly distinguish between the energy consumption response caused by changes in controllable parameters and the energy consumption background caused by equipment operating status, significantly improving the engineering interpretability of the model output. Finally, in the control execution phase, considering the writing granularity and engineering constraints such as quality, safety, and cycle time of the field control system, the target control action is determined through virtual execution evaluation and multiple screening mechanisms. This ensures that the generated control commands not only have clear energy-saving benefits but also fully comply with the actual execution conditions of the field equipment, greatly reducing the risk of parameter misadjustment and achieving a safe and reliable conversion of refined energy consumption modeling results into real-time optimized control actions. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a flowchart of the refined energy consumption modeling and real-time optimization control method based on process parameters according to the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the classification of process parameter sample set locations in an embodiment of the present invention.

[0018] Figure 3This is a flowchart of virtual execution evaluation and target control action screening in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram illustrating the recalculation relationship between the model training side and the real-time optimization side in an embodiment of the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] Figure 1 A flowchart of a method for refined energy consumption modeling and real-time optimization control based on process parameters according to at least one embodiment of the present disclosure is shown.

[0022] like Figure 1 As shown, in step 101, a sample set of process parameters for artificial intelligence modeling is generated.

[0023] Specifically, this step is used to generate a sample set of process parameters from existing systems on-site, which is required for subsequent training of the controllable energy consumption response AI model. For example... Figure 2 As shown, in practical implementation, the scope of field points entering the sample is first determined around the goal of energy consumption optimization and control. These field points are organized into four categories: controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators. Controllable process parameters are control quantities that can be actually written into PLCs, DCSs, SCADA systems, or frequency converters, such as temperature setpoints, pressure setpoints, pump frequency setpoints, fan frequency setpoints, valve opening commands, and heating power setpoints. Necessary state parameters are feedback quantities used to describe the current operating status of the equipment and participate in subsequent AI modeling, such as actual temperature, actual pressure, actual flow rate, motor current, and equipment operating status words. Energy consumption data comes from energy management systems, electricity meters, steam flow meters, gas meters, or compressed air flow meters. Constraint indicators come from process configuration files, equipment limit tables, quality standard documents, or control system limit tables, such as quality qualification boundaries, equipment safety boundaries, lower production cycle limits, and upper and lower limits of control parameters. If the MES batch records or operation logs contain personnel fields that are unrelated to process energy consumption control, then delete the field before forming the sample, so that the data entering the sample corresponds to equipment, process, energy metering, or process constraints.

[0024] Before forming samples, the system records the engineering units, sampling period, data source system, and validity markers for each point. Different types of fields, such as temperature, pressure, frequency, valve opening, flow rate, current, energy consumption, and output, are not directly added together. Instead, they are normalized and standardized according to their respective field types before entering model training, retaining the proportional coefficient or baseline value required for inverse normalization. If a point experiences communication interruption, out-of-bounds readings, missing units, or discontinuous timestamps within a process time slice, that time slice will not be used as a primary training sample or will be marked as a low-confidence sample. If a controllable process parameter lacks write permissions or write permissions are locked in the current production mode, that point will not be included in the controllable process parameter set and can only be saved as status or constraint information.

[0025] When specifically reading candidate points, the system reads the point name, address, read / write attributes, associated equipment, associated process segment, control loop number, sampling timestamp, historical records, and limit configuration from the PLC variable table, DCS tag number table, SCADA tag table, inverter communication address table, energy management system metering point table, process configuration file, and quality standard file to form a candidate point list. The points in this candidate point list are categorized hierarchically according to engineering attributes. First, the read / write attributes and control meaning of each point are read. If a point is marked as a setpoint, given value, output command, inverter frequency setpoint, or valve opening command in the PLC or DCS, and the point has write permissions to the control system, then the point is marked as a controllable process parameter. For example, in the hot air drying process, the "drying temperature setpoint," "circulating fan frequency setpoint," and "exhaust valve opening command" are derived from the DCS setpoint tag, the inverter communication address, and the valve actuator control point, respectively. All of these can be written into the field control system and are therefore marked as controllable process parameters. In the pumping process, the "pump frequency setpoint" and "outlet pressure setpoint" are derived from the PLC or inverter control address and are also marked as controllable process parameters. The marking results of these controllable process parameters directly determine that subsequent candidate parameter tuning actions can only be generated around these fields.

[0026] Subsequently, the read-only points are categorized into necessary state parameters. If a point originates from sensor feedback, inverter feedback register, or equipment status word, and is located in the same equipment, control loop, or process segment as a marked controllable process parameter, then that point is marked as a necessary state parameter. This determination is based on the naming relationship of equipment number, control loop number, process segment number, or process tag number. For example, "actual temperature of drying section" and "drying temperature setpoint" belong to the same drying section temperature control loop; "circulating fan current" and "circulating fan frequency setpoint" belong to the same circulating fan equipment; "actual pump outlet pressure" and "actual flow rate" and "pump frequency setpoint" belong to the same pumping loop. Therefore, these points are marked as necessary state parameters. Necessary state parameters are used to describe the equipment operating state after the controllable process parameters are applied, enabling the subsequent controllable energy consumption response AI model to distinguish between "energy consumption response caused by changes in controllable parameters" and "energy consumption background caused by the current equipment operating state." For example, if the fan current and actual air volume are both within a stable range after the fan frequency setpoint is reduced, the model can consider this operating condition as a normal response; if the fan frequency setpoint remains unchanged but the motor current rises abnormally, this status information can help the model identify that the current energy consumption change is more from the background of the operating status than from the active adjustment of controllable parameters.

[0027] Next, energy consumption metering points and constraint points are categorized. If a point originates from an electricity meter, steam flow meter, gas meter, compressed air flow meter, or energy management system, and its metering range covers the equipment, process segment, or production line where the marked controllable process parameters are located, then that point is marked as energy consumption data. For example, the electricity consumption of the drying section covers the process segment where the drying temperature setpoint, the circulating fan frequency setpoint, and the exhaust valve opening command are located, and is therefore marked as the energy consumption data of that process segment; the electricity consumption of the pump group covers the pump group where the pump frequency setpoint is located, and is therefore marked as the energy consumption data of the pumping process. If the energy consumption metering point is a cumulative reading, then the energy consumption result of that time segment is obtained based on the reading difference between adjacent process time segments when forming the sample record; if there is also a production count, then the energy consumption of that time segment is associated with the corresponding production output to form the unit product energy consumption. If a field originates from a process configuration file, quality standard file, equipment operating boundary, or control system limit table, and is used to limit controllable process parameters or target control actions, then it is marked as a constraint indicator. For example, the upper limits for outlet moisture content, furnace temperature, pipeline pressure, motor current, fan frequency, minimum flow rate, and production cycle time are all saved as constraint indicators. These constraint indicators are only saved with the sample records in this step and are used for subsequent virtual execution evaluation and target control action selection.

[0028] After completing the point classification, the classified points are organized into sample records according to process time slices. Process time slices are preferentially determined using energy metering cycles or production stability control cycles. For processes with thermal inertia or material residence time, such as drying, heating, and cooling, a single stability control cycle can be used. For processes with faster response, such as pumping, conveying, and compressed air supply, energy metering cycles can be used. Each sample record includes controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators within the same process time slice. If a controllable process parameter changes multiple times within a process time slice, the control system confirms that the setpoint has taken effect and remains stable. If a necessary state parameter experiences short-term fluctuations within a time slice, the recorded value used by the field monitoring system for stable operation judgment is used. If energy consumption data comes from cumulative electrical energy, cumulative steam volume, or cumulative gas volume, the difference between cumulative readings of adjacent time slices is used as the energy consumption result for that time slice. Taking the hot air drying process as an example, a sample record may include the drying temperature setpoint, circulating fan frequency setpoint, dehumidification valve opening command, actual temperature of the drying section, circulating fan current, power consumption increment of the drying section, upper limit of outlet moisture content, and upper limit of drying temperature within a certain stable time segment; taking the pumping process as an example, a sample record may include the pump frequency setpoint, outlet pressure setpoint, actual flow rate, motor current, pump power consumption increment, upper limit of pipeline pressure, and minimum flow requirement.

