A Deep Learning-Based Multi-Equipment Energy Consumption Optimization Method for Air Separation Units

By using an improved Deep Koopman neural network based on deep learning, energy consumption optimization of multiple devices in air separation units is achieved. This solves the problems of insufficient energy consumption prediction accuracy and unclear control boundaries in existing technologies, improves the accuracy of energy consumption prediction and the safety of control sequences, and clarifies the impact of local energy consumption increases on overall energy saving.

CN123116803APending Publication Date: 2026-11-17HANGZHOU OXYGEN PLANT GRP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611279055.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing energy consumption control methods for air separation units lack sufficient prediction accuracy during load changes, lack intersection of equipment control boundaries and implicit state reachability screening, making it difficult to identify the offsetting effect of local energy consumption increases on overall energy saving results. Furthermore, multi-equipment collaborative optimization focuses on total energy consumption changes while ignoring the marginal energy saving contribution of individual equipment.

Method used

By employing a deep learning-based approach and constructing an improved Deep Koopman neural network, we can achieve decoupling of production and demand loads, energy quality duality verification, and load manifold correction. This generates equipment control boundaries, performs candidate control sequence screening, and energy-saving contribution analysis, thereby improving the accuracy of energy consumption prediction and control safety.

Benefits of technology

It improves the accuracy of energy consumption prediction and cross-condition adaptability of air separation units under steady-state, load increase and load decrease conditions, ensures the safety and executability of control sequences, clarifies the offsetting effect of local consumption increase on overall energy saving results, and has high-precision and clear multi-equipment energy saving decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116803A_ABST
    Figure CN123116803A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based method for optimizing the energy consumption of multiple devices in an air separation unit, relating to the field of deep learning technology. The method includes: preprocessing multi-source operating data to generate standardized air separation operating data; dividing load conditions and establishing energy consumption baselines to generate baseline energy consumption data; performing dual-tower main cooling latent heat pairing to generate dual-tower main cooling latent heat pairing data; generating air separation unit prediction data through an improved Deep Koopman neural network; determining control boundaries and perturbing control variables to generate candidate multi-device control data; performing candidate control reachability screening to generate reachable candidate control data; analyzing the energy-saving contributions of multiple devices to generate multi-device marginal energy-saving contribution data; and sorting and selecting control sequences to generate energy consumption optimization results. This invention achieves multi-device collaborative control under load variation conditions, reducing the overall energy consumption of the air separation unit and the imbalance of the dual-tower main cooling latent heat, and improving the reachability of control sequences.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a deep learning-based method for optimizing the energy consumption of multiple devices in an air separation unit. Background Technology

[0002] With the increasing demand for oxygen, nitrogen, and argon in metallurgical, chemical, and industrial gas supply scenarios, air separation units are gradually developing towards larger-scale, continuous, and low-energy-consumption operation. Air separation units involve multiple stages, including air compression, pre-cooling and purification, expansion refrigeration, distillation separation, and product compression. There is a strong coupling between equipment power, process conditions, and product yield and quality. Current energy consumption control mainly relies on DCS setpoints, operational experience, and unit product power consumption indicators. Some solutions use regression models or neural networks to predict energy consumption and then adjust equipment operating parameters based on the prediction results.

[0003] Existing methods often establish energy consumption baselines based on fixed load ranges, failing to adequately characterize energy consumption changes during load increases, decreases, and demand fluctuations. Product supply load and equipment energy efficiency status are often intertwined in the prediction representation. In the dual-tower main cooling system, nitrogen condensation and liquid oxygen evaporation involve heat transfer delays and latent heat imbalances; using only temperature, pressure, and flow rate data is insufficient to reflect the thermo-mass coupling state of the main cooling system. Existing prediction models lack consistency verification of energy and material transfer; state manifold shifts caused by load changes reduce prediction accuracy across operating conditions. Candidate control sequences lack intersection of equipment control boundaries and latent state reachability screening, leading to issues such as exceeding safety limits or failing to reach the target operating condition. Multi-equipment collaborative optimization focuses on total energy consumption changes, lacking analysis of the marginal energy-saving contribution of individual equipment, making it difficult to identify the offsetting effect of localized energy increases on overall energy-saving results.

[0004] Therefore, how to provide a deep learning-based method for optimizing the energy consumption of multiple devices in air separation units is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for optimizing the energy consumption of multiple devices in an air separation unit. This invention integrates the baseline energy consumption under operating conditions, the pairing of latent heat of the main cooling towers, and an improved Deep Koopman neural network to complete the decoupling of production and demand loads, energy quality duality verification, and load manifold correction. It also performs candidate control boundary intersection, reachability screening, and energy-saving contribution analysis, and has the advantages of accurate energy consumption prediction, safe control, stable cross-operating condition adaptation, and clear energy-saving contribution.

[0006] According to an embodiment of the present invention, a deep learning-based method for optimizing the energy consumption of multiple devices in an air separation unit includes: Collect and preprocess multi-source operational data to generate standardized space separation operational data; Based on standardized air separation operation data, load conditions are divided and an energy consumption baseline model is established to generate baseline energy consumption data for the operating conditions. Perform dual-tower main cold latent heat pairing on standardized air separation operation data, calculate the dual-tower main cold phase change latent heat and latent heat loss, and generate dual-tower main cold latent heat pairing data. An improved Deep Koopman neural network with an orthogonal decoupling layer for production and demand load, a dual verification layer for energy and quality operators, and a load manifold curvature correction layer was constructed. Standardized air separation operation data and dual-tower main cooling latent heat pairing data were input to generate air separation unit prediction data. Based on the predicted data of the air separation unit, the equipment control boundary is obtained by intersection and the adjustable control quantity is disturbed to generate candidate multi-equipment control data. Perform candidate control reachability screening on candidate multi-device control data, filter out unreachable control sequences based on state evolution operators and control action operators, and generate reachable candidate control data; Based on the baseline energy consumption data of the operating condition, a multi-device energy-saving contribution analysis is performed on the achievable candidate control data to generate multi-device marginal energy-saving contribution data; Based on the marginal energy-saving contribution data of multiple devices, the achievable candidate control data are sorted, a control sequence is selected, and energy consumption optimization results are generated.

[0007] Optionally, the multi-source operational data includes: Equipment operating status data, control action data, control configuration data, safe operation data, energy consumption metering data, and product production, demand, and quality data; Equipment operating status data includes: temperature, pressure, flow rate, liquid level, speed, current, and valve position; Control action data includes: control setpoint, control loop output value, guide vane opening, valve opening, pump speed, and expander speed; Control configuration data includes: upper limit of control quantity, lower limit of control quantity, allowable rate of change, and interlocking restrictions; Safe operation data includes: alarm status, interlock status, anti-surge status, and equipment safety limits; Energy consumption metering data includes: cumulative power consumption of high-voltage incoming lines, equipment power, drive motor current, and time-of-use energy consumption; Product production, demand, and quality data include: product flow rate, cumulative output, demand, product purity, carbon dioxide content, and pressure nitrogen trace oxygen content for oxygen, nitrogen, and argon.

[0008] Optionally, generating standardized space separation operation data includes: Data on equipment operation status, control actions, control configuration, and safe operation is collected from the DCS; energy consumption data is collected from the energy consumption metering terminal; and product production, demand, and quality data are collected from the product metering terminal to generate multi-source operation data. Based on the DCS timestamp, multi-source operation data is aggregated and aligned in a one-minute cycle, missing values ​​are linearly interpolated, out-of-bounds values ​​are corrected according to control configuration data, and cleaning operation data is generated. Remove duplicate and highly correlated redundant measurement points from the cleaning operation data, retain control configuration data and safety operation data, normalize the values, and generate standardized space separation operation data.

[0009] Optionally, the generation of baseline energy consumption data includes: Product flow and energy consumption metering data are extracted from standardized air separation operation data. Production conversion, periodic aggregation and unit energy consumption conversion are performed on the product flow and energy consumption metering data to generate oxygen-converted comprehensive production and unit oxygen-converted comprehensive power consumption. Based on the changing direction and continuous state of the total oxygen production, the load change process is identified, and the steady-state operating condition, the load increase operating condition, and the load decrease operating condition are divided. The division results are correlated with the unit oxygen production power consumption to generate load operating condition data. Safe operation data is extracted from standardized air separation operation data, correlated with load condition data and selected as effective operation samples, and the energy consumption baseline model is trained using the effective operation samples. Baseline energy consumption is calculated according to load condition to generate baseline energy consumption data for operating conditions.

[0010] Optionally, generating the dual-tower primary cold latent heat pairing data includes: Operating status data of the dual-tower main refrigeration equipment were extracted from standardized air separation operation data to construct a dual-tower main refrigeration status sequence. Read the nitrogen and oxygen working fluid property data from the air separation unit process design document, match the dual-tower main cooling state sequence with the nitrogen and oxygen working fluid property data according to temperature and pressure, determine the nitrogen condensation phase boundary and the liquid oxygen evaporation phase boundary, and calculate and generate the dual-tower main cooling phase change latent heat sequence. A time-shift scan is performed on the latent heat of phase change of the dual-tower main cooler. The response delay of the main cooler is determined based on the time-shift correlation value and the direction of change of the main cooler liquid level. The latent heat of phase change sequences on both sides are aligned according to the response delay of the main cooler to generate a pairing sequence of latent heat of the dual-tower main cooler. The latent heat of nitrogen condensation and the latent heat of liquid oxygen evaporation in the latent heat pairing sequence of the dual-tower main cooling system are calculated by mean and difference to generate the latent heat of phase change and latent heat loss of the dual-tower main cooling system. These are then associated with the state sequence of the dual-tower main cooling system by timestamp to generate the latent heat pairing data of the dual-tower main cooling system.

[0011] Optionally, the generation of air separation unit prediction data includes: An improved Deep Koopman neural network is constructed, which includes an orthogonal decoupling layer for production, demand and load, a dual verification layer for energy and mass dual operators, and a load manifold curvature correction layer. The standardized air separation operation data and the dual-tower main cold latent heat pairing data are converted into the initial hidden state by the encoder. The production demand load orthogonal decoupling layer performs orthogonal separation of production demand load and energy efficiency state on the initial hidden state, generating the production demand load hidden state and the energy efficiency hidden state, and forming the orthogonal decoupling hidden state data. Based on the control action data, a controlled Koopman evolution is performed on the orthogonally decoupled hidden state data, and a predicted hidden state is generated through the state evolution operator and the control action operator. Based on standardized air separation operation data and dual tower main cold latent heat pairing data, the dual-operator dual verification layer performs bidirectional verification of energy transfer and material transfer on the predicted hidden state, adjusts the predicted hidden state according to the bidirectional verification deviation, and generates energy and quality verification hidden state data. The load manifold curvature correction layer performs local curvature identification and tangential state correction on the energy quality verification hidden state data based on the changes in the hidden state of production and demand load at adjacent time points, and generates curvature correction hidden state data. The curvature correction latent state data is input into the decoder, state decoding is performed, and air separation unit prediction data is generated. An improved Deep Koopman neural network was trained by dividing standardized space separation operation data and dual-tower main cold latent heat pairing data into training, validation, and test sets according to time sequence. The training set was input into the improved Deep Koopman neural network to calculate prediction error, orthogonal decoupling error, energy-mass duality error, and curvature correction error, and these were added together to generate a joint loss. The network parameters were updated in reverse based on the joint loss. The joint loss was calculated based on the validation set. The parameter update was terminated when the absolute value of the difference between five consecutive adjacent joint losses was less than 0.001. The prediction error was calculated using the test set.

