Multi-machine regulation and control method and device for electrically-driven compressed air station

Abstract

The invention provides a multi-machine regulation and control method and device for an electrically-driven compressed air station, and relates to the technical field of compressor regulation and control. A digital twin model is constructed, a plurality of optimization targets are fused, and an upper-layer energy management decision maker and a middle-layer strategy generator are trained in a layered offline manner; and then, the trained upper-layer energy management decision maker, the trained middle-layer strategy generator and the trained bottom-layer execution controller are used for jointly controlling the electrically-driven compressed air station, so that the dependence on artificial experience scheduling can be greatly reduced, the self-adaptive capability is improved, the gas consumption prediction information and the electricity price prediction information are fully utilized, and the energy utilization rate of the electric-driven compressed air station is improved. According to the method, the unit air volume energy consumption, the comprehensive operation cost, the equipment service life and the carbon emission intensity can be considered, the rationality of starting, stopping and load distribution of all the air compressors is improved, the stability and robustness of operation of the whole electrically-driven compressed air station are improved, and synchronous optimization of energy conservation, health and reliability of the electrically-driven compressed air station is achieved.

Description

Technical Field

[0001] This invention relates to the field of compressor control technology, and in particular to a multi-unit control method and device for an electrically driven compressed air station. Background Technology

[0002] An industrial electrically driven compressed air station includes multiple air compressors driven by electric motors. The types of air compressors can include screw compressors and centrifugal compressors. Screw compressors can include fixed frequency screw compressors and variable frequency screw compressors.

[0003] The existing electric compressed air station has the following problems in the control of each air compressor: 1) The scheduling rules rely on manual experience and lack adaptive ability; 2) It does not make full use of economic information such as time-of-use electricity pricing, demand-based electricity charges, and outsourced maintenance costs; 3) The start-up and shutdown and load distribution of each air compressor are unreasonable, which not only leads to high energy consumption per unit volume of air, but also the frequent start-up and shutdown of the air compressor and high-pressure ratio operation will affect its lifespan and reliability; 4) Under the premise of ensuring workshop pressure and flow, the carbon emission intensity is still too high. Summary of the Invention

[0004] This invention provides a multi-machine control method and device for an electrically driven compressed air station, which addresses the deficiencies in related technologies.

[0005] This invention provides a multi-unit control method for an electrically driven compressed air station, comprising: A digital twin model of an electrically driven compressed air station is constructed; the digital twin model includes a cascaded coupling structure of data-driven models of each air compressor unit in the electrically driven compressed air station and pressure response models at the end of the pipeline network. Based on the digital twin model, an offline training environment is built, and the multi-machine control task of the electric compressed air station is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and a supervised learning algorithm and an imitation learning algorithm are used to train the middle-level policy generator offline. Each optimization objective includes a unit compressed air energy consumption target, a comprehensive operating cost target, and an equipment lifespan and health target. Based on the trained upper-level energy management decision-maker, the current status information, air consumption prediction information, and electricity price prediction information of the electric-driven compressed air station are used to determine the air compressor unit operation strategy and target range information of the electric-driven compressed air station. Based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information are applied to constrain and optimize the air compressor unit operation strategy, generate a control strategy, and based on the bottom-level execution controller, apply the control strategy to control the electric-driven compressed air station.

[0006] According to the multi-machine control method of an electrically driven compressed air station provided by the present invention, the step of applying a deep reinforcement learning algorithm to perform offline training on the upper-level energy management decision-maker in the offline training environment based on each of the optimization objectives includes: In the offline training environment, a reward function is constructed based on each of the optimization objectives, and a state space is constructed based on the current state information, the gas consumption prediction information, and the electricity price prediction information. An action space is constructed based on the air compressor unit operation strategy. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the electric compressed air station. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline.

[0007] According to a multi-unit control method for an electrically driven compressed air station provided by the present invention, the steps for constructing the data-driven model of each air compressor unit include: For any air compressor unit, obtain the operating status parameters of the air compressor in the air compressor unit, the inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical characteristics of the air compressor unit, and determine the actual flow rate and actual power of the air compressor unit. The operating status parameters, inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical features are input into the data-driven model of any air compressor unit to obtain the predicted flow rate and predicted power output by the data-driven model of any air compressor unit. Based on the actual flow rate, the actual power, the predicted flow rate, and the predicted power, the data-driven model of any air compressor unit is trained. Once the training process is complete, the construction process of the data-driven model of any air compressor unit is finished.