[0029] For processes with response lag, this step stores the correspondence between controllable process parameter time slices and energy consumption result time slices in the process parameter sample set. This correspondence comes from on-site timestamps, equipment stabilization cycles, or lag settings in the process configuration. Specifically, when a controllable process parameter undergoes a stable change within a certain time slice, that time slice is recorded as the parameter action time slice; then, based on the process stabilization cycle, the energy consumption results in one or more subsequent time slices are selected as the energy consumption response results corresponding to that parameter action time slice. For example, in a heating furnace, after the furnace temperature setpoint is adjusted, the actual furnace temperature and gas consumption often stabilize only in the following few time slices. Therefore, the sample record stores the correspondence between the time slice containing the furnace temperature setpoint and the subsequent stable gas consumption time slices. In the drying process, after changes in fan frequency or heating power, the outlet moisture content and unit product energy consumption may be reflected after the material passes through the drying section. Therefore, the sample record stores the correspondence between the time slice containing the fan frequency setpoint and heating power setpoint and the subsequent outlet moisture detection time slice and energy consumption metering time slice. By using this time-slice correspondence method, the controllable energy consumption response AI model for the next step can learn the energy consumption response after changes in controllable process parameters, rather than just learning the static correlation at the same moment.

[0030] The time slice correspondence also stores the method for determining the lag window and the applicable processes. For processes with significant thermal inertia, the lag window can be determined by the material residence time, temperature stabilization time, or the response cycle in the process configuration. For processes with faster responses, such as pumping and conveying, the lag window can be determined jointly by the energy metering cycle and the control system stability judgment cycle. If the same parameter change may result in energy consumption responses in multiple subsequent time slices, the system selects the energy consumption of the first time slice after entering the stable range or the average energy consumption of multiple stable time slices as the energy consumption response result, and saves this selection method in the sample record. This avoids the situation where the same controllable parameter change corresponds to multiple different energy consumption results without specifying the source of the value.

[0031] The final process parameter sample set consists of multiple sample records organized according to process time slices. Each sample record includes controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators, and records the source points and roles of each field. Controllable process parameters come from writable setpoints, given values, or execution instructions in PLCs, DCSs, SCADA systems, or frequency converters; necessary state parameters come from sensor feedback points such as temperature, pressure, flow, and current, or equipment status words; energy consumption data comes from electricity meters, steam flow meters, gas meters, compressed air flow meters, or energy management systems; constraint indicators come from process configuration files, quality standard documents, equipment operating boundaries, or field control system limit tables. This process parameter sample set serves as the input for the next step of training the controllable energy consumption response AI model. Through point classification and time slice organization oriented towards control objectives, the dispersed control variables, state variables, energy consumption metering values, and constraint boundaries in the field are organized into a structurally unified process parameter sample set, enabling the subsequent AI model to establish energy consumption response relationships around controllable process parameters.

[0032] In step 102, a controllable energy consumption response AI model is trained based on a process parameter sample set. The controllable energy consumption response AI model includes a controllable process parameter input channel, a necessary state parameter input channel, a state modulation layer, a controllable energy consumption response branch, and a state energy consumption background branch.

[0033] Specifically, this step uses the process parameter sample set formed in the previous step as input to train the controllable energy consumption response AI model, which will be directly called in subsequent steps. The process parameter sample set already contains controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators, and saves the correspondence between the controllable process parameter time slices and the subsequent stable energy consumption result time slices. Therefore, this step directly uses this sample structure for model training. In practice, the controllable process parameters in each sample record are sent to the controllable process parameter input channel, the necessary state parameters are sent to the necessary state parameter input channel, the energy consumption data is used as the supervised training target, and the constraint indicators are used as the process boundary of the sample along with the sample record. Taking the hot air drying process as an example, the sample record generated in step 101 includes the drying temperature setpoint, the circulating fan frequency setpoint, the dehumidification valve opening command, the actual temperature of the drying section, the circulating fan current, the power consumption increment of the drying section, the upper limit of outlet moisture content, and the upper limit of drying temperature. In this step, the first three writable fields are sent to the controllable process parameter input channel, the actual temperature and motor current are sent to the necessary state parameter input channel, the power consumption increment of the drying section is used as the training target, and the upper limit of outlet moisture content and the upper limit of drying temperature are saved as process boundaries along with the model configuration. The model trained in this way can determine the energy consumption response corresponding to "a slight decrease in the circulating fan frequency setpoint" under the current actual temperature and circulating fan current state of the drying section in subsequent steps, providing a model basis for the generation of candidate parameter tuning action sets.

[0034] Before training, each input field is converted into model input values ​​according to the units and normalization configuration saved in step 101, and the model output is then interpreted according to the same energy consumption caliber. The energy consumption caliber can be the energy consumption of a single process time slice, the energy consumption of a unit product, the energy consumption of a unit batch, or the energy consumption of a specific equipment group, but only one energy consumption caliber is used in the same model file; the controllable energy consumption response branch, the state energy consumption background branch, the training objective, and the predicted energy-saving response in subsequent candidate actions all use this caliber. If a certain process has electricity consumption, steam consumption, and gas consumption at the same time, the system can build corresponding models for each, or first convert them to a unified energy consumption caliber according to the project configuration before training, to avoid different energy types directly entering the same loss term without conversion.

[0035] In terms of training sample organization, training sample pairs are constructed according to the time slice correspondence stored in the process parameter sample set. For each parameter action sample, the controllable process parameter within the parameter action time slice is taken as the controllable input, the necessary state parameter within the same parameter action time slice is taken as the state input, and the energy consumption data within the subsequent stable energy consumption result time slice is taken as the energy consumption target. For processes with thermal inertia, material residence time, or equipment response delay, this sample organization method enables the model to learn the energy consumption response after the parameter action. For example, in a heating furnace, after the furnace temperature setpoint is stably adjusted in a certain time slice, the actual furnace temperature and gas consumption usually enter a new stable state in subsequent time slices. Therefore, the time slice where the furnace temperature setpoint is located is paired with the subsequent stable gas consumption time slice. In the pumping process, after the pump frequency setpoint changes, the response period of flow rate, motor current, and pump group power consumption is relatively short. Therefore, a shorter stable time window is used to form sample pairs. Constraint metrics are used to determine the available process boundaries of the samples. For example, samples with outlet moisture, equipment current, pipeline pressure, minimum flow rate, or production cycle time within the process allowable range are used for primary training. Samples close to the boundary retain boundary markers so that subsequent virtual execution evaluations can combine quality, safety, and cycle time boundaries to determine whether parameter tuning actions are available.

[0036] The controllable energy consumption response AI model adopts a dual-channel temporal structure, consisting of a controllable process parameter input channel, a necessary state parameter input channel, a linear mapping layer, a temporal modeling layer, a state modulation layer, a controllable energy consumption response branch, and a state energy consumption background branch. The controllable process parameter input channel receives fields marked as controllable process parameters in step 101, such as temperature setpoints, pressure setpoints, pump frequency setpoints, fan frequency setpoints, valve opening commands, and heating power setpoints. These fields first enter the linear mapping layer, which consists of trainable weights and biases, used to map multiple controllable fields within the same time slice into fixed-length controllable features. The necessary state parameter input channel receives fields marked as necessary state parameters in step 101, such as actual temperature, actual pressure, actual flow rate, motor current, and equipment operating status words. These fields enter another set of linear mapping layers to form state features. Each of the two linear mapping layers is followed by either a gated recurrent unit (ROU) or a one-dimensional temporal convolutional layer. When the process exhibits significant hysteresis and a stable period, the GU is used to extract the change dependencies between consecutive time slices. When the process response is fast and the sampling period is stable, the one-dimensional temporal convolutional layer is used to extract the change patterns within adjacent time slices. The GU includes an update gate, a reset gate, and candidate state calculation. The one-dimensional temporal convolutional layer includes a fixed-length convolutional kernel, a stride, and an activation layer. The specific structure is determined in the model configuration based on the process response time and sampling period.