[0012] Optionally, generating candidate multi-device control data includes: Control configuration data, safety operation data, and control action data are extracted from standardized air separation operation data, and configuration safety constraint boundaries are constructed based on the control configuration data and safety operation data. The predicted equipment operating status is extracted from the predicted data of the air separation unit. The corresponding control quantity range is calculated based on the control action operator to generate the predicted control boundary. The intersection of the configured safety constraint boundary and the predicted control boundary is calculated according to the equipment identification and the predicted time to generate the equipment control boundary. Extract the current value of the adjustable control quantity from the control action data, perform bidirectional disturbance on the adjustable control quantity within the equipment control boundary, combine the disturbance results of different equipment according to the predicted time, and generate candidate multi-equipment control data.

[0013] Optionally, generating reachable candidate control data includes: Candidate control sequences are extracted from candidate multi-device control data, associated with air separation unit prediction data according to equipment identification and prediction time, and generated as initial hidden state and target hidden state by encoder. The predicted hidden state trajectory is generated recursively using the initial hidden state and candidate control sequence as inputs through state evolution operator and control action operator; Based on the predicted hidden state trajectory, state evolution operator, and control action operator, a reachable hidden domain is constructed. The target hidden state is projected onto the reachable hidden domain, and the difference norm between the target hidden state and the projection result is calculated to generate the reachable residual. Candidate control sequences with reachable residuals greater than the preset reachable residual threshold are filtered out, and the remaining candidate control sequences are collected to generate reachable candidate control data.

[0014] Optionally, generating multi-device marginal energy-saving contribution data includes: Based on reachable candidate control data, the energy consumption of candidate control is predicted by an improved Deep Koopman neural network. The baseline energy consumption data of the load condition and the prediction time are associated with the baseline energy consumption to generate the candidate total energy saving. Remove the adjustable control quantity disturbance of each device one by one, fix the adjustable control quantity of other devices, and generate a single device control withdrawal sequence; The single-device withdrawal control sequence is input into an improved Deep Koopman neural network to generate withdrawal control energy consumption. The candidate control energy consumption is subtracted from the withdrawal control energy consumption to generate the marginal energy saving contribution of the device. Collect marginal energy-saving contributions and candidate total energy savings of equipment according to equipment identification, and generate multi-equipment marginal energy-saving contribution data.

[0015] Optionally, the generation of energy consumption optimization results includes: By associating the marginal energy-saving contribution data of multiple devices with the achievable candidate control data according to the candidate control sequence, the candidate total energy saving, the marginal energy-saving contribution of the devices, and the disturbance value of the adjustable control quantity are extracted. The absolute values ​​of the marginal energy-saving contributions of equipment with negative marginal energy-saving contributions are summed, and the summation is subtracted from the candidate total energy saving to generate a collaborative energy-saving score. Candidate control sequences are sorted in descending order according to their collaborative energy-saving scores. When the collaborative energy-saving scores are the same, they are sorted in ascending order according to the sum of the absolute values ​​of the adjustable control quantity disturbances, thus generating the candidate control sequence sorting results. The first candidate control sequence is selected from the candidate control sequence ranking results, and the corresponding candidate total energy saving and equipment marginal energy saving contribution are associated to generate energy consumption optimization results.

[0016] The beneficial effects of this invention are: This invention proposes a deep learning-based multi-equipment energy consumption optimization method for air separation units. By constructing a multi-source operating data processing, load condition energy consumption baseline modeling, dual-tower main cooling latent heat pairing, air separation unit state prediction, and candidate control sequence screening process, it improves the accuracy of energy consumption prediction for air separation units under steady-state, load-increasing, and load-decreasing conditions. Compared to optimization methods relying on fixed thresholds and single energy consumption indicators, this invention introduces an improved Deep Koopman neural network. It utilizes an orthogonal decoupling layer to separate production / demand load from energy efficiency status, a dual-operator dual-verification layer to correct energy and material transfer deviations, and a load manifold curvature correction layer to correct hidden state shifts caused by load changes, thereby improving cross-condition prediction stability and adaptability to complex load changes.

[0017] This invention characterizes the thermo-mass coupling state between nitrogen condensation and liquid oxygen evaporation by pairing the latent heat of the dual-tower main cooling system, reducing the interference of main cooling response delay and latent heat imbalance on energy consumption prediction results. It limits the equipment control boundary by configuring safety constraint boundaries and predictive control boundaries, and filters out unreachable control sequences based on state evolution operators, control action operators, and reachable hidden domains, improving the safety and executability of candidate control sequences. Furthermore, it calculates the candidate total energy saving and the marginal energy saving contribution of each equipment through multi-equipment energy saving contribution analysis, identifying the offsetting effect of local energy consumption increases on the overall energy saving result. This invention possesses advantages such as high energy consumption prediction accuracy, strong adaptability across operating conditions, high control sequence reachability, and clear multi-equipment energy saving decisions. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based energy consumption optimization method for multiple devices in an air separation unit, as proposed in this invention. Figure 2 This is a schematic diagram of the improved Deep Koopman neural network structure in the energy consumption optimization method for multiple devices in an air separation unit based on deep learning proposed in this invention. Figure 3 This is a flowchart of the candidate multi-device control data generation and candidate control reachability screening in the deep learning-based multi-device energy consumption optimization method for air separation units proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A deep learning-based method for optimizing the energy consumption of multiple devices in an air separation unit includes: Collect and preprocess multi-source operational data to generate standardized space separation operational data; Based on standardized air separation operation data, load conditions are divided and an energy consumption baseline model is established to generate baseline energy consumption data for the operating conditions. Perform dual-tower main cold latent heat pairing on standardized air separation operation data, calculate the dual-tower main cold phase change latent heat and latent heat loss, and generate dual-tower main cold latent heat pairing data. An improved Deep Koopman neural network with an orthogonal decoupling layer for production and demand load, a dual verification layer for energy and quality operators, and a load manifold curvature correction layer was constructed. Standardized air separation operation data and dual-tower main cooling latent heat pairing data were input to generate air separation unit prediction data. Based on the predicted data of the air separation unit, the equipment control boundary is obtained by intersection and the adjustable control quantity is disturbed to generate candidate multi-equipment control data. Perform candidate control reachability screening on candidate multi-device control data, filter out unreachable control sequences based on state evolution operators and control action operators, and generate reachable candidate control data; Based on the baseline energy consumption data of the operating condition, a multi-device energy-saving contribution analysis is performed on the achievable candidate control data to generate multi-device marginal energy-saving contribution data; Based on the marginal energy-saving contribution data of multiple devices, the achievable candidate control data are sorted, a control sequence is selected, and energy consumption optimization results are generated.

[0021] In this embodiment, the multi-source operational data includes: Equipment operating status data, control action data, control configuration data, safe operation data, energy consumption metering data, and product production, demand, and quality data; Equipment operating status data includes: temperature, pressure, flow rate, liquid level, speed, current, and valve position; Control action data includes: control setpoint, control loop output value, guide vane opening, valve opening, pump speed, and expander speed; Control configuration data includes: upper limit of control quantity, lower limit of control quantity, allowable rate of change, and interlocking restrictions; Safe operation data includes: alarm status, interlock status, anti-surge status, and equipment safety limits; Energy consumption metering data includes: cumulative power consumption of high-voltage incoming lines, equipment power, drive motor current, and time-of-use energy consumption; Product production, demand, and quality data include: product flow rate, cumulative output, demand, product purity, carbon dioxide content, and pressure nitrogen trace oxygen content for oxygen, nitrogen, and argon.

[0022] In this embodiment, generating standardized space separation operation data includes: Data on equipment operation status, control actions, control configuration, and safe operation is collected from the DCS; energy consumption data is collected from the energy consumption metering terminal; and product production, demand, and quality data are collected from the product metering terminal to generate multi-source operation data. Based on the DCS timestamp, multi-source operation data is aggregated and aligned in a one-minute cycle, missing values ​​are linearly interpolated, out-of-bounds values ​​are corrected according to control configuration data, and cleaning operation data is generated. Remove duplicate and highly correlated redundant measurement points from the cleaned operational data, retain control configuration data and safety operation data, normalize the values, and generate standardized spatial separation operational data, including: Standardized air separation operation data includes: Time index data, normalized equipment operating status data, normalized control action data, control configuration data retaining original values, safety operation data retaining original values, normalized energy consumption metering data, and normalized product production, demand, and quality data; The time index data includes the DCS timestamp, data source identifier, device identifier, and measurement point number; Normalized equipment operating status data includes temperature, pressure, flow rate, liquid level, speed, current, and valve position; The normalized control action data includes control setpoints, control loop outputs, guide vane opening, valve opening, pump speed, and expander speed; Control configuration data includes upper limit of control quantity, lower limit of control quantity, allowable rate of change, and interlocking restrictions; Safe operation data includes alarm status, interlock status, anti-surge status, and equipment safety limits; Normalized energy consumption metering data includes cumulative power consumption of high-voltage incoming lines, equipment power, drive motor current, and time-of-use energy consumption; Normalized product supply, demand, and quality data include product flow rate, cumulative output, demand, product purity, carbon dioxide content, and pressure nitrogen trace oxygen content for oxygen, nitrogen, and argon.