[0008] According to the present invention, a multi-unit control method for an electrically driven compressed air station is provided, wherein the input of the pipeline end pressure response model includes the predicted flow rate of each air compressor unit, the air compressor unit allocation structure information, the air storage tank pressure of the electrically driven compressed air station, and the flow rate of a representative pipeline flow measurement point of the electrically driven compressed air station; the flow rate of the representative pipeline flow measurement point reflects the change in air consumption of the pipeline of the electrically driven compressed air station. The output of the pipeline end pressure response model includes the predicted pressure of each key node of the electric compressed air station.

[0009] According to the multi-unit control method of an electrically driven compressed air station provided by the present invention, the step of constructing the data-driven model of any air compressor unit further includes: Obtain the theoretical power and physical constraints of pressure and flow rate for any of the air compressor units; The theoretical power and the physical constraints of pressure and flow are used as mechanistic priors to update the input or training loss of the data-driven model of any air compressor unit.

[0010] According to a multi-unit control method for an electrically driven compressed air station provided by the present invention, the prediction step of the air consumption prediction information includes: Obtain historical air consumption information, production plan characteristics, time characteristics, and environmental characteristics of the electric compressed air station; The historical gas consumption information, the production plan characteristics, the time characteristics, and the environmental characteristics are input into the gas consumption prediction model to obtain the gas consumption prediction information output by the gas consumption prediction model.

[0011] According to the present invention, a multi-unit control method for an electrically driven compressed air station, wherein the control strategy is applied based on a low-level execution controller to control the electrically driven compressed air station, and the method further includes: Calculate the instantaneous efficiency of each air compressor unit, and calculate the relative efficiency deviation based on the instantaneous efficiency; Based on the relative efficiency deviation, the health score of each air compressor unit is calculated, and the health score of each air compressor unit is statistically analyzed within a preset time period. If the health score of the target air compressor unit is lower than the preset score within the preset time period, the priority of the target air compressor unit will be reduced in the decision-making process of the upper-level energy management decision-maker after training.

[0012] According to the multi-machine control method of an electrically driven compressed air station provided by the present invention, the offline training of the mid-level policy generator based on supervised learning algorithm and imitation learning algorithm includes: The system obtains the air compressor unit operation strategy sample and target interval information sample obtained by the upper-level energy management decision-maker after training, and obtains the gas consumption change trend sample, the expert control strategy of each air compressor unit, and the working mode label. The air compressor unit operation strategy sample, the target interval information sample, and the air consumption change trend sample are input into the middle-level strategy generator to obtain the control strategy sample output by the middle-level strategy generator. Based on the underlying execution controller, the control strategy sample is applied to control the electric compressed air station, and the unit compressed air energy consumption and overall operating cost of the electric compressed air station are detected. The intermediate-level strategy generator is trained offline based on the expert control strategy, the control strategy samples, the unit compressed air energy consumption under different operating mode labels, and the comprehensive operating cost.