[0037] After the two channels form temporal features, the model sets a state modulation layer to modulate the energy consumption response of the necessary state parameters to the controllable process parameters. The state modulation layer first inputs the state features output from the necessary state parameter channels into a fully connected layer, then passes them through a Sigmoid activation layer to obtain modulation coefficients. These modulation coefficients are then applied element-wise to the controllable features output from the controllable process parameter channels to obtain the state-modulated controllable response features. This structure corresponds to common control phenomena in industrial settings: the same controllable parameter adjustment has different energy consumption effects under different equipment states. For example, when the pump frequency decreases by the same amount, if the current actual flow rate is sufficient, the pipeline pressure is stable, and the motor current is normal, the energy-saving response corresponding to this action is more explicit; if the current actual flow rate is close to the lower limit or the motor current fluctuates significantly, the controllable energy consumption response corresponding to this action will be weakened by the state modulation layer. Similarly, in the drying process, reducing the setpoint of the circulating fan frequency usually reduces fan power consumption, but when the actual temperature in the drying section is low or the outlet moisture content is close to the mass boundary, the state modulation layer will weaken the controllable response corresponding to this change in controllable parameter, making the subsequent candidate actions more closely aligned with process stability requirements.

[0038] The model output consists of two branches. The controllable energy consumption response branch receives the controllable response features modulated by the state and outputs the energy consumption response results corresponding to changes in controllable process parameters. The state energy consumption background branch receives the state features output from the necessary state parameter channels and outputs the energy consumption background results corresponding to the current operating state. Both branches are implemented using a two-layer fully connected network. The first layer is used to compress channel features and introduce nonlinear activation, and the second layer outputs the corresponding energy consumption results. The final predicted energy consumption is obtained by synthesizing the two branches. The output of the controllable energy consumption response branch is used in subsequent steps to identify controllable process parameters with energy-saving response directions, while the state energy consumption background branch is used to characterize the energy consumption level corresponding to background factors such as current equipment load, current temperature, pressure, flow rate, and motor status. Taking the pumping process as an example, when the pump frequency setpoint decreases but the actual flow rate still meets the minimum flow requirement, the controllable energy consumption response branch outputs the energy consumption change corresponding to this frequency adjustment; when the pump frequency setpoint remains stable but the pipeline pressure and motor current increase, the state energy consumption background branch outputs the increase in energy consumption background corresponding to this state change.

[0039] The outputs of the two branches are not two independent prediction targets, but rather serve the same energy consumption prediction and action evaluation task. The controllable energy consumption response branch expresses the energy consumption response caused by changes in writable parameters, while the state energy consumption background branch expresses the background energy consumption caused by load, environment, incoming materials, and equipment status. The sum of the two is compared with the sample energy consumption target. The model file stores the meaning and energy consumption unit of each output branch. During real-time runtime, step 103 only calls the controllable energy consumption response branch to detect candidate actions, while step 104 uses both the controllable response result and the current state background to determine whether the action is suitable for issuance.

[0040] The model training employs a joint loss oriented towards controllable responses. This formula is based on the squared error loss in supervised learning, originally representing the squared difference between the predicted and true values. In this application, predicted energy consumption is decomposed into the sum of the output of the controllable energy consumption response branch and the output of the state energy consumption background branch, thus the first term is derived from the original squared error. Further, by combining the parameter action time slice saved in step 101 with the subsequent stable energy consumption result time slice, the actual energy consumption change before and after the controllable parameter change can be obtained, thereby adding a second term to ensure that the controllable energy consumption change result output by the model is consistent with the actual energy consumption change result in the samples. Furthermore, for sample segments where controllable process parameters remain stable, the fluctuation of the controllable response branch is constrained to a smaller range, thus forming a third term, making the energy consumption background caused by state changes more representative of the state energy consumption background branch. All three terms are formed around the same energy consumption measurement caliber, and the weight parameters serve as scaling coefficients in the training configuration, ensuring that each term can be optimized together during training. The training loss is as follows: ; in, This indicates that when training a controllable energy consumption response AI model, at the first... The loss value calculated on each sample record is obtained by the model training program based on the sample input and model output. Represents the first in the process parameter sample set Each sample record corresponds to energy consumption data from energy management systems, electricity meters, steam flow meters, gas meters, compressed air flow meters, or unit product energy consumption results. Indicates the first The controllable energy consumption component corresponding to each sample record is calculated by the fully connected output layer of the controllable process parameter input channel, linear mapping layer, timing modeling layer, state modulation layer and controllable energy consumption response branch; Indicates the first The state energy consumption background component corresponding to each sample record is calculated by the fully connected output layer of the necessary state parameter input channel, linear mapping layer, temporal modeling layer and state energy consumption background branch; The weights of the controllable response consistency term are given by the training configuration and are used to control the constraint strength between the actual energy consumption change and the controllable energy consumption change of the model output. This represents the actual energy consumption change result obtained from the parameter application time slice saved in step 101 and the subsequent stable energy consumption result time slice; This represents the output difference of the controllable energy consumption response branch on two samples before and after the parameter change. It is the result of the controllable energy consumption change obtained by subtracting the controllable energy consumption components of the corresponding samples, rather than an independent value. New branch outputs in addition to; The weight of the static stability suppression term is given by the training configuration and is used to control the fluctuation amplitude of the controllable energy consumption response branch when the controllable process parameters are stable. Indicates the stability of controllable process parameters, when the first... A sample record is assigned a value of 1 when the controllable process parameter is in a stable state compared to its adjacent or corresponding sample, and a value of 0 when the controllable process parameter undergoes a stable change. This value is obtained from the time slice records of the controllable process parameters in the process parameter sample set and the control system confirmation records. Where... , , , and All use the same energy consumption caliber or the same normalized numerical caliber; and , where are the weight coefficients in the training configuration, used to adjust the relative influence of the three losses during the training process, and do not represent the physical quantities in the field. The first term in the formula is used to train the model to synthesize energy consumption results, the second term is used to train the controllable energy consumption response branch to express the energy consumption difference before and after parameter changes, and the third term is used to suppress invalid fluctuations of the controllable energy consumption component in the stable segment of controllable parameters, so that the energy consumption change caused by the necessary state parameter changes is more represented by the state energy consumption background branch.

[0041] The calculation of this joint loss can be performed directly within a specific process sample. Taking a training sample from a hot air drying process as an example, the energy consumption data for the drying section corresponding to a certain stable time slice... Take 120, and output the controllable energy consumption response branch of the model. Take 34, state energy consumption background branch output If we take 83, then the first term is The frequency setpoint of the circulating fan corresponding to this sample decreases steadily during the parameter application time slice, and the energy consumption reduction in subsequent stable time slices relative to the initial value before the change is... Taking -8, the model outputs the controllable energy consumption change result for this frequency change. Take -7, if If we take 0.6, then the second term is This sample belongs to a segment where controllable process parameters undergo stable changes. If we set the third term to 0, then the training loss for this sample is: Taking another sample in the same drying process as an example, the given value of the circulating fan frequency remains stable, while the actual temperature of the drying section and the circulating fan current fluctuate. The model's controllable energy consumption changes are as follows. Take 2, Take 1, if If we take 0.4, then the third term is This improvement reduces the fluctuation of the controllable energy consumption response branch in the model when the controllable parameters remain unchanged, allowing such energy consumption changes to be represented more by the state energy consumption background branch. Through the above calculations, the model training process allocates "energy consumption changes caused by controllable parameters" and "energy consumption changes caused by the operating state background" to different output branches.

[0042] During training, batch samples are organized according to equipment, process segment, or control loop. For sample segments where controllable parameters undergo stable changes, the training program extracts the energy consumption difference before and after the parameter change based on the time slice correspondence saved in step 101, and uses this difference as... Participate in training; for sample segments where controllable parameters remain stable but necessary state parameters change, the training program will... The value is set to 1, and the controllable energy consumption response branch is kept stable by constraining the static stability suppression term. Taking the pumping process as an example, if the actual flow rate still meets the minimum flow requirement and the pump power consumption decreases after the pump frequency setpoint is reduced, this sample is used to strengthen the expression of the controllable energy consumption response branch for pump frequency changes; if the pump frequency setpoint does not change but the pipeline pressure and motor current increase, this sample is used to train the state energy consumption background branch to express the increase in operating background energy consumption. Taking the heating furnace as an example, if the product temperature remains within the acceptable range and the gas consumption decreases after the furnace temperature setpoint is reduced, this sample is used to train the controllable energy consumption response branch; if the furnace temperature setpoint is stable but the actual furnace temperature fluctuates due to changes in material load, this sample is used to train the state energy consumption background branch. After the above training, the model can judge the strength of the energy-saving response of the same parameter adjustment action based on the current necessary state parameters in subsequent real-time operation.