[0023] In this embodiment, generating baseline energy consumption data for operating conditions includes: Product flow and energy consumption metering data are extracted from standardized air separation operation data. Production conversion, periodic aggregation, and unit energy consumption conversion are performed on the product flow and energy consumption metering data to generate comprehensive oxygen-converted production and comprehensive electricity consumption per unit of oxygen. The comprehensive oxygen yield and comprehensive electricity consumption per unit of oxygen are as follows: Read the product flow rate, high-voltage incoming line cumulative power consumption, and the minimum and maximum values ​​of the measurement points recorded during numerical normalization from the standardized air separation operation data, and restore the product flow rate and high-voltage incoming line cumulative power consumption to the acquired dimensions; The oxygen coefficients of each product are read from the process design documents of the air separation unit. The product flow rate is multiplied by the corresponding oxygen coefficient and then summed. The product flow rate is then aggregated in a one-hour cycle to generate the comprehensive oxygen yield. Calculate the difference between the first and last cumulative power consumption of the high-voltage incoming line within the same cycle, divide the difference by the oxygen-converted comprehensive output, and generate the unit oxygen-converted comprehensive power consumption. Based on the changing direction and continuous state of the total oxygen production, the load change process is identified, and the operating conditions are divided into steady-state, increasing load, and decreasing load conditions. The classification results are then correlated with the total electricity consumption per unit oxygen production to generate load condition data, including: The load condition data is generated as follows: The difference in oxygen-based total yield between adjacent polymerization cycles is calculated in chronological order, and the median absolute deviation of the difference sequence is set as the load change threshold. When the difference between two consecutive aggregation cycles is greater than the load change threshold, it is marked as an increase in load condition; when the difference between two consecutive aggregation cycles is less than the negative of the load change threshold, it is marked as a decrease in load condition; the remaining aggregation cycles are marked as steady-state conditions. The operating condition markers are associated with the unit oxygen-converting comprehensive power consumption according to the polymerization cycle to generate load operating condition data; Safe operation data is extracted from standardized air separation operation data, correlated with load condition data, and valid operation samples are selected. An energy consumption baseline model is trained using these valid operation samples, and baseline energy consumption is calculated according to load conditions to generate baseline energy consumption data. Generate baseline energy consumption data for operating conditions, specifically: Extract oxygen yield, product purity, equipment operating status data and control action data from effective operating samples, form baseline feature data according to timestamps, and set the unit oxygen yield power consumption corresponding to the same timestamp as the fitting target; The first 80% of the valid operating samples were assigned to the training samples in chronological order, and the last 20% of the valid operating samples were assigned to the validation samples. The average value of the unit oxygen-based comprehensive power consumption in the training samples was used as the initial baseline value. Calculate the residual between the unit oxygen-based total energy consumption of each training sample and the current baseline value. Divide the candidate groups according to the numerical order of each baseline feature. Calculate the sum of the squared differences between the residuals within each candidate group and the average residuals of the group. Select the group with the smallest sum of squared differences, add the average residuals of the corresponding group to the current baseline value, and update the baseline mapping parameters. After each round of parameter updates, the sum of squares of the differences between the predicted values ​​of the validation samples and the comprehensive energy consumption per unit oxygen is calculated, and the sum of squares is divided by the number of validation samples to generate the validation mean square error. The parameter update is terminated when the absolute value of the difference between the validation mean square errors of 5 consecutive rounds is less than 0.001, and the baseline mapping parameters with the smallest validation mean square error are retained. The energy consumption baseline model is generated by combining the baseline feature fields, initial baseline values, selected grouping method, average group residuals, and baseline mapping parameters. The baseline characteristic data corresponding to steady-state operating conditions, load increase operating conditions, and load decrease operating conditions are respectively input into the energy consumption baseline model to calculate the predicted unit oxygen-based comprehensive power consumption. The average value of the predicted unit oxygen-based comprehensive power consumption within the same load condition is calculated to generate the baseline energy consumption for each load condition. This baseline energy consumption is then associated with the load condition identifier to generate the operating condition baseline energy consumption data.

[0024] In this embodiment, generating the dual-tower primary cold latent heat pairing data includes: Operating status data of the dual-tower main refrigeration unit was extracted from standardized air separation operation data to construct a dual-tower main refrigeration status sequence, in which: The dual-tower master cold state sequence is constructed as follows: Read the DCS timestamp, equipment identifier, and measurement point number from the time index data in the standardized air separation operation data, and locate the operation measurement points corresponding to the nitrogen condensation area, liquid oxygen evaporation area, and main cooling liquid level of the dual tower main cooler according to the equipment identifier and measurement point number; Extract the temperature, pressure, flow rate and liquid level corresponding to the operating measurement points from the normalized equipment operating status data, read the minimum and maximum values ​​of the measurement points recorded during the normalization of the values, and restore the extracted data to the collected dimensions. The arithmetic mean of the temperature measurement points in the nitrogen condensation area under the same equipment identifier and the same DCS timestamp is calculated to generate the nitrogen condensation side temperature. The arithmetic mean of the temperature measurement points in the liquid oxygen evaporation area is calculated to generate the liquid oxygen evaporation side temperature. The nitrogen condensation side temperature, liquid oxygen evaporation side temperature, and corresponding pressure, flow rate, and main cooling liquid level are associated with the equipment identifier and DCS timestamp, and arranged sequentially according to the DCS timestamp to generate a dual-tower main cooling status sequence. The nitrogen and oxygen working fluid property data were read from the air separation unit process design documents. The dual-tower main cooling state sequence was matched with the nitrogen and oxygen working fluid property data according to temperature and pressure to determine the nitrogen condensation phase boundary and the liquid oxygen evaporation phase boundary. The latent heat of phase change sequence of the dual-tower main cooling system was calculated and generated, where: The latent heat sequence of the dual-tower primary cold phase transition is calculated and generated as follows: The nitrogen and oxygen working fluid property data are obtained from the process design documents delivered by the air separation unit equipment design unit. The working fluid name, temperature, pressure, saturated gas phase enthalpy, saturated liquid phase enthalpy, standard state gas density and saturated liquid density are read from the nitrogen and oxygen working fluid property table to form nitrogen property node set and oxygen property node set. Extract the equipment identifier, DCS timestamp, nitrogen condenser side temperature, nitrogen condenser side pressure, liquid oxygen evaporator side temperature, liquid oxygen evaporator side pressure, and working fluid flow rates on both sides from the dual-tower main cooling state sequence; Calculate the normalized difference between the working fluid temperature and the corresponding physical property node temperature on both sides, and the normalized difference between the working fluid pressure and the corresponding physical property node pressure on both sides. Add the normalized difference of temperature and the normalized difference of pressure to generate the nitrogen temperature and pressure matching deviation and the oxygen temperature and pressure matching deviation. The nitrogen property node with the smallest temperature and pressure matching deviation was selected as the nitrogen condensation phase boundary, and the oxygen property node with the smallest temperature and pressure matching deviation was selected as the liquid oxygen evaporation phase boundary. The saturated gas phase enthalpy, saturated liquid phase enthalpy, and working fluid density corresponding to the two phase boundaries were read. Subtracting the saturated liquid phase enthalpy from the saturated gas phase enthalpy respectively yields the unit latent heat of nitrogen condensation and the unit latent heat of liquid oxygen evaporation; Based on the flow measurement unit, the flow rate on the nitrogen condensation side and the flow rate on the liquid oxygen evaporation side are converted into the working fluid mass flow rate; The mass flow rates of the working fluids on both sides are converted into working fluid mass per minute, and multiplied with the corresponding unit latent heat to generate the latent heat of nitrogen condensation and the latent heat of liquid oxygen evaporation. The latent heat of nitrogen condensation and the latent heat of liquid oxygen vaporization are associated with equipment identification and DCS timestamp, and arranged sequentially according to DCS timestamps to generate a dual-tower main cold phase change latent heat sequence. A time-shift scan is performed on the latent heat of phase change sequence of the dual-tower main cooler. The response delay of the main cooler is determined based on the time-shift correlation value and the direction of the main cooler liquid level change. The latent heat of phase change sequences on both sides are aligned according to the main cooler response delay to generate a paired sequence of latent heat of the dual-tower main cooler, wherein: Generate the dual-tower master cold latent heat pairing sequence as follows: The latent heat of phase change of the dual-tower main cooling system is associated with the dual-tower main cooling system state sequence according to the equipment identifier and DCS timestamp. The main cooling liquid level is extracted and the liquid level difference between adjacent times is calculated to divide the liquid level continuously rising interval and the liquid level continuously falling interval. Within each continuous liquid level change range, the liquid oxygen evaporation latent heat sequence is moved in 1-minute increments. The correlation value of the phase change latent heat on both sides is calculated for each time shift. The positive latent heat difference corresponds to a decrease in liquid level, and the negative latent heat difference corresponds to an increase in liquid level, which are used as directional matching conditions. Select the time shift with the largest correlation value from the time shifts that meet the direction matching conditions, calculate the median of the selection results for each interval, and generate the main cold response delay; The liquid oxygen evaporation latent heat sequence is moved according to the main cold response delay. The time periods on both sides of the sequence that do not overlap are deleted. The sequence is then paired according to the equipment identifier and DCS timestamp to generate a pairing sequence of the dual tower main cold latent heat. The latent heat of nitrogen condensation and the latent heat of liquid oxygen evaporation in the paired sequence of the dual-tower main cooling system are calculated by mean and difference to generate the latent heat of phase change and latent heat loss of the dual-tower main cooling system. These are then correlated with the dual-tower main cooling system state sequence by timestamp to generate paired latent heat data for the dual-tower main cooling system, where: Generate the dual-tower main cold latent heat pairing data, specifically as follows: Extract equipment identifiers, DCS timestamps, nitrogen condensation latent heat, and liquid oxygen evaporation latent heat from the dual-tower main latent heat pairing sequence; Add the latent heat of nitrogen condensation and the latent heat of liquid oxygen evaporation corresponding to the same equipment identifier and the same DCS timestamp, and divide by 2 to generate the latent heat of phase change of the dual-tower main cold. Subtracting the latent heat of liquid oxygen evaporation from the latent heat of nitrogen condensation yields a latent heat loss measure. Based on the equipment identifier and DCS timestamp, the temperature, pressure, flow rate, and main cooling liquid level in the dual-tower main cooling status sequence are associated, and the latent heat of phase change, latent heat loss, and associated status data of the dual-tower main cooling are collected to generate dual-tower main cooling latent heat pairing data.