[0013] The present invention also provides a multi-unit control device for an electrically driven compressed air station, comprising: The model building module is used to build a digital twin model of the electric-driven compressed air station; the digital twin model includes a cascaded coupling structure of the data-driven model of each air compressor unit in the electric-driven compressed air station and the pressure response model at the end of the pipeline network. The offline training module is used to build an offline training environment based on the digital twin model, and to transform the multi-machine control task of the electric compressed air station into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and a supervised learning algorithm and an imitation learning algorithm are used to train the middle-level policy generator offline. Each optimization objective includes a unit compressed air energy consumption target, a comprehensive operating cost target, and an equipment lifespan and health target. The upper-level decision module is used to determine the operating strategy and target range information of the air compressor unit of the electric-driven compressed air station based on the trained upper-level energy management decision-maker and by applying the current status information, air consumption prediction information and electricity price prediction information of the electric-driven compressed air station. The control module is used to optimize the operation strategy of the air compressor unit based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information, generate a control strategy, and control the electric-driven compressed air station based on the bottom-level execution controller and the control strategy.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-machine control method of the electrically driven compressed air station as described above.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-machine control method of the electrically driven compressed air station as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-machine control method of the electrically driven compressed air station as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The multi-unit control method and apparatus for electrically driven compressed air stations provided by this invention constructs a digital twin model, integrating multiple optimization objectives such as unit compressed air energy consumption, comprehensive operating cost, and equipment lifespan and health. It performs hierarchical offline training of the upper-level energy management decision-maker and the middle-level strategy generator. Then, using the trained upper-level energy management decision-maker, middle-level strategy generator, and lower-level execution controller, the electric compressed air station is jointly controlled. This significantly reduces reliance on manual experience-based scheduling and improves adaptability. Furthermore, it fully utilizes air consumption and electricity price forecasts, balancing unit air consumption energy consumption, comprehensive operating cost, equipment lifespan, and carbon emission intensity. This improves the rationality of start-up and shutdown of each air compressor and load allocation, enhancing the stability and robustness of the entire electric compressed air station operation, and achieving simultaneous optimization of energy saving, health, and reliability. This method can be flexibly applied to different sites or load scenarios, possessing good industrialization and self-evolution capabilities. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the multi-machine control method for an electrically driven compressed air station provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of the multi-machine control device for the electric-driven compressed air station provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a flowchart illustrating a multi-unit control method for an electrically driven compressed air station provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: S1, Construct a digital twin model of the electric-driven compressed air station; the digital twin model includes a cascaded coupling structure of the data-driven model of each air compressor unit in the electric-driven compressed air station and the pressure response model at the end of the pipeline network. S2, Based on the digital twin model, an offline training environment is built, and the multi-machine control task of the electric compressed air station is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and a supervised learning algorithm and an imitation learning algorithm are used to train the middle-level policy generator offline. Each optimization objective includes a unit compressed air energy consumption target, a comprehensive operating cost target, and an equipment lifespan and health target. S3, based on the trained upper-level energy management decision-maker, apply the current status information, air consumption prediction information and electricity price prediction information of the electric-driven compressed air station to decide the air compressor unit operation strategy and target range information of the electric-driven compressed air station. S4. Based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information are applied to constrain and optimize the air compressor unit operation strategy, generate a control strategy, and based on the bottom-level execution controller, the control strategy is applied to control the electric-driven compressed air station.

[0024] Specifically, the multi-machine control method for an electric compressed air station provided in this embodiment of the invention is executed by a multi-machine control device for the electric compressed air station. This device can be configured in the central control system of the electric compressed air station. The central control system can run on a computer, which can be a local computer or a cloud computer. The local computer can be a computer, tablet, etc., without specific limitations.

[0025] First, execute step S1 to construct a digital twin model of the electrically driven compressed air station. Here, an electrically driven compressed air station refers to a system that uses electric motors to drive multiple air compressors to provide compressed air to a factory or workshop. It may include air compressor units, air tanks, piping networks, end-of-line air-using equipment, and related auxiliary equipment. The air compressor unit may include an air compressor, drive motor and starting equipment, a unit-attached cooling system, and intake / exhaust piping and local auxiliary equipment tightly coupled to the unit.

[0026] Air compressors can include fixed frequency screw air compressors, variable frequency screw air compressors, centrifugal air compressors, etc. Different air compressors can have different rated power and rated flow.

[0027] Drive motors and starting devices may include electric motors, soft starters, or frequency converters.

[0028] The unit's auxiliary cooling system may include cooling water pumps, cooling fans, oil coolers, etc., directly associated with the air compressor unit. Additionally, it may include electric heating equipment.

[0029] The intake / exhaust piping and local auxiliary equipment closely coupled to the unit may include silencers, local filters, and unit outlet check valves that are close to the unit.

[0030] Pipeline networks can include trunk networks and branch networks.

[0031] Pneumatic equipment at the end of a pipeline network can include multiple production lines, intermittent pneumatic equipment, and pulsed pneumatic equipment. Multiple production lines are pneumatic units at the workshop / production line level, and can include stamping production lines, welding production lines, painting production lines, assembly lines, etc. Each production line may contain multiple continuous pneumatic equipment and intermittent / pulsated pneumatic equipment. Continuous pneumatic equipment can include pneumatic actuators, pneumatic clamps, spray guns, etc.