[0043] If a change in a controllable parameter in a training sample is accompanied by drastic fluctuations in multiple state parameters, exceeding of quality boundaries, or exceeding of equipment safety boundaries, the training program will not treat this sample as a normal controllable response sample, but will instead mark it as a boundary sample or remove it from the main training batch. If an energy consumption metering point is missing, jumps, or its cumulative value returns to zero in the corresponding time slice, the training program will not use that time slice for calculation. Through the above processing, the controllable response samples used for model training are consistent with the executable parameter tuning actions in the field, reducing the likelihood that abnormal operating condition samples will cause the model to learn unexecutable or unsafe changes as energy-saving actions.

[0044] After training, the controllable energy consumption response AI model is solidified into a model file containing the order of input fields, field roles, network structure parameters, and the meaning of output branches. The order of input fields is inherited from the role records in the process parameter sample set. Controllable process parameters enter the controllable process parameter input channel, necessary state parameters enter the necessary state parameter input channel, energy consumption data is used as the training target, and constraint indicators are saved according to the model's configuration. The model file stores the weights of the linear mapping layer, the temporal modeling layer, the state modulation layer, the controllable energy consumption response branch weights, and the state energy consumption background branch weights. During real-time operation, the controllable process parameters and necessary state parameters under the current operating condition are input into the model in the same field order. The model outputs the controllable energy consumption response result and the state energy consumption background result under the current state. The next step directly uses the controllable energy consumption response result to identify controllable process parameters with energy-saving response directions and generates a set of candidate parameter tuning actions around these controllable process parameters.

[0045] This step outputs a result: a controllable energy consumption response AI model. This model consists of a controllable process parameter input channel, a necessary state parameter input channel, a linear mapping layer, a timing modeling layer, a state modulation layer, a controllable energy consumption response branch, and a state-energy consumption background branch. The model's input fields, training objectives, and constraint boundaries all come from the process parameter sample set output in step 101. The controllable process parameters are used to learn the adjustable energy consumption response, the necessary state parameters are used to characterize the current operating context, the energy consumption data is used for supervised training, and the constraint indicators are used to limit the effective boundaries of the samples and the available boundaries of subsequent actions. Through the above training method, the structured samples formed in step 101 are transformed into an AI model capable of expressing "controllable process parameter changes—energy consumption response." This model will directly serve as the input for generating the candidate parameter tuning action set in step 103.

[0046] In step 103, the current operating condition data is input into the controllable energy consumption response artificial intelligence model for response detection, generating a candidate parameter tuning action set. The candidate parameter tuning action set includes at least one candidate parameter tuning action and at least one predicted energy-saving response of the candidate parameter tuning action.

[0047] Specifically, this step uses the controllable energy consumption response AI model output from the previous step as input to generate a set of candidate parameter tuning actions required for subsequent virtual execution evaluation. The controllable energy consumption response AI model has already fixed the controllable process parameter input channel, necessary state parameter input channel, state modulation layer, controllable energy consumption response branch, state energy consumption background branch, input field order, field role, and constraint indicators saved with the model configuration. Therefore, this step directly reads the current operating condition data according to the field order of the model. In specific implementation, the current controllable process parameter values ​​are read from the PLC, DCS, SCADA, or frequency converter, such as temperature setpoint, pressure setpoint, pump frequency setpoint, fan frequency setpoint, valve opening command, and heating power setpoint; the current necessary state parameter values ​​are read from sensor feedback points or equipment status words, such as actual temperature, actual pressure, actual flow rate, motor current, and equipment operating status words; and constraint indicators are read from the model configuration, such as the upper and lower limits of controllable process parameters, maximum single adjustment range, rate of change, write cycle, quality qualification boundary, equipment safety boundary, and production cycle lower limit. The aforementioned current operating condition data are input into the same controllable energy consumption response AI model according to the field roles used in step 102 training, ensuring that the candidate action generation process maintains the same process meaning as the model training process. Taking the hot air drying process as an example, the current operating condition includes the drying temperature setpoint, the circulating fan frequency setpoint, the dehumidification valve opening command, the actual temperature of the drying section, the circulating fan current, and the outlet moisture detection value. This step generates candidate adjustment actions based on the first three writable fields and uses the latter three state fields to participate in the model response judgment. Taking the pumping process as an example, the current operating condition includes the pump frequency setpoint, the outlet pressure setpoint, the actual flow rate, the motor current, and the actual outlet pressure. This step generates candidate actions based on the pump frequency setpoint and the outlet pressure setpoint and uses the actual flow rate, the motor current, and the actual outlet pressure to participate in the model's judgment of the energy-saving response of the actions.

[0048] When reading current operating data, the system simultaneously reads the timestamp and validity flag of each field. If the current controllable process parameters are not in automatic control mode, write permissions are locked, the actuator is in a fault state, or communication with necessary state parameters is interrupted, then no candidate action will be generated based on that field. If the input fields required by the energy consumption response AI model are missing, this step will not perform model response detection, but will instead output an empty candidate parameter tuning action set and maintain the original control settings. In this way, candidate actions are only generated when the current data is complete, the control system is writable, and the field is in a state that allows optimization.

[0049] The generation of candidate parameter tuning actions employs a continuous processing approach: "current response reading, adjustment grid point generation, model response detection, action disturbance calculation, and candidate action retention." First, the current controllable process parameters and necessary state parameters are input into the controllable energy consumption response AI model. The model's controllable energy consumption response branch outputs the baseline controllable energy consumption response result under the current operating condition. Then, for each controllable process parameter, candidate target values ​​are generated based on the control system's allowed write granularity, upper and lower limits, maximum single adjustment amplitude, and rate of change. These candidate target values ​​are derived from the set grid points that the field control system can actually execute. For example, the inverter frequency setpoint is generated according to the inverter's allowed set granularity, the valve opening command according to the actuator's allowed opening granularity, and the temperature setpoint according to the DCS's allowed set granularity. Each candidate parameter tuning action consists of an action object, a current value, a target value, and an action direction; the action object comes from the controllable process parameter fields stored in the controllable energy consumption response AI model. For single-device, single-loop control scenarios, a candidate action changes one controllable process parameter. For control loops that require synchronous changes in the process, linked actions are generated based on the same-loop markers in the model configuration. For example, in the drying process, the heating power setpoint and the circulating fan frequency setpoint can form a small-scale synchronous action; in the combustion control scenario, the gas valve opening and the combustion fan frequency can form the same action group. Each target value in the linked action comes from the allowable range of its respective control system, and the current and target values ​​of each parameter within the action group are retained for the next step to evaluate quality, safety, and cycle time boundaries item by item.

[0050] The model response detection for candidate actions employs the finite difference approach. Finite difference originates from the classic numerical analysis method of "keeping other conditions consistent and comparing the difference in the target quantity before and after the variable change." This step adapts it to the industrial energy consumption control scenario, limiting the target quantity to the output of the controllable energy consumption response branch obtained in step 102. Specifically, while maintaining the current necessary state parameters as inputs, the current controllable process parameters are input into the model to obtain the baseline controllable energy consumption response result. Then, the controllable process parameters involved in the candidate parameter tuning action are replaced with the target values, while keeping the remaining controllable process parameters and necessary state parameters unchanged. These are then input into the same model to obtain the controllable energy consumption response result after the action. The difference between the two represents the potential decrease in controllable energy consumption response under the current state conditions. ; in, This represents a candidate parameter adjustment action, consisting of the action object, the current value, and the target value. The action object comes from the controllable process parameter field saved in the controllable energy consumption response AI model, the current value comes from the current set value of the PLC, DCS, SCADA, or frequency converter, and the target value comes from the adjustment grid point within the write range allowed by the control system. Indicates candidate parameter tuning action The corresponding predicted energy-saving response is used to determine whether the action should be included in the candidate parameter tuning action set. It represents the baseline controllable energy consumption response result under the current operating conditions, and is output by the controllable energy consumption response branch of the controllable energy consumption response AI model under the input of current controllable process parameters and current necessary state parameters. Indicates the action of candidate parameter tuning After the target value replaces the current controllable process parameters, the controllable energy consumption response result after the action is output by the controllable energy consumption response branch of the same controllable energy consumption response AI model. , and Using the same energy consumption response caliber, this difference can be used to compare energy-saving responses between different candidate actions under the same current operating condition. If the model output uses normalized energy consumption response values, the system first compares candidate actions under the same normalization caliber, and then converts them into field-readable energy-saving response records according to the inverse normalization parameters in the model configuration; if... Not less than ,but It does not indicate energy saving and will not be included in the priority candidates.