[0025] In this embodiment, generating air separation unit prediction data includes: An improved Deep Koopman neural network is constructed, which includes an orthogonal decoupling layer for production, demand, and load; a dual verification layer for energy and mass operators; and a load manifold curvature correction layer, wherein: The improved Deep Koopman neural network is constructed as follows: The original Deep Koopman neural network retains its multi-layer fully connected state encoding, Koopman A linear state evolution, Koopman B linear control action, and multi-step recursive prediction process, and defines its structure as an encoder, a state evolution operator, and a control action operator. The original process of directly splicing the original state and the encoding result to form the embedded state is replaced by an orthogonal decoupling process. An orthogonal decoupling layer of production, demand and load is set at the encoder output. Standardized air separation operation data and dual-tower main cold latent heat pairing data are received. The hidden state channels are divided according to product demand, comprehensive output and energy consumption status. Orthogonal projection is performed on the relevant components between channels to generate orthogonal decoupling hidden state data. The controlled linear evolution path, which is the sum of the encoded state after mapping by the state evolution operator and the control action data after mapping by the control action operator, is retained. The orthogonally decoupled hidden state data is then connected to the state evolution operator to generate the predicted hidden state. Disconnect the direct connection from the predicted hidden state to the state recovery end, set up an energy-mass dual operator dual verification layer after the controlled linear evolution path, establish energy transfer verification branch and material transfer verification branch, calculate bidirectional verification deviation based on standardized air separation operation data and dual tower main cold latent heat pairing data, write the deviation correction result back to the predicted hidden state, and generate energy-mass verification hidden state data. A load manifold curvature correction layer is set between the dual verification layer of the energy and mass dual operator and the state recovery end. Based on the hidden state of production and demand load at adjacent time points, the local tangential direction and curvature of the load change trajectory are identified. The hidden state of energy and mass verification is projected onto the local tangential space and the deviation component is corrected to generate curvature correction hidden state data. The state recovery projection is determined as the decoder, the curvature-corrected hidden state data is input into the decoder, and the multi-step recursive connection is retained to form an improved Deep Koopman neural network. Standardized air separation unit (ASU) operation data and dual-tower main cooling latent heat pairing data are converted into an initial latent state via an encoder. The production-demand-load orthogonal decoupling layer performs orthogonal separation of production-demand-load and energy efficiency states on the initial latent state, generating production-demand-load latent states and energy efficiency latent states, which together form orthogonal decoupled latent state data, wherein: The production-demand-load orthogonal decoupling layer specifically includes: Production and demand characteristic slot area: Receive the initial hidden state output by the encoder, and write the corresponding representations of product flow, cumulative output and demand into the production and demand load slot according to the mapping position of the encoder input field. Write the corresponding representations of energy consumption metering data, equipment operation status data and dual tower main cold latent heat pairing data into the energy efficiency status slot. Construction of branches for load orthogonal basis: Calculate the production-demand difference between the flow rate and demand of each product, extract the change direction of the comprehensive oxygen production at adjacent time moments, map the production-demand difference and change direction to the initial hidden state dimension, and construct the production-demand-load orthogonal basis; Orthogonal residual projection kernel: Project the initial implicit state in the production-demand-load slot to the production-demand-load orthogonal basis to generate the production-demand-load implicit state; project the initial implicit state in the energy efficiency state slot to the orthogonal complement space of the production-demand-load orthogonal basis to generate the energy efficiency implicit state. Decoupling representation aggregation end: Perform dimensional alignment on the hidden states of production and demand load and energy efficiency, arrange them according to DCS timestamp and equipment identifier, and generate orthogonal decoupling hidden state data; The orthogonally decoupled hidden state data is composed as follows: Extract equipment identifiers, DCS timestamps, equipment operating status data, control action data, energy consumption metering data, and product production, demand, and quality data from standardized air separation operation data. Extract the latent heat of phase change and latent heat loss of the dual-tower main cooling system from the dual-tower main cooling system latent heat pairing data. Align the data with equipment identifiers and DCS timestamps and assemble coded input data according to a fixed field order. The encoder performs multi-layer fully connected mapping and non-linear activation on the encoded input data, mapping the input fields to a unified hidden state dimension to generate the initial hidden state; The production and demand characteristic sub-slot extracts the product flow rate, cumulative output and demand of oxygen, nitrogen and argon from the product production and demand and quality data, generates a production and demand query characterization through independent fully connected mapping, and generates an energy efficiency query characterization from equipment operation status data, control action data, energy consumption metering data, latent heat of phase change of the dual tower main cold phase change and latent heat loss measurement. The similarity between each channel in the initial hidden state and the production and demand query representation and the energy efficiency query representation is calculated in the production and demand characteristic slot area. Channels with production and demand similarity not less than energy efficiency similarity are written into the production and demand load slot, and the remaining channels are written into the energy efficiency status slot. The load orthogonal basis constructs branches to calculate the product flow rate and demand difference of oxygen, nitrogen and argon respectively, calculates the cumulative output change under adjacent DCS timestamps, and maps the production and demand difference and the cumulative output change to the initial hidden state dimension to form the product production and demand direction vector and the load change direction vector. The load orthogonal basis is constructed by sequentially eliminating parallel components between direction vectors and performing length normalization to generate an orthogonal basis for production and demand loads. The orthogonal residual projection kernel projects the latent states in the production and demand load slots item by item along the orthogonal basis of production and demand load, adds up the projection results to generate the latent states of production and demand load, and subtracts the projection components along the direction of the orthogonal basis of production and demand load from the latent states in the energy efficiency state slots to generate the energy efficiency latent states orthogonal to the latent states of production and demand load. The decoupled characterization aggregation end performs dimensional calibration on the hidden states of production and demand load and energy efficiency respectively, records the channel positions of the two types of hidden states, arranges them according to equipment identification and DCS timestamp, and splices the two types of hidden states to form orthogonal decoupled hidden state data. Controlled Koopman evolution is performed on the orthogonally decoupled hidden state data based on the control action data. Predicted hidden states are generated through state evolution operators and control action operators, where: Controlled Koopman evolution is performed on the orthogonally decoupled hidden state data, specifically as follows: Extract the device identifier, DCS timestamp, production and demand load hidden state, energy efficiency hidden state and corresponding channel position from the orthogonal decoupled hidden state data. Combine the two types of hidden states under the same device identifier and the same DCS timestamp according to the channel position to generate the current hidden state. Extract control setpoints, control loop outputs, guide vane openings, valve openings, pump speeds, and expander speeds from standardized air separation operation data. Associate the current hidden state with the equipment identifier and DCS timestamp, and arrange them in a fixed field order to generate the current control vector. The state evolution operator receives the current hidden state, performs a linear mapping on the current hidden state through an unbiased linear layer, and generates autonomous state evolution components without applying the current control action. The control action operator receives the current control vector and maps each control action to the current hidden state dimension through an unbiased linear layer, generating the control action component corresponding to the current control action. The autonomous state evolution component and the control action component are added channel by channel to generate the predicted latent state corresponding to the next DCS timestamp, and the channel positions of the production demand load latent state and the energy efficiency latent state are retained. The generated predicted hidden state is used as the current hidden state of the next controlled Koopman evolution. It is associated with the control action data corresponding to the next DCS timestamp. The state evolution operator mapping, control action operator mapping and channel-by-channel addition are repeatedly executed to generate the predicted hidden state corresponding to each prediction time. Arrange the predicted hidden states according to the device identifier and the predicted time to generate a sequence of predicted hidden states; The dual-operator verification layer for energy and mass transfer performs bidirectional verification of the predicted latent state based on standardized air separation operation data and dual-tower main latent heat pairing data. It adjusts the predicted latent state according to the bidirectional verification deviation, generating energy and mass transfer verification latent state data, where: The energy-mass dual operator dual verification layer specifically includes: Energy transfer calculation branch: Receives predicted latent state, energy consumption metering data and latent heat pairing data of dual tower main cooling, extracts equipment power, time-of-use energy consumption, latent heat of phase change of dual tower main cooling and latent heat loss, and generates energy transfer verification characterization. Material balance calculation branch: Receives predicted latent state and standardized air separation operation data, extracts product flow rate, demand, product purity and main cooling liquid level, calculates product production and demand changes and main cooling liquid level changes, and generates material transfer verification characterization; Dual bias cross kernel: Projects the energy transfer verification characterization and the material transfer verification characterization to the same hidden state dimension, performs forward verification from energy characterization to material characterization and reverse verification from material characterization to energy characterization, and generates bidirectional verification bias; Verification residual write-back channel: Based on the bidirectional verification deviation, determine the correction amplitude of the energy channel and material channel in the predicted hidden state, and write the correction residual back to the corresponding hidden state position; Energy quality characterization output: Perform dimension alignment and time arrangement on the hidden state after writing back to generate energy quality verification hidden state data; Generate energy quality verification hidden state data, specifically as follows: Extract equipment identifiers, prediction times, production and demand load hidden state channels, and energy efficiency hidden state channels from the predicted hidden state. Then, associate standardized air separation operation data and dual-tower main cold latent heat pairing data according to the equipment identifiers and prediction times. The energy transfer calculation branch extracts the cumulative power of the high-voltage incoming line and the equipment power from the energy consumption metering data, calculates the difference in the cumulative power of the high-voltage incoming line at adjacent times, and converts the equipment power into the equipment energy consumption for 1 minute. It also extracts the latent heat of the dual-tower main cooling phase change and the latent heat loss from the dual-tower main cooling phase change latent heat pairing data, performs unified dimension mapping on the corresponding data, and generates an energy transfer benchmark characterization. The energy transfer calculation branch extracts the predicted energy efficiency representation from the predicted hidden state energy efficiency hidden state channel, generates the predicted energy transfer representation through fully connected mapping, and performs difference calculation between the predicted energy transfer representation and the energy transfer benchmark representation to generate the energy transfer deviation. The material balance calculation branch extracts the inlet air flow and main cooling liquid level from the equipment operation status data, and extracts the product flow, demand and purity of oxygen, nitrogen and argon from the product production, demand and quality data. It calculates the difference between the inlet air flow and the total product flow, the difference between the product flow and the demand, and the change in the main cooling liquid level at adjacent times, and generates a material transfer benchmark characterization. The material balance calculation branch extracts the predicted production and demand representation from the predicted implicit state channel of production and demand load, generates the predicted material transfer representation through fully connected mapping, and performs difference calculation between the predicted material transfer representation and the material transfer benchmark representation to generate the material transfer deviation. The dual bias cross kernel maps the predicted energy transfer characterization to the material characterization dimension, performs forward verification with the predicted material transfer characterization to generate the energy-to-material verification bias, and maps the predicted material transfer characterization to the energy characterization dimension, performs reverse verification with the predicted energy transfer characterization to generate the material-to-energy verification bias. The dual deviation cross kernel performs dimensional alignment and splicing of energy transfer deviation, material transfer deviation, energy to material verification deviation, and material to energy verification deviation to generate bidirectional verification deviation. The verification residual write-back channel maps the bidirectional verification deviation to the prediction hidden state dimension. The correction component corresponding to the energy transfer deviation is written into the energy efficiency hidden state channel, and the correction component corresponding to the material transfer deviation is written into the production demand load hidden state channel to generate the post-verification prediction hidden state. The energy quality characterization output terminal performs channel position recovery on the predicted hidden state after verification, arranges it according to the device identifier and prediction time, and generates energy quality verification hidden state data. The load manifold curvature correction layer performs local curvature identification and tangential state correction on the energy quality verification latent state data based on the changes in the latent state of production and demand load at adjacent time points, generating curvature-corrected latent state data, wherein: The load manifold curvature correction layer specifically includes: Local load neighborhood index area: Receives energy quality verification hidden state data and production and demand load hidden state, extracts the hidden state of the current time and the adjacent times before and after according to the DCS timestamp, and constructs the local load neighborhood; Tangential base plane construction of branches: Calculate the direction of change of the hidden state of adjacent production and demand loads in the local load neighborhood, perform orthogonalization processing on the direction of change, and construct a local tangential base plane; Curvature offset discrimination kernel: compare the directional changes of adjacent local tangential base planes, calculate the degree of deviation of the energy quality verification hidden state relative to the current local tangential base plane, and generate a curvature offset characterization; Normal residual correction channel: Decompose the energy quality verification hidden state into tangential and normal components, adjust the normal component according to the curvature offset characterization, and recombine it with the tangential component; Curvature correction output: Perform time position recovery and dimension alignment on the recombined hidden states to generate curvature correction hidden state data; Local curvature identification and tangential state correction are performed on the latent state data of energy quality verification, specifically as follows: The local load neighborhood index area extracts equipment identifiers, prediction times, production and demand load hidden state channels, and energy efficiency hidden state channels from the energy quality verification hidden state data. It is arranged according to equipment identifiers and prediction times, and extracts the production and demand load hidden states of the current prediction time and the adjacent prediction times before and after it to construct the local load neighborhood. The local load neighborhood index area extracts the hidden state of production and demand load at the beginning of the sequence for the first three consecutive prediction times, and extracts the hidden state of production and demand load at the end of the sequence for the last three consecutive prediction times, forming the local load neighborhood corresponding to the beginning and end positions. The tangential base plane constructs a branch by subtracting the previous implicit state of production and demand load from the current implicit state of production and demand load, generating the backward load change direction, and by subtracting the current implicit state of production and demand load from the next implicit state of production and demand load, generating the forward load change direction. The tangential base plane constructs branches that perform length normalization on the forward and backward load change directions, eliminate the parallel components between them, generate mutually orthogonal local tangential base vectors, and construct local tangential base planes from the local tangential base vectors. The curvature offset discrimination kernel calculates the directional difference between the forward load change direction and the backward load change direction, generates the local load curvature, and calculates the average value of the hidden state of production and demand load in the local load neighborhood to generate the local load center. The curvature offset discrimination kernel projects the offset of the current production and demand load hidden state relative to the local load center onto the local tangential base plane to generate tangential state components. The offset that does not enter the local tangential base plane is determined as the normal residual. The local load curvature and the normal residual are combined to generate the curvature offset characterization. The normal residual correction channel converts the curvature offset characterization into a normal correction component through a fully connected mapping. It subtracts the normal correction component from the current production and demand load latent state, retains the tangential state component, and generates a curvature-corrected production and demand load latent state. The normal residual correction channel maps the normal correction component to the energy efficiency latent state channel, adjusts the energy efficiency latent state associated with load changes, and generates the curvature correction energy efficiency latent state. The curvature correction output terminal combines the latent state of curvature correction production and demand load and the latent state of curvature correction energy efficiency according to the original channel position, associates the equipment identifier with the prediction time and arranges them in sequence to generate curvature correction latent state data; The curvature-corrected latent state data is input into the decoder, state decoding is performed, and predictive data for the air separation unit is generated, where: Generate predictive data for air separation units, specifically: Extract the device identifier, prediction time, curvature correction production and demand load hidden state, and curvature correction energy efficiency hidden state from the curvature correction hidden state data, and assemble the decoding input data according to the original channel position; The decoder performs fully connected mapping and nonlinear activation on the decoded input data to generate predicted values ​​for equipment operating status, energy consumption, and product production, demand, and quality. Read the minimum and maximum values ​​of the measurement points recorded during numerical normalization, restore each predicted value to the acquired dimensions, and associate it with the equipment identifier and prediction time to generate air separation unit prediction data; An improved Deep Koopman neural network was trained by dividing standardized space separation operation data and dual-tower main cold latent heat pairing data into training, validation, and test sets according to time sequence. The training set was input into the improved Deep Koopman neural network to calculate prediction error, orthogonal decoupling error, energy-mass duality error, and curvature correction error, which were then summed to generate a joint loss. The network parameters were updated in reverse based on the joint loss. The joint loss was calculated based on the validation set, and parameter updates were terminated when the absolute value of the difference between five consecutive adjacent joint losses was less than 0.001. The prediction error was calculated using the test set, where: The training and improvement of the Deep Koopman neural network are as follows: The standardized spatial division operation data and dual-tower main cold latent heat paired data are sorted according to DCS timestamps. The first 70% is assigned to the training set, the middle 15% to the validation set, and the last 15% to the test set to maintain temporal isolation. Multi-step prediction samples are formed by extracting the current operating status and the operating status and control actions of the next 15 consecutive moments based on the DCS timestamp, and samples containing missing time points, alarm status and interlock status are deleted. The training samples are input into the improved Deep Koopman neural network, and forward computation is performed through the encoder, the production-demand-load orthogonal decoupling layer, the state evolution operator, the control action operator, the energy-mass dual operator dual verification layer, the load manifold curvature correction layer, and the decoder. The parameters to be updated include the original Deep Koopman network parameters, the production and demand characteristic trough region, the load orthogonal basis construction branch, the orthogonal residual projection kernel, the decoupling characterization convergence end, the energy transfer calculation branch, the material balance calculation branch, the dual deviation cross kernel, the verification residual write-back channel, the energy and quality characterization output end, the local load neighborhood index region, the tangential base surface construction branch, the curvature offset discrimination kernel, the normal residual correction channel, and the curvature correction output end parameters; Calculate the mean square error between the decoded predicted value and the actual operating value to generate the prediction error; calculate the mean square value of the inner product of the implicit state of production, demand and load and the implicit state of energy efficiency to generate the orthogonal decoupling error; calculate the mean square value of the bidirectional verification deviation to generate the energy quality dual error; calculate the mean square value of the normal residual after curvature correction to generate the curvature correction error. The four errors are summed to generate the joint loss. The Adam optimizer is used to update the network parameters in reverse with a learning rate of 0.001. After each training round, the joint loss is calculated based on the validation set, and the network parameters with the smallest joint loss in the validation set are saved. When the absolute value of the joint loss difference between five consecutive adjacent validation sets is less than 0.001, the parameter update is terminated, the saved network parameters are loaded, and the prediction error is calculated using the test set. This invention optimizes the structure of the controlled Deep Koopman neural network by introducing an orthogonal decoupling layer for production and demand load, a dual verification layer for energy and mass dual operators, and a load manifold curvature correction layer into the state encoding, controlled linear evolution, and multi-step recursive path, thereby enhancing the ability to predict hidden states of variable loads. The production-demand-load orthogonal decoupling layer sets up a production-demand characteristic sub-zone, a load orthogonal basis construction branch, an orthogonal residual projection kernel, and a decoupling characterization convergence end. Based on the product flow rate, cumulative output, and demand of oxygen, nitrogen, and argon, the production-demand-load orthogonal basis is constructed to separate the implicit state of production-demand-load from the implicit state of energy efficiency, reduce the mixing of product demand changes and load rises and falls on energy efficiency characterization, and enhance the consistency of energy consumption prediction under different load conditions. The dual verification layer of the energy and quality dual operator sets up an energy transfer calculation branch, a material balance calculation branch, a dual deviation cross kernel, a verification residual write-back channel, and an energy and quality characterization output terminal. It uses the high-voltage incoming line cumulative power, equipment power, product flow rate, main cooling liquid level, latent heat of phase change of the dual tower main cooling and latent heat loss to construct a two-way verification relationship. It performs deviation write-back on the predicted components that violate energy transfer and material balance, suppressing energy consumption drift and material state mismatch in multi-step recursion. The load manifold curvature correction layer sets up a local load neighborhood index area, a tangential base surface construction branch, a curvature offset discrimination kernel, a normal residual correction channel, and a curvature correction output end. It constructs a local tangential base surface based on the hidden state of production and demand loads at adjacent times, identifies the curvature offset at the load switching position, corrects the normal component that deviates from the local load manifold, and improves the state tracking capability during the load transition phase. The improved Deep Koopman neural network retains linear hidden domain evolution and multi-step prediction capabilities, and increases the decoupling of production and demand, energy quality verification and load curvature adaptation capabilities, providing stable air separation unit prediction data for candidate multi-equipment control data generation, reachability screening and energy saving contribution analysis.