[0032] Related auxiliary equipment may include waste heat recovery systems. A waste heat recovery system refers to equipment or systems that recover and utilize high-temperature waste heat from the air compressor unit and its auxiliary cooling system. This may include: waste heat recovery heat exchangers in the unit's oil or cooling water circuits, and heat exchange systems that use the recovered heat for domestic hot water, process hot water, or low-temperature heating processes.

[0033] A digital twin model is a system-level discrete-time simulation model that integrates black-box or gray-box data-driven methods. It constructs a virtual environment reflecting the actual operating behavior of an electrically driven compressed air station by collecting and modeling operational data from multiple air compressor units and real-time and historical information on pipeline terminal pressure. This digital twin model employs a hierarchical cascaded coupling structure, including cascaded coupling structures of data-driven models for each air compressor unit and pipeline terminal pressure response models, providing dynamic interactive simulation support for subsequent control. The data-driven models of each air compressor unit and the pipeline terminal pressure response models are cascaded and coupled by using the output of the previous stage as the input of the next stage, thus forming a unified digital twin model capable of characterizing the dynamic correlation between unit scheduling, total air supply capacity, and pipeline pressure distribution. During the discrete-time simulation, at each moment, the unit-level response of each air compressor unit is first calculated according to the air compressor unit operating strategy. Then, the unit-level responses are aggregated and input into the pipeline terminal pressure response model to obtain the system-level pressure response, which is used as the basis for state updates, constraint judgments, and reward calculations at the next moment.

[0034] The data-driven model of an air compressor unit refers to an input-output function relationship model established using lightweight deep networks, gradient boosting trees, and other machine learning models, with historical operating data as the main information source. Given the unit's operating status, inlet / outlet pressure, inlet / outlet temperature, frequency parameters, operating mode code, and historical statistical characteristics, it outputs the predicted flow rate, predicted power, and / or predicted outlet temperature of the corresponding air compressor unit. The inputs to the data-driven model can include the operating status of the air compressor within the unit, inlet / outlet pressure, inlet / outlet temperature, frequency parameters, current operating mode code, and historical statistical characteristics. The outputs are unit-level response quantities such as predicted flow rate, predicted power, predicted efficiency, and predicted outlet temperature.

[0035] Lightweight deep networks can include 2-3 layer multilayer perceptrons (MLPs), and gradient boosting trees can be XGBoost / LightGBM, etc.

[0036] The operating status parameters of an air compressor can include its start / stop status, loading / unloading status, and the gear or fixed frequency setting of a line frequency compressor.

[0037] Inlet and outlet pressures include inlet intake pressure and outlet exhaust pressure. Inlet and outlet temperatures include inlet intake temperature and outlet exhaust temperature.

[0038] The frequency parameter is the variable frequency or equivalent speed of the air compressor unit.

[0039] The current operating mode code refers to the identification information of the current operating mode. The current operating mode can include local manual, local automatic, remote automatic, etc., and can also include constant pressure mode, constant flow mode, energy saving mode, etc.

[0040] Historical statistical features can include average flow rate, pressure, power, slope, variance, etc. over the previous 5 / 10 / 30 minutes, which are used to reflect the recent dynamic process of the air compressor unit.

[0041] The unit-level data-driven model for the kth air compressor unit can be represented as: ; in: Let be the input vector of the unit-level data-driven model of the k-th air compressor unit at time t; For predicting traffic flow; To predict power; To predict the outlet exhaust temperature; For the unit-level data-driven model of the k-th air compressor unit, These are the model parameters for the data-driven model of the k-th air compressor unit.

[0042] The pipeline terminal pressure response model aims to reflect the dynamic correspondence between the total output of the generating units and the pressure at the end of the pipeline. The total output of the generating units can be represented by the predicted flow rate of each air compressor unit, and the pressure at the end of the pipeline can be represented by the predicted pressure of each key node at the end of the pipeline. Here, key nodes are locations in the pipeline that have a significant impact on production or are sensitive to pressure fluctuations, such as remote nodes and branches supplying core equipment.

[0043] The pressure response model at the end of the pipeline network can be a system-level time series model, such as LSTM, Temporal Convolution Network, etc., which can capture complex dynamic correspondences such as gas pulsation, pipeline capacity buffer, and unit hysteresis.