[0051] This formula is derived from the finite difference response term. The original form of the finite difference is "the objective quantity before the change minus the objective quantity after the change," which corresponds to the following in this step. Taking the frequency setpoint of the circulating fan in the hot air drying process as an example, the baseline controllable energy consumption response is obtained after inputting the current operating condition into the model. The control system allows the frequency of the circulating fan to be adjusted slightly at a set granularity; this step generates candidate actions. Its action object is the given frequency of the circulating fan, and the target value is the next allowable grid point below the current frequency. After replacing the current frequency, input the same controllable energy consumption response AI model to obtain the controllable energy consumption response result after the action. ,but When the minimum effective predicted energy-saving response configured on-site is 1.0, this action enters the candidate parameter tuning action set. Another candidate action... Slightly lower the drying temperature setpoint; if the model output... ,but If the predicted energy-saving response of this action is lower than the minimum effective predicted energy-saving response, it will not be considered a priority candidate under the current operating conditions. Taking the pumping process as an example, if the controllable energy consumption response output by the model decreases significantly after the pump frequency setpoint is reduced by one adjustment grid point, then this action will be included in the candidate parameter tuning action set; if the decrease in controllable energy consumption response caused by reducing the outlet pressure setpoint is insufficient, then this action will not be included in the candidate parameter tuning action set under the current operating conditions.

[0052] After net energy-saving response calculation, candidate actions enter the action boundary trimming process. Action boundary trimming directly uses constraint indicators inherited from the process parameter sample set in the controllable energy consumption response AI model configuration. For each candidate parameter adjustment action, it is first determined whether the target value is within the upper and lower limits of the controllable process parameter, then it is determined whether the difference between the target value and the current value meets the requirements of the maximum adjustment amplitude and rate of change in a single operation, and finally it is determined whether the controllable process parameter is within the current control system's allowed write cycle. For linked actions, it is determined whether each controllable process parameter within the action group meets its own boundaries, and it is confirmed that the action group belongs to the same equipment, the same control loop, or a preset linkage relationship. For example, the action of lowering the setpoint of the circulating fan frequency needs to meet the lower limit of the fan frequency, the minimum air volume requirement, and the inverter write cycle; the action of lowering the setpoint of the drying temperature needs to meet the lower limit of the drying temperature, the outlet moisture boundary, and the temperature change rate limit; the action of lowering the setpoint of the pump frequency needs to meet the lower limit of the pump frequency, the minimum flow requirement, and the pipeline pressure boundary. After action boundary trimming, each retained action has a writable target value and clear field execution conditions.

[0053] Action boundary pruning also records the reason for each rejected action, such as target value exceeding the limit, exceeding the maximum adjustment range in a single operation, not reaching the write cycle, insufficient associated quality boundary, insufficient equipment safety boundary, or actuator being unwritable. Rejected actions do not participate in subsequent virtual execution evaluations, nor are they issued in the next step. If all target grid points for the same controllable process parameter are rejected, the corresponding action for that parameter will no longer be retained in the candidate parameter tuning action set to avoid invalid evaluation of unwritable actions in subsequent steps.

[0054] After action boundary trimming, this step confirms the state consistency of candidate actions based on the current necessary state parameters. State consistency confirmation invokes the state modulation structure in the controllable energy consumption response AI model from step 102, where the current necessary state parameters influence the strength of the controllable response of candidate actions. For the same candidate action, when the current state input is within a stable operating range, the state-modulated controllable energy consumption response is clearer; when the current state input is close to the quality, safety, or cycle time boundaries, the state-modulated controllable response weakens. Taking the drying process as an example, when the outlet moisture detection value is far from the quality qualification boundary, the actual temperature of the drying section is stable, and the circulating fan current is normal, a slight reduction in the circulating fan frequency setpoint usually yields a higher net energy saving response; when the outlet moisture detection value is close to the boundary or the actual temperature is low, the same frequency reduction action will yield a lower net energy saving response in the model. Taking the pumping process as an example, when the actual flow rate is higher than the minimum flow requirement and the pipeline pressure is stable, the pump frequency reduction action retains a relatively clear candidate value; when the actual flow rate is close to the minimum flow requirement or the motor current fluctuates significantly, the net energy saving response of this action will be weakened. This process ensures that the candidate action set for parameter tuning is already combined with the current operating conditions during the generation phase, and the next step is to continue to perform virtual execution evaluation on the candidate actions.

[0055] When multiple candidate actions are generated for the same controllable process parameter, this step adopts a retention strategy prioritizing predicted energy-saving effectiveness and low-disturbance priority. If multiple target values ​​exist for the same action and all predicted energy-saving responses reach the minimum effective predicted energy-saving response, the action with the smaller adjustment range is retained first. If the predicted energy-saving response of a small action is insufficient, but a slightly larger action reaches the minimum effective predicted energy-saving response, the smaller action that meets the requirements is retained. If the predicted energy-saving responses of multiple actions are close, the action with fewer actuator adjustments, the action direction consistent with the previous control cycle, and a shorter write cycle is retained first. Taking valve opening commands as an example, if both lowering the opening by one grid point and two grid points can generate an effective predicted energy-saving response, this step prioritizes retaining the candidate action of lowering by one grid point. If the response of one grid point is insufficient, but two grid points reach the minimum effective predicted energy-saving response, the action of lowering by two grid points is retained, and the adjustment range is marked in the candidate action record. This record will be used for virtual execution evaluation in the next step.

[0056] Each action in the candidate parameter adjustment action set is recorded according to a unified format. The action record includes the action object, current value, target value, action direction, adjustment range, net energy-saving response, associated necessary state parameters, and associated constraint indicators. The action object comes from the controllable process parameter input channel field in the controllable energy consumption response AI model, such as the setpoint for circulating fan frequency, pump frequency, valve opening command, and heating power; the current value comes from the real-time reading of the PLC, DCS, SCADA, or inverter's current setting; the target value comes from the adjustment grid point within the control system's allowable range; the action direction indicates upward or downward adjustment; the adjustment range is determined by the current value and the target value; and the predicted energy-saving response is determined by... The calculations are as follows: The necessary state parameters are derived from the necessary state parameter input channel fields in the controllable energy consumption response AI model, such as actual temperature, actual pressure, actual flow rate, and motor current; the associated constraint indicators are derived from the quality qualification boundary, equipment safety boundary, production cycle lower limit, or parameter upper and lower limits saved in the model configuration. Taking the candidate action of reducing the frequency of the circulating fan in the drying process as an example, its action record includes the current frequency, target frequency, direction of reduction, frequency change amplitude, net energy saving response, actual temperature of the drying section, circulating fan current, upper limit of outlet moisture content, and minimum airflow requirement. Similarly, taking the candidate action of reducing the pump frequency in the pumping process as an example, its action record includes the current frequency, target frequency, direction of reduction, frequency change amplitude, net energy saving response, actual flow rate, motor current, pipeline pressure boundary, and minimum flow requirement.