[0026] In this embodiment, generating candidate multi-device control data includes: Control configuration data, safety operation data, and control action data are extracted from standardized air separation operation data. Based on the control configuration data and safety operation data, a configuration safety constraint boundary is constructed, where: Based on the control configuration data and safe operation data, a configuration security constraint boundary is constructed, specifically as follows: Extract equipment identifiers, DCS timestamps, and measurement point numbers from standardized air separation operation data; extract control setpoints, control loop output values, guide vane openings, valve openings, pump speeds, and expander speeds from control action data; and generate the current values ​​of adjustable control quantities. Associate the control configuration data with the equipment identification and measuring point number, read the upper limit of control quantity, lower limit of control quantity, allowable rate of change and interlock limit from the control configuration data, and form the static boundary of control quantity by combining the lower limit of control quantity and the upper limit of control quantity; The allowable rate of change is converted into the allowable change amount per minute. The lower bound of the rate of change is generated by subtracting the allowable change amount per minute from the current value of the adjustable control quantity, and the upper bound of the rate of change is generated by adding the allowable change amount per minute to the current value of the adjustable control quantity. The control quantity rate of change boundary is formed. Read alarm status, interlock status, anti-surge status and equipment safety limits from safe operation data, and associate the equipment safety limits with the corresponding adjustable control quantities according to the equipment identification and measuring point number to form the control quantity safety limit boundary; When a triggering state exists in the alarm state, interlock state, and anti-surge state, the corresponding interlock limit is read. When the interlock limit is prohibited from increasing, the upper limit of the boundary is set to the current value of the adjustable control quantity. When the interlock limit is prohibited from decreasing, the lower limit of the boundary is set to the current value of the adjustable control quantity. When the interlock limit is prohibited from adjusting, both the upper and lower limits of the boundary are set to the current value of the adjustable control quantity. Perform intersection calculations on the static boundary, rate of change boundary, and safety limit boundary of the control quantity, take the maximum value among the lower limits of each boundary as the lower bound of the configuration safety, and take the minimum value among the upper limits of each boundary as the upper bound of the configuration safety. Delete adjustable control quantity records where the lower bound of the configuration safety is greater than the upper bound of the configuration safety. Arrange the device identifier, measurement point number, lower bound of the configuration safety, upper bound of the configuration safety, and interlocking limit corresponding to the remaining adjustable control quantities according to the DCS timestamp to generate the configuration safety constraint boundary. The predicted operating status of the air separation unit is extracted from the predicted data. The corresponding control variable range is calculated based on the control action operator to generate the predicted control boundary. The intersection of the configured safety constraint boundary and the predicted control boundary is calculated according to the equipment identifier and the predicted time to generate the equipment control boundary, where: The device control boundary is generated as follows: Extract equipment identifiers, prediction times, predicted equipment operating status, and adjustable control values ​​associated with the generation of predicted equipment operating status from the air separation unit prediction data; extract the lower bound of configuration safety, the upper bound of configuration safety, and the equipment safety limit from the configuration safety constraint boundaries. According to the equipment identification and measurement point number, the control action vector corresponding to each adjustable control quantity is read from the weight matrix of the control action operator. A unit increment is applied to each adjustable control quantity, and after being mapped by the control action operator, it is added to the corresponding predicted hidden state and converted into the disturbance prediction equipment operating state by the decoder. Subtract the operating status of the predicted equipment from the operating status of the disturbance prediction equipment to generate the sensitivity of the control effect of each adjustable control quantity on temperature, pressure, flow rate, liquid level, speed, current and valve position. Calculate the lower limit margin between each predicted equipment operating status and the corresponding equipment safety lower limit, and calculate the upper limit margin between the corresponding equipment safety upper limit and the predicted equipment operating status; When the control sensitivity is greater than 0, the negative allowable change of the adjustable control quantity is calculated based on the lower limit margin, and the positive allowable change of the adjustable control quantity is calculated based on the upper limit margin. When the control sensitivity is less than 0, the negative allowable change of the adjustable control quantity is calculated based on the upper limit margin, and the positive allowable change of the adjustable control quantity is calculated based on the lower limit margin. When the control action sensitivity is 0 and the predicted equipment operating state is within the equipment safety limit, no state restriction is added to the adjustable control quantity; when the control action sensitivity is 0 and the predicted equipment operating state exceeds the equipment safety limit, the corresponding adjustable control quantity record is deleted. The allowable variation ranges generated by the operating states of various predictive devices associated with the same adjustable control quantity are intersected. The intersection result is added to the value of the adjustable control quantity to generate the lower bound and upper bound of predictive control, which together form the predictive control boundary. Configure safety constraint boundaries and predictive control boundaries according to the equipment identification, measurement point number and prediction time. Take the maximum value of the configured safety lower bound and the predictive control lower bound as the equipment control lower bound, and take the minimum value of the configured safety upper bound and the predictive control upper bound as the equipment control upper bound. Delete records where the lower control boundary of the equipment is greater than the upper control boundary of the equipment. Arrange the remaining lower control boundary of the equipment, upper control boundary of the equipment, equipment identifier, measuring point number and prediction time to generate the equipment control boundary. The current value of the adjustable control quantity is extracted from the control action data. A bidirectional disturbance is applied to the adjustable control quantity within the equipment control boundary. The disturbance results from different equipment are combined according to the predicted time to generate candidate multi-equipment control data, where: Generate candidate multi-device control data, specifically as follows: Extract the device identifier, DCS timestamp, measurement point number and control action value from the control action data, and determine the control actions that can match the device control boundary as adjustable control quantities; For each prediction time, select the DCS timestamp record that is no later than the prediction time and has the smallest time difference, and determine the corresponding control action value as the current value of the adjustable control quantity. Associating the equipment control boundary with the equipment identifier, measurement point number, and predicted time, read the lower and upper control boundaries of the equipment. The positive usable amplitude is generated by subtracting the current value from the upper control boundary, and the negative usable amplitude is generated by subtracting the lower control boundary from the current value. Based on the current value, add half and all of the positive available amplitude respectively to generate two positive disturbance values. Subtract half and all of the negative available amplitude respectively to generate two negative disturbance values. Combine the current value, the two positive disturbance values, and the two negative disturbance values ​​to form a set of candidate values ​​for adjustable control quantities. Delete candidate values ​​that exceed the device control boundary, merge candidate values ​​with the same value, and record the disturbance direction and disturbance amplitude corresponding to each candidate value; Arrange the candidate values ​​of adjustable control quantities according to the prediction time, equipment identification and measuring point number. Set the adjustable control quantity corresponding to each prediction time as a combination column. Construct a mixed level orthogonal combination table based on the number of candidate values ​​in each combination column. The number of times each candidate value appears in the corresponding combination column is consistent. The difference in the number of candidate value pairings between any two combination columns is not greater than 1. Candidate values ​​are extracted row by row from the mixed-level orthogonal combination table, and candidate values ​​of different devices are connected according to the predicted time to form a candidate multi-device control sequence. Remove candidate multi-device control sequences with duplicate content, record the candidate control sequence number, device identifier, measurement point number, prediction time, current value, disturbance direction, disturbance amplitude, and candidate value, and generate candidate multi-device control data.