[0044] In the digital twin model, the data-driven model for each air compressor unit calculates and encodes the unit-level response quantities corresponding to the inputs and outputs, thereby obtaining the total output characteristics and air compressor unit allocation structure information. The total output characteristics and air compressor unit allocation structure information, along with the air storage tank pressure and the flow rate at representative flow measurement points, are input into the pipeline end pressure response model. This model then outputs the predicted pressure at each key node.

[0045] The pressure response model at the end of the pipeline network can be expressed as: ; in: Let be the vector composed of the predicted pressure of each key node at time t; Assign structural information to the air compressor unit at time t; Let t be the pressure of the gas storage tank at time t; Let be the flow rate at a representative pipeline flow measurement point at time t; g is the pressure response model at the end of the pipeline. These are the model parameters for the pressure response model at the end of the pipeline network.

[0046] Furthermore, the digital twin model can be uniformly represented as: ; in, For digital twin models, The operating strategy of the air compressor unit at time t.

[0047] The digital twin model of the electric compressed air station needs to meet the constraints of the electric compressed air station, including pressure constraints, unit physical constraints and safety constraints. The pressure constraint is that the predicted pressure of each key node at the end of the pipeline network is greater than or equal to the lower pressure limit and less than or equal to the upper pressure limit at each moment within the preset sampling period.

[0048] The physical constraints of the generating units include the upper and lower limits of single-unit flow rate, the upper and lower limits of single-unit frequency, the minimum start-up time, and the minimum downtime.

[0049] Safety constraints include prohibiting repeated start-ups and shutdowns within a short period of time and avoiding prolonged periods of near-full load operation for all units simultaneously.

[0050] Step S1, by constructing a digital twin model of the electrically driven compressed air station, provides a highly realistic virtual carrier for all subsequent steps, capable of reversing real operational risks. This effectively enhances the versatility of control strategies and their transferability in new scenarios, breaking the limitations of relying on empirical formulas or single-point trial algorithms, and laying the foundation for better overall energy efficiency management.

[0051] By using a pipeline end-pressure response model, subsequent upper-level energy management decision-makers can anticipate the comprehensive impact of air compressor unit operation strategies on pipeline end-pressure distribution and energy consumption. The air compressor unit operation strategy refers to the start-up and shutdown states and load redistribution strategies of different air compressor units determined by the upper-level energy management decision-makers. This includes the start-up and shutdown sequence and timing of air compressor units, main / auxiliary unit role switching logic, loading / unloading switching logic, and load redistribution strategies among units. The start-up and shutdown sequence and timing can include which air compressor unit starts first, which air compressor unit starts last, and under what pressure / air consumption conditions which air compressor unit is shut down. The main / auxiliary unit role switching logic can include which variable frequency compressor acts as the main unit responsible for pressure regulation and when to switch the main unit role to optimize energy efficiency or balance lifespan. The loading / unloading switching logic can include the pressure range within which fixed frequency compressors load / unload, and how different air compressor units alternate loading to reduce the number of start-ups and shutdowns. The load redistribution strategy between units can include how to allocate the frequency and load rate of each unit when multiple frequency converters coexist, so as to avoid all air compressor units being fully loaded at the same time for a long time.

[0052] Then, step S2 is executed to build an offline training environment using the digital twin model of the electric compressed air station. This offline training environment refers to a reinforcement learning and policy learning simulation platform built based on the digital twin model. It can reproduce real scenarios such as multi-machine scheduling, pipeline pressure, air consumption prediction, energy efficiency and economic constraints in an electric compressed air station, and provide a reproducible virtual time-series environment for algorithm simulation interaction, reinforcement learning, and supervised / imitation learning training.

[0053] Digital twin models use a preset sampling period Operation. At each moment within the preset sampling period, based on the air compressor unit operation strategy output by the upper-level energy management decision-maker, the data-driven model of each air compressor unit is invoked to calculate the unit-level response of each unit. Further, based on the unit-level response of each unit, the total output characteristics and unit allocation structure information of the units are determined. Using the total output characteristics, unit allocation structure information, air tank pressure, and flow rate at representative pipeline flow measurement points as inputs, the pipeline end pressure response model is invoked to output the predicted pressure of each key node. The predicted pressure of each key node is compared with the upper and lower pressure limits, and constraints such as start-stop intervals and minimum start-stop durations are judged to form a set of actionable actions after reward / penalty or safety filtering. This set of actionable actions is fed back to the next moment, thus forming an offline training environment for offline training and strategy evaluation.