[0057] This step outputs a candidate set of parameter tuning actions. This set consists of several candidate actions, each generated by the controllable energy consumption response AI model output in step 102 under the current operating conditions. Each action includes the action object, current value, target value, action direction, adjustment range, predicted energy-saving response, associated necessary state parameters, and associated constraint indicators. The actions in the candidate set are generated around the controllable process parameter fields marked in step 101 and solidified by the model in step 102. The predicted energy-saving response comes from the difference in output from the controllable energy consumption response branch of the controllable energy consumption response AI model. This candidate set of actions will serve as input for the next step of virtual execution evaluation and the issuance of target control actions. Through this step, the controllable energy consumption response AI model formed in step 102 is transformed into a finite, low-disturbance candidate control action that meets the conditions for on-site writing, allowing the refined energy consumption modeling results to enter the real-time optimization control stage.

[0058] If no action satisfies write permissions, boundary constraints, and minimum effective predicted energy-saving response under the current operating conditions, the candidate parameter tuning action set can be empty. An empty candidate parameter tuning action set is also a valid output of this step, indicating that no real-time optimization write will be performed under the current operating conditions. The next step will output based on this, maintaining the currently set result, rather than forcibly generating the target control action.

[0059] In step 104, at least one candidate parameter tuning action in the candidate parameter tuning action set is virtually executed, evaluated, and screened to determine the target control action, and the target control action is sent to the field control system.

[0060] Specifically, such as Figure 3As shown, this step takes the candidate parameter tuning action set output in the previous step as input, and combines it with the controllable energy consumption response AI model solidified in step 102 and the complete current operating condition read in real time by the field control system to perform virtual execution evaluation on each candidate parameter tuning action in the candidate parameter tuning action set, and determines the target control action that can be written into the field control system. Each candidate parameter tuning action in the candidate parameter tuning action set already includes the action object, current value, target value, action direction, adjustment range, predicted energy-saving response, associated necessary state parameters, and associated constraint indicators; the complete current operating condition includes the current values ​​of all controllable process parameters and necessary state parameters read in real time by the field. In specific implementation, for each candidate parameter tuning action, its target value replaces the corresponding controllable process parameter value in the complete current operating condition, the remaining controllable process parameters adopt the current values ​​read in real time by the field, the associated necessary state parameters retain the current feedback values, and are input into the model according to the field order of the controllable energy consumption response AI model solidified in step 102. Controllable process parameters enter the controllable process parameter input channel, and associated necessary state parameters enter the necessary state parameter input channel. The state modulation layer modulates the controllable response characteristics of the candidate action according to the current necessary state parameters. The controllable energy consumption response branch outputs the controllable energy consumption response result after the action, and the state energy consumption background branch outputs the energy consumption background result corresponding to the current operating state. This virtual execution evaluation is used to determine whether the energy consumption response of the candidate action is still effective under the current operating background. Quality, safety, and cycle time conditions are conservatively screened using the current necessary state parameters, associated constraint indicators, and boundary margin requirements configured on site. Taking the hot air drying process as an example, if the candidate action is to reduce the setpoint of the circulating fan frequency, then in the virtual working condition, only the setpoint of the circulating fan frequency is replaced with the target frequency. The drying temperature setpoint and the dehumidification valve opening command use the real-time current values ​​on site. The actual temperature of the drying section, the circulating fan current, and the outlet moisture detection value are used as the current state background input model. The model outputs the controllable energy consumption response and state energy consumption background under the target frequency accordingly. Taking the pumping process as an example, if the candidate action is to reduce the pump frequency setpoint, then only the pump frequency setpoint is replaced in the virtual working condition, and the other controllable process parameters adopt the real-time current values ​​on site. The actual flow rate, motor current and actual outlet pressure are used as the background input model of the current state, so that the model can determine whether the energy-saving response of the action is still available under the current flow and pressure conditions.

[0061] If the candidate parameter tuning action set output in step 103 is empty, or the timestamp of the current complete operating condition data is inconsistent with the data when the candidate action was generated, this step will not construct a virtual operating condition or issue a new target control action, but will output the result that maintains the current control settings. If the field control system mode switches from automatic to manual after the candidate action is generated, the actuator fault state changes, or the associated constraint index goes out of bounds, this step will also stop issuing the action and record the reason for stopping.

[0062] After the virtual operating condition is constructed, this step determines the target control action in the order of "hard constraint screening, energy-saving effectiveness screening, boundary margin screening, and low-disturbance optimization". Hard constraint screening is used to determine whether the candidate parameter tuning actions have the basic conditions for issuance. For each candidate parameter tuning action, first read its target value and the upper and lower limits of the parameters corresponding to the action object to determine whether the target value is within the writable range; then read the quality qualification boundary, equipment safety boundary, and production cycle lower limit associated with the action, and combine them with the currently associated necessary state parameters to determine whether the current operating condition is within the allowable range. The quality qualification boundary comes from the quality standard document or process configuration file, such as the upper limit of outlet moisture in the drying process, the product qualification temperature range in the heating furnace, and the thickness deviation range in the coating process; the equipment safety boundary comes from the equipment operating boundary or control system limit table, such as the upper limit of motor current, the upper limit of pipeline pressure, the upper limit of furnace temperature, and the upper and lower limits of valve opening; the production cycle lower limit comes from the process configuration file, production line cycle configuration, or control system limit table, such as the minimum flow requirement, the minimum throughput requirement, and the equipment allowable write cycle. In specific judgments, the associated necessary state parameters in the candidate action record are used to represent the current operating condition. The controllable energy consumption response AI model in step 102 is used to output the energy consumption response trend after the action. The associated constraint indicators in the candidate action record are used to determine whether the current operating condition meets the boundary conditions for executing the action. For example, a candidate action to reduce the frequency of the circulating fan needs to simultaneously meet the following conditions: the target frequency is not lower than the lower limit of the fan frequency, the current outlet moisture does not exceed the upper limit, the current circulating fan current does not exceed the upper limit, and the current air volume meets the minimum air volume requirement. A candidate action to reduce the pump frequency needs to simultaneously meet the following conditions: the target frequency is not lower than the lower limit of the pump frequency, the current actual flow rate is not lower than the minimum flow rate requirement, the current pipeline pressure does not exceed the boundary, and the current motor current does not exceed the upper limit. A candidate action to reduce the furnace temperature setpoint needs to simultaneously meet the following conditions: the furnace temperature setpoint is within the allowable range, the current product temperature is within the qualified range, and the current furnace temperature is within the process stability range. Actions that pass the hard constraint screening enter the subsequent energy-saving effectiveness judgment, and actions that do not meet any hard constraints are removed from the candidate parameter adjustment action set.

[0063] After hard constraint screening, this step screens the remaining candidate actions for energy-saving effectiveness. Energy-saving effectiveness is directly determined using the predicted energy-saving response recorded in step 103 for the candidate parameter tuning actions. This predicted energy-saving response is obtained by the difference between the baseline controllable energy consumption response output by the controllable energy consumption response AI model in step 103 and the controllable energy consumption response after the action, representing the controllable energy consumption reduction that the candidate action can bring under the current operating conditions. In practice, the net energy-saving response recorded for each candidate action is read and compared with the minimum effective predicted energy-saving response configured on-site. The minimum effective predicted energy-saving response comes from the process configuration or energy-saving control threshold and is used to exclude actions whose energy-saving effect is below the perceptible range on-site. For example, in the hot air drying process, if the net energy-saving response of reducing the circulating fan frequency by one allowable grid point reaches the minimum effective predicted energy-saving response, the action is retained; if the net energy-saving response of slightly reducing the drying temperature setpoint is below this threshold, even if the action meets the parameter boundaries, it will not be included in the target control action selection range. Taking the pumping process as an example, if the net energy-saving response of reducing the pump frequency reaches the threshold and the minimum flow rate still meets the requirements, then the action is retained; if the net energy-saving response from adjusting the outlet pressure setpoint is weak, then the action is removed from subsequent selections. Energy-saving effectiveness screening ensures that target control actions are based on a clear and controllable energy consumption response, rather than simply being issued based on the direction of parameter changes.