[0027] In this embodiment, generating reachable candidate control data includes: Candidate control sequences are extracted from candidate multi-device control data, associated with air separation unit prediction data according to device identifier and prediction time, and generated as initial and target hidden states via encoders, wherein: The initial hidden state and the target hidden state are generated as follows: Extract candidate control sequence number, device identifier, measurement point number, prediction time, current value and candidate value from candidate multi-device control data, collect and record according to candidate control sequence number, and sort in ascending order according to prediction time to generate candidate control sequence; The earliest prediction time in each candidate control sequence is determined as the starting prediction time, and the latest prediction time is determined as the target prediction time. Extract equipment identifiers, prediction times, predicted equipment operating statuses, predicted energy consumption measurements, and predicted product output, demand, and quality values ​​from the air separation unit's forecast data. Then, associate each equipment identifier and prediction time with the candidate control sequence to generate candidate control status association records. Extract the predicted equipment operating status, predicted energy consumption metering value, and predicted product output, demand, and quality value corresponding to the initial prediction time from the candidate control state association record, arrange them in the order of equipment identification and encoder input fields, and generate initial state coding data; Extract the predicted equipment operating status, predicted energy consumption metering value, and predicted product output, demand, and quality value corresponding to the target prediction time from the candidate control state association record, and arrange them in the same equipment order and field order as the initial state coding data to generate target state coding data; The initial state encoded data is input into the improved Deep Koopman neural network encoder that has completed parameter training, and feature mapping and latent space projection are performed to generate the initial hidden state. The target state encoding data is input into the improved Deep Koopman neural network encoder that has completed parameter training. The target hidden state is generated by performing feature mapping and latent space projection using the same encoder parameters as the initial hidden state. The initial hidden state, the target hidden state, and the candidate control sequence are stored together according to the candidate control sequence number. Using the initial hidden state and candidate control sequence as input, the predicted hidden state trajectory is recursively generated through state evolution operators and control action operators, where: Generate the predicted hidden state trajectory, specifically as follows: Extract the candidate control sequence number, equipment identifier, measurement point number, prediction time and candidate value from the candidate control sequence. Arrange the candidate values ​​corresponding to the same prediction time into candidate control vectors according to the order of the control variable fields determined by the equipment identifier and measurement point number. Arrange the candidate control vectors in ascending order according to the prediction time, set the initial hidden state corresponding to the starting prediction time as the recursion starting point, and read the state evolution operator parameters and control action operator parameters in the improved Deep Koopman neural network after parameter training is completed. The initial hidden state is input into the state evolution operator, and a linear mapping is performed according to the hidden state dimension to generate the initial autonomous evolution hidden vector. Read the candidate control vector from the first prediction time after the initial prediction time, input the candidate control vector into the control action operator, and generate the control response latent vector according to the mapping relationship between the control quantity field and the latent state dimension. The initial autonomous evolutionary latent vector and the control response latent vector are added item by item according to the latent state dimension to generate the predicted latent state corresponding to the current prediction time. The predicted hidden state corresponding to the current prediction time is used as the input of the next recursive process. The next autonomous evolution hidden vector is generated through the state evolution operator. The candidate control vector corresponding to the next prediction time is generated through the control action operator to generate the next control response hidden vector. The terms are added one by one to generate the predicted hidden state corresponding to the next prediction time. The hidden state is recursively repeated according to the predicted time until all the predicted times included in the candidate control sequence are completed. The initial hidden state and the predicted hidden state corresponding to each predicted time are arranged in chronological order and associated with the candidate control sequence number to generate the predicted hidden state trajectory. Based on the predicted hidden state trajectory, state evolution operator, and control action operator, a reachable hidden domain is constructed. The target hidden state is projected onto the reachable hidden domain, and the difference norm between the target hidden state and the projection result is calculated to generate the reachable residual, where: Generate reachable residuals, specifically: Extract candidate control sequence numbers, start prediction time, target prediction time, and prediction hidden states corresponding to each prediction time from the predicted hidden state trajectory, and set the prediction hidden state corresponding to the target prediction time as the center of the reachable hidden domain. Based on the candidate control sequence number, associate the candidate multi-device control data, and collect the candidate values ​​according to the device identifier, measurement point number and prediction time. Select the minimum and maximum candidate values ​​corresponding to the same control quantity at the same prediction time. The lower limit of the control offset is generated by subtracting the candidate value in the current candidate control sequence from the minimum candidate value, and the upper limit of the control offset is generated by subtracting the candidate value in the current candidate control sequence from the maximum candidate value, thus forming the control offset interval; Extract the control action vector corresponding to each adjustable control quantity from the control action operator. Based on the number of recursions between the prediction time of the control quantity and the target prediction time, input the control action vector into the state evolution operator one by one to generate the reachable direction vector corresponding to each adjustable control quantity at the target prediction time. Arrange the reachable direction vectors according to the predicted time, equipment identification and measurement point number to form a reachable direction matrix. Associate each reachable direction vector with the corresponding control offset interval. The control offset values ​​within each control offset interval are used as weights for the reachable direction vector. The weighted reachable direction vectors are summed, and the summation results are superimposed on the center of the reachable hidden domain. The reachable hidden domain is formed by the hidden states corresponding to all control offset values. Subtract the target hidden state from the center of the reachable hidden domain item by item according to the hidden state dimension to generate the target offset vector. Perform singular value decomposition on the reachable direction matrix and construct the generalized inverse. Multiply the generalized inverse with the target offset vector to generate the initial projection coefficients. The initial projection coefficients are compared with the corresponding control offset intervals one by one. The initial projection coefficients that exceed the control offset intervals are fixed to the nearest lower or upper control offset limit. The reachable direction components corresponding to the fixed projection coefficients are deducted from the target offset vector. The projection coefficients are recalculated for the remaining reachable direction vectors until all projection coefficients are within the corresponding control offset intervals. The target hidden state is projected into the center of the reachable hidden domain by multiplying each reachable direction vector by its corresponding projection coefficient and summing the results. The target hidden state and the projection result are subtracted item by item according to the hidden state dimension. The squares of each difference are summed and the square root of the sum is calculated. The calculation result is associated with the candidate control sequence number to generate the reachable residual. Candidate control sequences with reachable residuals greater than a preset reachable residual threshold are filtered out. The remaining candidate control sequences are then aggregated to generate reachable candidate control data, where: Generate reachable candidate control data, specifically: Extract the actual execution control sequences that did not trigger alarms and interlocks from the validation set of the improved Deep Koopman neural network, generate the validation reachable residual according to the reachable residual calculation process, and set the 95th percentile value of the validation reachable residual as the preset reachable residual threshold. According to the candidate control sequence number, associate the reachable residual with the candidate multi-device control data, delete the candidate control sequence whose reachable residual is greater than the preset reachable residual threshold, and retain the candidate control sequence whose reachable residual is less than or equal to the preset reachable residual threshold; The retained candidate control sequences, reachable residuals, and candidate control sequence numbers are aggregated to generate reachable candidate control data.

[0028] In this embodiment, generating multi-device marginal energy-saving contribution data includes: Based on reachable candidate control data, an improved Deep Koopman neural network is used to predict candidate control energy consumption. Baseline energy consumption data is correlated with the load condition and the predicted time. The candidate total energy saving is generated by subtracting the candidate control energy consumption from the baseline energy consumption, where: The candidate total node energy is generated as follows: Extract candidate control sequence numbers, prediction times, and candidate values ​​from reachable candidate control data, and form candidate control vector sequences according to prediction times; Data records corresponding to the initial prediction time are extracted from standardized air separation operation data and dual-tower main cold latent heat pairing data. An improved Deep Koopman neural network is trained by inputting candidate control vector sequences to generate predicted high-voltage incoming line cumulative power and predicted product flow. Based on the 1-hour cycle aggregation prediction results, the difference in the cumulative power of the high-voltage incoming line predicted in adjacent aggregation cycles is calculated to generate candidate control energy consumption. The predicted product flow rate is converted into the predicted oxygen-based comprehensive output using the oxygen conversion factor. Candidate load conditions are determined based on the direction of change and the duration of the predicted oxygen-based comprehensive output. The baseline energy consumption is generated by multiplying the baseline unit oxygen-based comprehensive power consumption by the predicted oxygen-based comprehensive output, based on the candidate load conditions and the baseline energy consumption data associated with the predicted time. Subtract the candidate control energy consumption from the baseline energy consumption to generate the periodic section energy. Sum the periodic section energies corresponding to the same candidate control sequence number to generate the candidate total section energy. Remove the adjustable control disturbances for each device one by one, fix the adjustable control variables for other devices, and generate a single-device control withdrawal sequence, wherein: Generate a single-device revocation control sequence, specifically as follows: Extract candidate control sequence number, equipment identifier, measurement point number, prediction time, current value and candidate value from the reachable candidate control data, and collect them into a complete candidate control sequence according to the candidate control sequence number; Compare the current and candidate values ​​of each adjustable control quantity of the same device, and group the devices with numerical differences into a set of devices to be withdrawn. Select one device identifier from the set of devices to be withdrawn one by one, copy the corresponding complete candidate control sequence, replace all candidate values ​​of the selected device at each prediction time with the current value, and leave the candidate values ​​of the devices not withdrawn unchanged. Set a withdrawal control sequence number for each withdrawal result, and record the original candidate control sequence number, withdrawal device identifier, measurement point number, prediction time, and replacement control value; Repeat the equipment withdrawal process until all equipment in the set of equipment to be withdrawn is completed, and generate a single equipment withdrawal control sequence. The single-device withdrawal control sequence is input into an improved Deep Koopman neural network to generate withdrawal control energy consumption. The marginal energy saving contribution of the device is generated by subtracting the candidate control energy consumption from the withdrawal control energy consumption. The marginal energy-saving contribution of generating equipment is as follows: Extract the control sequence number, original candidate control sequence number, withdrawing device identifier, predicted time, and replacement control value from the single device withdrawal control sequence; Data records corresponding to the initial prediction time are extracted from standardized air separation operation data and dual-tower main cold latent heat pairing data. An improved Deep Koopman neural network is trained with single-device withdrawal control sequence input to generate the predicted high-voltage incoming line cumulative power. Based on the cumulative power of the high-voltage incoming line predicted by the 1-hour cycle aggregation, the difference in cumulative power between adjacent aggregation cycles is calculated to generate the energy consumption for withdrawal control. The candidate control energy consumption is associated with the original candidate control sequence number and the predicted time. The energy consumption of withdrawn control is subtracted from the energy consumption of candidate control to generate the periodic marginal energy saving contribution. Sum the periodic marginal energy-saving contributions corresponding to the same withdrawn equipment identifier to generate the equipment's marginal energy-saving contribution; By aggregating the marginal energy-saving contribution and candidate total energy savings of each equipment according to its identifier, multi-equipment marginal energy-saving contribution data is generated, including: Generate multi-device marginal energy-saving contribution data, specifically: Extract candidate control sequence numbers and corresponding values ​​from candidate total energy savings, and extract original candidate control sequence numbers, withdrawn equipment identifiers and corresponding values ​​from equipment marginal energy saving contributions; The candidate total energy saving is associated with the marginal energy saving contribution of the equipment according to the candidate control sequence number, and the marginal energy saving contribution of the equipment corresponding to the same candidate control sequence is arranged according to the equipment identifier; Record candidate control sequence numbers, equipment identifiers, equipment marginal energy-saving contributions, and candidate total energy savings to generate multi-equipment marginal energy-saving contribution data.