[0064] After screening for energy-saving effectiveness, this step performs boundary margin screening on the retained actions. Boundary margin screening is performed separately for quality boundaries, safety boundaries, and cycle time boundaries, without merging these three types of boundaries into a single comprehensive score. For quality boundaries, the remaining distance between the current quality indicator and the quality qualification boundary is calculated or read, such as the margin between outlet moisture and the upper limit, the margin between product temperature and the lower limit, and the margin between coating thickness and the deviation boundary. For equipment safety boundaries, the remaining distance between the current equipment status and the equipment operation boundary is calculated or read, such as the margin between motor current and the upper limit, the margin between pipeline pressure and the upper limit, and the margin between furnace temperature and the upper or lower limit. For production cycle time boundaries, the remaining distance between the current production line capacity and the lower limit of the cycle time is calculated or read, such as the margin between actual flow rate and the minimum flow rate requirement, and the margin between production line throughput and the minimum throughput. Each type of boundary margin is compared separately with its corresponding boundary. Only actions whose quality margin, safety margin, and cycle time margin are all not lower than the corresponding minimum margin requirement enter the low-disturbance selection stage. Taking the drying process as an example, reducing the frequency of the circulating fan is retained if there is still sufficient margin between the outlet moisture level and the upper limit, the circulating fan current is within a safe range, and the minimum airflow requirement is still met. If the outlet moisture level is already very low, this action, even if it has a net energy-saving response, will not be included in the final selection. Taking the pumping process as an example, reducing the pump frequency is retained if there is still margin between the actual flow rate and the minimum flow rate requirement, and both the pipeline pressure and motor current are within a safe range. If the actual flow rate is close to the lower limit, this action is eliminated. Boundary margin screening allows this step to judge process quality, safe operation, and cycle stability as independent engineering conditions, avoiding the simple merging of boundaries with different physical meanings.

[0065] After hard constraint screening, energy-saving effectiveness screening, and boundary margin screening, if multiple available candidate actions still exist, this step uses a low-disturbance selection rule to determine the target control action. The low-disturbance selection is performed sequentially according to the stability requirements of the field control execution: first, actions with smaller adjustment amplitudes are prioritized; if multiple actions have similar adjustment amplitudes, actions with the same direction as the previous control cycle are prioritized; if the direction of the action is still indistinguishable, actions with shorter actuator write cycles and smoother responses are prioritized; if multiple actions still exist, actions with higher net energy-saving responses are selected. The adjustment amplitude is derived from the difference between the target value and the current value in the candidate parameter tuning action record; the direction of the action is derived from the upward or downward adjustment markers in the candidate parameter tuning action record; the actuator write cycle is derived from the control system configuration or actuator configuration file; and the net energy-saving response is derived from the candidate action record in step 103. Taking valve opening commands as an example, if both lowering the opening by one grid point and lowering it by two grid points pass the boundary margin screening, then lowering the opening by one grid point is preferred; if the net energy-saving response of one grid point is lower than the threshold while two grid points reach the threshold, then the action of lowering the opening by two grid points is selected. Taking the setpoint of the circulating fan frequency as an example, if both lowering the opening by one grid point and lowering it by two grid points are available, and both have clear energy-saving responses, then lowering the opening by one grid point is preferred to reduce fan speed fluctuations. Taking the setpoint of the pump frequency as an example, if both lowering the pump frequency and lowering the outlet pressure setpoint are available, then the action with smaller flow and pressure disturbances and more stable execution cycles is preferred. For linkage actions within the same control loop of the same equipment, before selecting the least disturbed action, it is confirmed that each target value in the action group meets its respective boundary, and the linkage relationship of the action group comes from the same loop mark in the controllable energy consumption response AI model configuration in step 102. For example, the opening degree of the gas valve and the frequency of the combustion fan in combustion control can be issued as a linkage action, but both must meet the corresponding opening degree, frequency, furnace pressure and temperature boundaries; the reduction of the frequency of the circulating fan and the fine adjustment of the opening degree of the dehumidification valve in drying control can form a linkage action within the same control loop, but the upper limit of outlet moisture, the minimum air volume requirement and the equipment current boundary must be met at the same time.

[0066] After the target control action is determined, this step converts it into a control command that the field control system can recognize. The control command includes the target control action's action object, target value, write address, write time, action direction, and adjustment range. The write address comes from the PLC variable address, DCS tag number, SCADA tag, or inverter communication address corresponding to the action object in the candidate parameter tuning action record; the target value comes from the target value in the candidate parameter tuning actions; the write time uses the allowed write time of the current control cycle; the action direction and adjustment range are recorded with the instruction. For PLC scenarios, the target control action is written to the corresponding register or variable address; for DCS scenarios, the target control action is written to the corresponding setpoint tag number; for inverter scenarios, the target control action is written to the frequency setpoint register; for valve actuator scenarios, the target control action is written to the opening setpoint address. Before writing, the target value is confirmed to be within the control system's allowed range according to the parameter boundaries in the candidate action record, and then the field control system executes the target control action. Taking the drying process as an example, if the target control action is to lower the setpoint of the circulating fan frequency by one allowable grid point, the target frequency is written to the frequency setpoint register of the circulating fan inverter. Taking the pumping process as an example, if the target control action is to slightly lower the setpoint of the pump frequency, the target frequency is written to the frequency setpoint register of the pump inverter. Taking the heating furnace as an example, if the target control action is to slightly lower the furnace temperature setpoint, the target temperature is written to the DCS furnace temperature setpoint tag. After the control command is issued, this step outputs the target control action, completing the on-site execution stage of real-time optimized control.

[0067] After the control command is written, the system reads the corresponding PLC variables, DCS tag number, SCADA tag, inverter feedback register, or actuator feedback value to confirm whether the target value takes effect within the allowed confirmation time. If the feedback value reaches the target value or enters the allowable deviation range, the target control action is recorded as executed; if the writing fails, feedback times out, the feedback value deviates from the target value, or the actuator returns to a fault state, the execution record of this target control action is canceled, and the control settings before the writing are maintained or restored. This confirmation process ensures that the target control action not only has model evaluation results but also a closed-loop feedback from the field control system.

[0068] This step outputs a result: the target control action. This target control action is obtained from candidate actions in the candidate parameter tuning action set after virtual execution evaluation, hard constraint screening, energy-saving effectiveness screening, boundary margin screening, and low-disturbance optimization. It includes the action object, target value, write address, write time, action direction, and adjustment range. The action object of the target control action comes from the controllable process parameter fields marked in step 101 and solidified by the controllable energy consumption response AI model in step 102. The target value comes from the candidate parameter tuning action set formed in step 103. The predicted energy-saving response comes from the candidate action attributes recorded in step 103. The associated necessary state parameters and associated constraint indicators come from the candidate action records. This target control action is issued to the PLC, DCS, SCADA, frequency converter, or valve actuator for execution, enabling the candidate action to complete the transformation from model evaluation object to field control command. When no candidate action passes the screening or the write confirmation fails, this step outputs the result of maintaining the current setting or not executing the target action, and records the reason for not executing. When the target control action is written and confirmed, this step records the target value and feedback value after execution. Through this step, the candidate parameter tuning action set formed in step 103 is further transformed into a target control action that meets the current operating condition boundary, has energy-saving benefits, and can be written into the field control system, thus completing the execution point of the refined energy consumption modeling and real-time optimization control method based on process parameters.

[0069] In one specific embodiment, as shown in the table below, this method is applied to the energy consumption optimization control of the hot air drying process. The field control system records the total energy consumption, the setpoint of the circulating fan frequency, the setpoint of the drying temperature, the detected outlet moisture content, the circulating fan current, and the detection results of the candidate action responses for the corresponding control cycle within the same statistical period. The total energy consumption and the controllable energy consumption response output by the model both use the same energy consumption statistical caliber, and the target values ​​of the candidate actions are all within the allowable write range on-site. This embodiment selects a training sample where the controllable parameter changes, a segment where the controllable parameter remains stable, and two candidate parameter tuning actions for recalculation to illustrate that the controllable energy consumption response AI model training constraints in step 102 and the candidate parameter tuning action screening process in step 103 can form verifiable intermediate results.