[0029] In this embodiment, the generation of energy consumption optimization results includes: By associating the marginal energy-saving contribution data of multiple devices with the achievable candidate control data according to the candidate control sequence, the candidate total energy saving, the marginal energy-saving contribution of each device, and the adjustable control quantity disturbance value are extracted, among which: Extract candidate total energy savings, equipment marginal energy-saving contribution, and adjustable controllable disturbance values, specifically as follows: Extract candidate control sequence numbers, equipment identifiers, candidate total energy savings, and equipment marginal energy savings from multi-equipment marginal energy saving contribution data; Extract candidate control sequence number, device identifier, measurement point number, prediction time, current value, and candidate value from reachable candidate control data; Subtract the current value from the candidate value to generate an adjustable control disturbance value with the disturbance direction; Based on the two types of data associated with candidate control sequence number and equipment identifier, the candidate total energy saving is recorded according to the candidate control sequence number, the marginal energy saving contribution of equipment is recorded according to the equipment identifier, and the adjustable control quantity disturbance value is recorded according to the measurement point number and the prediction time. For equipment with negative marginal energy-saving contributions, the absolute values ​​of their marginal energy-saving contributions are summed. The total energy saved in the candidate projects is then subtracted from the summed values ​​to generate a collaborative energy-saving score, where: Generate a collaborative energy-saving score, specifically as follows: Collect the marginal energy-saving contribution and total energy saving of the equipment according to the candidate control sequence number; Select the marginal energy-saving contribution of equipment with a value less than 0, take the absolute value of each selection result and sum them to generate a negative contribution penalty value. If there is no negative value record, the negative contribution penalty value is recorded as 0. The negative contribution penalty value is subtracted from the candidate total energy saving value to generate the collaborative energy saving score of the corresponding candidate control sequence; The collaborative energy-saving score is associated with and stored with the candidate control sequence number; Candidate control sequences are sorted in descending order of their collaborative energy-saving scores. If the collaborative energy-saving scores are the same, they are sorted in ascending order of the sum of the absolute values ​​of the adjustable control quantity disturbances. This generates the candidate control sequence ranking result, where: The candidate control sequence sorting results are generated as follows: Collect collaborative energy-saving scores and adjustable control disturbance values ​​according to the candidate control sequence number; Take the absolute value of the adjustable control quantity disturbance value corresponding to each device, each measuring point and each prediction time in the same candidate control sequence and sum them to generate the total control disturbance. Candidate control sequences are arranged from largest to smallest according to the collaborative energy-saving score. When the collaborative energy-saving scores are the same, they are arranged from smallest to largest according to the total amount of control disturbance. When the collaborative energy-saving score and the total amount of control disturbance are the same, the candidate control sequences are arranged in ascending order according to their numbers. Set the sorting order according to the arrangement order, record the candidate control sequence number, collaborative energy saving score, total control disturbance and sorting order, and generate the candidate control sequence sorting result; The first candidate control sequence is selected from the candidate control sequence ranking results, and associated with the corresponding candidate total energy saving and equipment marginal energy saving contribution to generate energy consumption optimization results, where: The energy consumption optimization results are generated as follows: Extract the candidate control sequence number, collaborative energy-saving score, and total control disturbance of the candidate control sequence with the first sorting position from the candidate control sequence sorting results; Based on the candidate control sequence number, the reachable candidate control data is associated, and the equipment identifier, measurement point number, prediction time, current value, candidate value and adjustable control quantity disturbance value are extracted. The data are then sorted in ascending order according to the prediction time to generate the target control sequence. By associating the marginal energy-saving contribution data of multiple devices with the candidate control sequence number, the candidate total energy saving is extracted, and the marginal energy-saving contribution of each device is extracted according to its identifier. The target control sequence, candidate total energy saving, equipment marginal energy saving contribution, collaborative energy saving score and total control disturbance are correlated and recorded to generate energy consumption optimization results.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a large-scale industrial gas production scenario. The air separation unit is equipped with compression, expansion, conveying, distillation, and dual-tower main refrigeration equipment, with rated oxygen, nitrogen, and argon outputs of 55,000 Nm³ / h, 32,000 Nm³ / h, and 1,200 Nm³ / h, respectively. Traditional methods adjust equipment setpoints based on product flow deviations. During load increases, the main air compressor increases load first, while the expander and liquid oxygen pump respond later. After the product flow reaches the required value, the high-pressure incoming line power continues to increase, and the latent heat imbalance rate of the dual-tower main refrigeration unit exceeds 3%.

[0031] The simulation process received 30 consecutive days of DCS data, energy consumption data, and product measurement data, with a sampling period of 1 minute, encompassing 43,200 time points and 186 measurement points. Missing values ​​accounted for 0.41%, out-of-bounds values ​​for 0.09%, and there were 34 duplicate or highly correlated measurement points. After aggregation, imputation, out-of-bounds correction, and normalization, 152 measurement points were retained, resulting in 39,840 valid samples. The training set, validation set, and test set each contained 27,888, 5,976, and 5,976 samples, respectively.

[0032] Training sample 004816 corresponds to steady-state conditions, with an oxygen product flow rate of 50240 Nm³ / h, a demand of 50300 Nm³ / h, a high-pressure incoming line power of 38.21 MW, a main air compressor guide vane opening of 77.1%, and an expander speed of 26790 r / min. The supervised values ​​after 15 minutes are 50310 Nm³ / h and 38.24 MW. Training sample 018742 corresponds to increased load conditions, with an oxygen product flow rate of 50320 Nm³ / h, a demand of 50800 Nm³ / h, a high-pressure incoming line power of 38.74 MW, a main coolant level of 68.4%, and a latent heat loss of 4700 kJ / min. The supervised values ​​after 15 minutes are 51040 Nm³ / h, 68.9%, and 39.08 MW. The traditional DeepKoopman neural network and the improved Deep Koopman neural network use the same partitioning.

[0033] The sample consisted of 648 1-hour cycles, with 487 cycles in steady state, 86 in load increase, and 75 in load decrease conditions. The baseline unit oxygen-based comprehensive power consumption was 0.612 kWh / Nm³, 0.638 kWh / Nm³, and 0.621 kWh / Nm³, respectively. After matching the dual-tower main cooling state sequence with the nitrogen and oxygen working fluid property data, the main cooling response delay was 4 min. The latent heat of nitrogen condensation at the two main cooling points was 146,800 kJ / min and 139,400 kJ / min, respectively; the latent heat of liquid oxygen evaporation was 142,100 kJ / min and 135,600 kJ / min, respectively; and the latent heat loss was 4,700 kJ / min and 3,800 kJ / min, respectively.

[0034] During training, the production-demand-load orthogonal decoupling layer calls the production-demand characteristic distribution area, the load orthogonal basis construction branch, the orthogonal residual projection kernel, and the decoupling characterization convergence end; the energy-quality dual operator dual verification layer calls the energy transfer calculation branch, the material balance calculation branch, the dual deviation cross kernel, the verification residual write-back channel, and the energy-quality characterization output end; the load manifold curvature correction layer calls the local load neighborhood index area, the tangential base surface construction branch, the curvature offset discrimination kernel, the normal residual correction channel, and the curvature correction output end. The joint loss decreased from 0.0924 to 0.0068, and training was terminated after the difference was less than 0.001 for 5 consecutive rounds. The energy consumption, product flow, and load state prediction errors of the method of this invention are 0.76%, 0.93%, and 1.08%, respectively, while those of the traditional Deep Koopman method are 1.84%, 1.67%, and 2.96%, respectively, with coefficients of determination of 0.982 and 0.941, respectively.

[0035] A closed-loop simulation was used to simulate a sudden surge in product demand. Oxygen demand increased from 50,000 Nm³ / h to 55,000 Nm³ / h within 40 minutes, and nitrogen demand increased from 29,000 Nm³ / h to 31,500 Nm³ / h. At the start of the event, the high-voltage incoming line power was 38.92 MW, the main air compressor guide vane opening was 78.0%, the expander speed was 26,850 r / min, the liquid oxygen pump speed was 2,980 r / min, the oxygen compressor guide vane opening was 74.0%, and the nitrogen compressor guide vane opening was 69.0%. After intersecting the safety constraint boundary and the predictive control boundary, the equipment control boundaries for the five control variables were 76.4% to 80.2%, 26,600 r / min to 27,100 r / min, 2,920 r / min to 3,040 r / min, 72.5% to 76.0%, and 67.5% to 71.0%, respectively.

[0036] Five levels of bidirectional disturbance values ​​are generated within the equipment control boundary, and 625 candidate multi-device control sequences are formed through a hybrid horizontal orthogonal combination. The state evolution operator and control action operator recursively generate predicted latent state trajectories, with the reachable residual threshold being the 95th percentile of the verified reachable residual, 0.083. 512 sequences are retained and 113 are deleted after reachability screening. Four consecutive 15-step predictions are performed on the retained sequences to form 1 hour of candidate control energy consumption.

[0037] After removing the disturbances from each device, the top-ranked sequence was adjusted to 76.8% for the main air compressor guide vane, 26970 r / min for the expander, 2945 r / min for the liquid oxygen pump, 73.2% for the oxygen compressor guide vane, and 69.4% for the nitrogen compressor guide vane. The baseline energy consumption for 1 hour was 39.24 MWh, the candidate control energy consumption was 37.86 MWh, and the candidate total energy saving was 1.38 MWh. The marginal energy saving contributions of the five devices were 0.58 MWh, 0.29 MWh, 0.17 MWh, 0.24 MWh, and -0.03 MWh, respectively, with a synergistic energy saving score of 1.35 MWh.