[0070] As can be seen from the above recalculation, on the training side, the total energy consumption target in the samples with controllable parameter changes is... There are calculable residuals between the controllable energy consumption response branch and the state energy consumption background branch, and the response consistency term reflects the actual energy consumption change. Controllable response of model output Constraints are imposed, so the model training result does not simply output an uninterpretable total energy consumption prediction; in the stable segment of controllable parameters, the static stability suppression term reduces invalid fluctuations in the controllable response branch when parameters do not change effectively, thereby enhancing the engineering interpretability of the output of step 102. On the control side, the frequency reduction action of the circulating fan... To achieve the minimum effective predicted energy-saving response, the drying temperature setpoint is lowered. If the threshold is not met, it means that step 103 can perform recalculation screening of different candidate actions based on the same benchmark operating condition and exclude actions with insufficient energy-saving response from the priority candidates.

[0071] Figure 4 This diagram illustrates the recalculation relationship between the model training side and the real-time optimization side in this embodiment. The left side corresponds to the relationship between the total energy consumption target, the controllable energy consumption response branch, the state energy consumption background branch, the response consistency term, and the static stability suppression term in step 102. The right side corresponds to the relationship between the baseline controllable energy consumption response, the controllable energy consumption response after the candidate action, the minimum effective predicted energy-saving response threshold, and the screening conclusion in step 103. This diagram is generated by recalculating the values ​​in the test records of this embodiment according to the calculation process disclosed in the text. It is used to help explain the source of intermediate variables, the destination of variables, and the judgment basis for retaining or eliminating candidate actions in this method.

[0072] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for refined energy consumption modeling and real-time optimization control based on process parameters, characterized in that, include: Acquire field location data, categorize the field location data into controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators, and organize them into a process parameter sample set according to process time slices. The process parameter sample set records the correspondence between the parameter action time slice of the controllable process parameter and the subsequent stable energy consumption result time slice. A controllable energy consumption response artificial intelligence model is trained based on the process parameter sample set. The controllable energy consumption response artificial intelligence model includes a controllable process parameter input channel, a necessary state parameter input channel, a state modulation layer, a controllable energy consumption response branch, and a state energy consumption background branch. The controllable energy consumption response artificial intelligence model also includes a linear mapping layer and a time-series modeling layer; The controllable process parameter input channel and the necessary state parameter input channel are respectively connected to the linear mapping layer and the time series modeling layer to output controllable features and state features respectively; The state modulation layer inputs the state features into the fully connected layer and processes them through an activation function to obtain modulation coefficients. The modulation coefficients are then applied element by element to the controllable features to obtain the state-modulated controllable response features. The controllable energy consumption response branch receives the controllable response features and outputs the controllable energy consumption response results, and the state energy consumption background branch receives the state features and outputs the energy consumption background results. The current operating condition data is input into the controllable energy consumption response artificial intelligence model for response detection, and a candidate parameter tuning action set is generated. The candidate parameter tuning action set includes at least one candidate parameter tuning action and the predicted energy-saving response of the at least one candidate parameter tuning action. A virtual execution evaluation is performed on at least one candidate parameter tuning action in the candidate parameter tuning action set to obtain an evaluation result; based on the evaluation result, the at least one candidate parameter tuning action is screened to determine the target control action, and the target control action is sent to the field control system.

2. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The classification of the field location data into controllable process parameters, necessary state parameters, energy consumption data, and constraint indicators includes: The set value or execution instruction with write permission to the control system is marked as the controllable process parameter; The sensor feedback or device status word that is in the same device or the same control loop as the controllable process parameter is marked as the necessary status parameter. The energy metering results of the equipment whose metering range covers the controllable process parameters are marked as the energy consumption data; The process configuration values ​​or limit table data used to limit the controllable process parameters or the target control actions are marked as the constraint indicators.

3. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The process parameter sample set records the correspondence between the time slices of the controllable process parameters and the time slices of the subsequent stable energy consumption results, including: When the controllable process parameter changes stably within the time slice of the parameter action, the energy consumption result in the subsequent time slice is selected as the energy consumption response result corresponding to the subsequent stable energy consumption result time slice according to the process stabilization cycle. When energy consumption response occurs in subsequent time slices after the same controllable process parameter changes, the average energy consumption of the first time slice after entering the stable range or the stable time slice is selected as the energy consumption response result.

4. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The joint loss used to train the controllable energy consumption response artificial intelligence model includes a synthetic energy consumption error term, a controllable response consistency term, and a static stability suppression term; The synthesized energy consumption error term is the error between the sum of the controllable energy consumption response result and the energy consumption background result and the energy consumption data; The controllable response consistency term is the error between the output difference of the controllable energy consumption response branch before and after parameter change and the actual energy consumption change result. The static stability suppression term is a penalty term that constrains the fluctuation of the controllable energy consumption response branch when the controllable process parameters remain stable.

5. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The step of inputting current operating condition data into the controllable energy consumption response artificial intelligence model for response detection and generating a candidate parameter tuning action set includes: Based on the write granularity and parameter upper and lower limits allowed by the control system, candidate target values ​​are generated for the controllable process parameters, and the candidate parameter tuning action is constructed, which includes the action object, the current value, and the candidate target value. The current operating condition data is input into the controllable energy consumption response artificial intelligence model to obtain the baseline controllable energy consumption response result; Replace the values ​​of the controllable process parameters in the current operating condition data with the values ​​of the candidate targets, while keeping the necessary state parameters in the current operating condition data unchanged, and input them into the controllable energy consumption response artificial intelligence model to obtain the controllable energy consumption response results after the action; The predicted energy-saving response is calculated based on the difference between the baseline controllable energy consumption response and the controllable energy consumption response after the action. Based on the constraint index, the candidate parameter tuning actions are pruned to retain those that meet the writing conditions, forming the candidate parameter tuning action set.

6. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The virtual execution evaluation of at least one candidate parameter tuning action in the candidate parameter tuning action set includes: For each candidate parameter tuning action in the candidate parameter tuning action set, replace the value of the corresponding controllable process parameter in the current operating condition data with the target value of the candidate parameter tuning action to construct a virtual operating condition. The virtual operating condition is input into the controllable energy consumption response artificial intelligence model to obtain the controllable energy consumption response result after the action of the candidate parameter tuning action and the state energy consumption background result after the action. Based on the controllable energy consumption response results after the action and the state energy consumption background results after the action, the effectiveness of the candidate parameter tuning action in the current operating context is evaluated.

7. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The step of screening the at least one candidate parameter tuning action based on the evaluation result to determine the target control action includes: Hard constraint screening is performed on the at least one candidate parameter tuning action to eliminate candidate parameter tuning actions that do not meet the upper and lower limits of parameters or correlation constraint indicators; The candidate parameter tuning actions retained after hard constraint screening are screened for energy-saving effectiveness, and candidate parameter tuning actions whose predicted energy-saving response is lower than the minimum effective predicted energy-saving response are eliminated. After screening for energy-saving effectiveness, the remaining candidate parameter tuning actions are screened for boundary margins, and candidate parameter tuning actions with mass margin, safety margin, or cycle time margin lower than the corresponding minimum margin requirement are eliminated. After screening the boundary margin, the candidate parameter tuning actions are selected with low disturbance, and the candidate parameter tuning action with the smallest adjustment amplitude or whose action direction is consistent with the previous control cycle is determined as the target control action.

8. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, The step of sending the target control action to the field control system includes: The target control action is converted into a control instruction, and the control instruction includes the action object, target value, write address and write time of the target control action; Write the control command into the corresponding register or tag number of the field control system; Read the actuator feedback value. When the actuator feedback value reaches the target value or enters the allowable deviation range within the allowable confirmation time, confirm that the target control action has been executed. When writing fails or the actuator feedback value deviates from the target value, the execution record of the target control action is canceled and the control settings before writing are restored.

9. The method for refined energy consumption modeling and real-time optimization control based on process parameters according to claim 1, characterized in that, Each sample record in the process parameter sample set contains the controllable process parameters, the necessary state parameters, the energy consumption data, and the constraint indicators within the same process time slice; When the controllable process parameter changes during the process time slice, the control system confirms the effect and maintains a stable set value as the controllable process parameter in the sample record. When there is a response lag in the process, the sample record saves the lag window and the source of values ​​for the time slice of the controllable process parameters and the time slice of the subsequent stable energy consumption results.

Images