[0038] Under the same demand curve and initial conditions, the traditional rule-based control method, by adjusting five control variables to 79.4%, 26720 r / min, 3015 r / min, 75.1%, and 70.2%, results in an energy consumption of 39.12 MWh per hour. In a continuous 72-hour comparative simulation, the average power of the traditional method and the method of this invention were 38.92 MW and 37.61 MW, respectively; the comprehensive power consumption per unit oxygen equivalent was 0.628 kWh / Nm³ and 0.606 kWh / Nm³, respectively; the load stabilization time was 48 min and 31 min, respectively; the maximum oxygen flow deviation was 2.73% and 0.82%, respectively; and the dual-tower main refrigeration latent heat imbalance rate was 3.06% and 1.32%, respectively. The traditional method experienced four high-load alarms, while the method of this invention had zero alarms and interlocks, resulting in a 3.50% reduction in comprehensive energy consumption.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the energy consumption of multiple devices in an air separation unit based on deep learning, characterized in that, include: Collect and preprocess multi-source operational data to generate standardized space separation operational data; Based on standardized air separation operation data, load conditions are divided and an energy consumption baseline model is established to generate baseline energy consumption data for the operating conditions. Perform dual-tower main cold latent heat pairing on standardized air separation operation data, calculate the dual-tower main cold phase change latent heat and latent heat loss, and generate dual-tower main cold latent heat pairing data. An improved Deep Koopman neural network with an orthogonal decoupling layer for production and demand load, a dual verification layer for energy and quality operators, and a load manifold curvature correction layer was constructed. Standardized air separation operation data and dual-tower main cooling latent heat pairing data were input to generate air separation unit prediction data. Based on the predicted data of the air separation unit, the equipment control boundary is obtained by intersection and the adjustable control quantity is disturbed to generate candidate multi-equipment control data. Perform candidate control reachability screening on candidate multi-device control data, filter out unreachable control sequences based on state evolution operators and control action operators, and generate reachable candidate control data; Based on the baseline energy consumption data of the operating condition, a multi-device energy-saving contribution analysis is performed on the achievable candidate control data to generate multi-device marginal energy-saving contribution data; Based on the marginal energy-saving contribution data of multiple devices, the achievable candidate control data are sorted, a control sequence is selected, and energy consumption optimization results are generated.

2. The energy consumption optimization method for multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The multi-source operational data includes: Equipment operating status data, control action data, control configuration data, safe operation data, energy consumption metering data, and product production, demand, and quality data; Equipment operating status data includes: temperature, pressure, flow rate, liquid level, speed, current, and valve position; Control action data includes: control setpoint, control loop output value, guide vane opening, valve opening, pump speed, and expander speed; Control configuration data includes: upper limit of control quantity, lower limit of control quantity, allowable rate of change, and interlocking restrictions; Safe operation data includes: alarm status, interlock status, anti-surge status, and equipment safety limits; Energy consumption metering data includes: cumulative power consumption of high-voltage incoming lines, equipment power, drive motor current, and time-of-use energy consumption; Product production, demand, and quality data include: product flow rate, cumulative output, demand, product purity, carbon dioxide content, and pressure nitrogen trace oxygen content for oxygen, nitrogen, and argon.

3. The energy consumption optimization method for multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generation of standardized space division operation data includes: Data on equipment operation status, control actions, control configuration, and safe operation is collected from the DCS; energy consumption data is collected from the energy consumption metering terminal; and product production, demand, and quality data are collected from the product metering terminal to generate multi-source operation data. Based on the DCS timestamp, multi-source operation data is aggregated and aligned in a one-minute cycle, missing values ​​are linearly interpolated, out-of-bounds values ​​are corrected according to control configuration data, and cleaning operation data is generated. Remove duplicate and highly correlated redundant measurement points from the cleaning operation data, retain control configuration data and safety operation data, normalize the values, and generate standardized space separation operation data.

4. The method for optimizing the energy consumption of multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generated baseline energy consumption data includes: Product flow and energy consumption metering data are extracted from standardized air separation operation data. Production conversion, periodic aggregation and unit energy consumption conversion are performed on the product flow and energy consumption metering data to generate oxygen-converted comprehensive production and unit oxygen-converted comprehensive power consumption. Based on the changing direction and continuous state of the total oxygen production, the load change process is identified, and the steady-state operating condition, the load increase operating condition, and the load decrease operating condition are divided. The division results are correlated with the unit oxygen production power consumption to generate load operating condition data. Safe operation data is extracted from standardized air separation operation data, correlated with load condition data and selected as effective operation samples, and the energy consumption baseline model is trained using the effective operation samples. Baseline energy consumption is calculated according to load condition to generate baseline energy consumption data for operating conditions.

5. The method for optimizing the energy consumption of multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generation of dual-tower primary cold latent heat pairing data includes: Operating status data of the dual-tower main refrigeration equipment were extracted from standardized air separation operation data to construct a dual-tower main refrigeration status sequence. Read the nitrogen and oxygen working fluid property data from the air separation unit process design document, match the dual-tower main cooling state sequence with the nitrogen and oxygen working fluid property data according to temperature and pressure, determine the nitrogen condensation phase boundary and the liquid oxygen evaporation phase boundary, and calculate and generate the dual-tower main cooling phase change latent heat sequence. A time-shift scan is performed on the latent heat of phase change of the dual-tower main cooler. The response delay of the main cooler is determined based on the time-shift correlation value and the direction of change of the main cooler liquid level. The latent heat of phase change sequences on both sides are aligned according to the response delay of the main cooler to generate a pairing sequence of latent heat of the dual-tower main cooler. The latent heat of nitrogen condensation and the latent heat of liquid oxygen evaporation in the latent heat pairing sequence of the dual-tower main cooling system are calculated by mean and difference to generate the latent heat of phase change and latent heat loss of the dual-tower main cooling system. These are then associated with the state sequence of the dual-tower main cooling system by timestamp to generate the latent heat pairing data of the dual-tower main cooling system.

6. The energy consumption optimization method for multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generated air separation unit prediction data includes: An improved Deep Koopman neural network is constructed, which includes an orthogonal decoupling layer for production, demand and load, a dual verification layer for energy and mass dual operators, and a load manifold curvature correction layer. The standardized air separation operation data and the dual-tower main cold latent heat pairing data are converted into the initial hidden state by the encoder. The production demand load orthogonal decoupling layer performs orthogonal separation of production demand load and energy efficiency state on the initial hidden state, generating the production demand load hidden state and the energy efficiency hidden state, and forming the orthogonal decoupling hidden state data. Based on the control action data, a controlled Koopman evolution is performed on the orthogonally decoupled hidden state data, and a predicted hidden state is generated through the state evolution operator and the control action operator. Based on standardized air separation operation data and dual tower main cold latent heat pairing data, the dual-operator dual verification layer performs bidirectional verification of energy transfer and material transfer on the predicted hidden state, adjusts the predicted hidden state according to the bidirectional verification deviation, and generates energy and quality verification hidden state data. The load manifold curvature correction layer performs local curvature identification and tangential state correction on the energy quality verification hidden state data based on the changes in the hidden state of production and demand load at adjacent time points, and generates curvature correction hidden state data. The curvature correction latent state data is input into the decoder, state decoding is performed, and air separation unit prediction data is generated. An improved Deep Koopman neural network was trained by dividing standardized space separation operation data and dual-tower main cold latent heat pairing data into training, validation, and test sets according to time sequence. The training set was input into the improved Deep Koopman neural network to calculate prediction error, orthogonal decoupling error, energy-mass duality error, and curvature correction error, and these were added together to generate a joint loss. The network parameters were updated in reverse based on the joint loss. The joint loss was calculated based on the validation set. The parameter update was terminated when the absolute value of the difference between five consecutive adjacent joint losses was less than 0.

001. The prediction error was calculated using the test set.

7. The energy consumption optimization method for multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generation of candidate multi-device control data includes: Control configuration data, safety operation data, and control action data are extracted from standardized air separation operation data, and configuration safety constraint boundaries are constructed based on the control configuration data and safety operation data. The predicted equipment operating status is extracted from the predicted data of the air separation unit. The corresponding control quantity range is calculated based on the control action operator to generate the predicted control boundary. The intersection of the configured safety constraint boundary and the predicted control boundary is calculated according to the equipment identification and the predicted time to generate the equipment control boundary. Extract the current value of the adjustable control quantity from the control action data, perform bidirectional disturbance on the adjustable control quantity within the equipment control boundary, combine the disturbance results of different equipment according to the predicted time, and generate candidate multi-equipment control data.

8. The method for optimizing the energy consumption of multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generation of reachable candidate control data includes: Candidate control sequences are extracted from candidate multi-device control data, associated with air separation unit prediction data according to equipment identification and prediction time, and generated as initial hidden state and target hidden state by encoder. The predicted hidden state trajectory is generated recursively using the initial hidden state and candidate control sequence as inputs through state evolution operator and control action operator; Based on the predicted hidden state trajectory, state evolution operator, and control action operator, a reachable hidden domain is constructed. The target hidden state is projected onto the reachable hidden domain, and the difference norm between the target hidden state and the projection result is calculated to generate the reachable residual. Candidate control sequences with reachable residuals greater than the preset reachable residual threshold are filtered out, and the remaining candidate control sequences are collected to generate reachable candidate control data.

9. The energy consumption optimization method for multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generation of multi-device marginal energy-saving contribution data includes: Based on reachable candidate control data, the energy consumption of candidate control is predicted by an improved Deep Koopman neural network. The baseline energy consumption data of the load condition and the prediction time are associated with the baseline energy consumption to generate the candidate total energy saving. Remove the adjustable control quantity disturbance of each device one by one, fix the adjustable control quantity of other devices, and generate a single device control withdrawal sequence; The single-device withdrawal control sequence is input into an improved Deep Koopman neural network to generate withdrawal control energy consumption. The candidate control energy consumption is subtracted from the withdrawal control energy consumption to generate the marginal energy saving contribution of the device. Collect marginal energy-saving contributions and candidate total energy savings of equipment according to equipment identification, and generate multi-equipment marginal energy-saving contribution data.

10. The method for optimizing the energy consumption of multiple devices in an air separation unit based on deep learning according to claim 1, characterized in that, The generated energy consumption optimization results include: By associating the marginal energy-saving contribution data of multiple devices with the achievable candidate control data according to the candidate control sequence, the candidate total energy saving, the marginal energy-saving contribution of the devices, and the disturbance value of the adjustable control quantity are extracted. The absolute values ​​of the marginal energy-saving contributions of equipment with negative marginal energy-saving contributions are summed, and the summation is subtracted from the candidate total energy saving to generate a collaborative energy-saving score. Candidate control sequences are sorted in descending order according to their collaborative energy-saving scores. When the collaborative energy-saving scores are the same, they are sorted in ascending order according to the sum of the absolute values ​​of the adjustable control quantity disturbances, thus generating the candidate control sequence sorting results. The first candidate control sequence is selected from the candidate control sequence ranking results, and the corresponding candidate total energy saving and equipment marginal energy saving contribution are associated to generate energy consumption optimization results.