[0054] The multi-unit control task of an electric compressed air station refers to the joint control task of each air compressor unit in the station. This multi-unit control task is transformed into multiple optimization objectives. These objectives may include unit compressed air energy consumption targets, overall operating cost targets, and equipment lifespan and health targets. The unit compressed air energy consumption target can be the system-level minimization of unit compressed air energy consumption. Unit compressed air energy consumption represents the average electrical energy consumed per unit of compressed air in the electric compressed air station, which can be expressed as the ratio of electrical energy consumed by the station to the total air consumption over a period of time. ; in, This represents the unit compressed air energy consumption for the statistical time period set T, with units of 1. , This represents the station-level power consumption of the electric compressed air station at time t, in kW. This represents the total station-level air volumetric flow rate of the electric compressed air station at time t, in units of... , The length of the preset sampling period, in hours (h).

[0055] The overall operating cost objective is to minimize overall operating costs, which need to take into account peak and off-peak electricity prices, demand charges, and outsourced maintenance costs.

[0056] The overall operating cost can be expressed as: ; Where C represents the total operating cost of the statistical time period set T. Let be the electricity price at time t. The maintenance cost is estimated based on the number of operating hours. Fines for pressure drops or overpressure.

[0057] Equipment life and health targets are used to improve equipment life and reliability. By controlling start-up and shutdown frequency, running time, overload time ratio, and temperature over-limit time, they can be introduced as life penalty items.

[0058] Furthermore, carbon emission and process steady-state related targets can be characterized through unit compressed air energy consumption targets and comprehensive operating cost targets. These targets are used to reduce carbon emission intensity while meeting workshop pressure and flow requirements. For example, carbon emission intensity per unit volume of air can be indirectly reflected through energy consumption and carbon emission factors; reducing energy consumption can reduce carbon emission intensity per unit volume of air. Penalties or constraints can be used to reflect process pressure stability, ensuring stable process pressure and reducing pressure fluctuations.

[0059] In summary, by introducing multiple optimization objectives, the multi-machine control task of the electric compressed air station not only considers energy consumption itself, but also multiple interrelated objectives such as operating costs, equipment lifespan and health, and process stability.

[0060] Subsequently, using the various optimization objectives, the optimization problem for the upper-level energy management decision-maker can be constructed: ; in, The operating strategy of the air compressor unit at time t. The normalized unit compressed air energy consumption at time t. The normalized overall operating cost at time t, This is the lifetime penalty term for normalized time t. All are dimensionless. These are all dimensionless weighting coefficients for the corresponding terms.

[0061] It can be obtained through the following formula: ; in, The energy consumption per unit of compressed air at time t. This is an energy consumption reference value, which can be the average energy consumption per unit of compressed air at historical moments, or a pre-defined value.

[0062] It can be obtained through the following formula: ; in, The total operating cost at time t, This is a cost reference value, which can be the average of the comprehensive operating costs at historical moments or a pre-given value.

[0063] It can be obtained through the following formula: ; in, The lifetime penalty term at time t. The lifespan reference value can be the average of lifespan penalties at historical moments or a pre-defined value.

[0064] It can be obtained through the following formula: ; ; ; ; ; in, For the start / stop penalty normalization term, For the load factor over-limit normalization term, This is the normalization term for exhaust temperature exceeding the limit. The minimum downtime violates the normalization term; This represents the total number of start-ups and shutdowns of all air compressor units within the statistical time period from the start of the time period to time t. This represents a reference value for the number of starts and stops, which can be taken as the maximum allowed number of starts and stops for the statistical period, the historical average, or an empirical threshold. This represents the load rate of the k-th air compressor unit at time t. This represents the upper limit of the load rate for the k-th air compressor unit. The reference value for overload rate can be the allowable overload margin, the difference between the empirical upper limit and the limit, or a preset positive number. This represents the outlet exhaust temperature of the k-th air compressor unit at time t. Let be the upper limit of the outlet exhaust temperature of the k-th air compressor unit. This is a reference value for temperature exceeding limits; the allowable temperature rise margin or an empirical threshold can be used. This represents the downtime duration of the k-th air compressor unit from the start of the statistical period to time t within the period containing time t. Minimum downtime These are all penalty weights for the corresponding items.

[0065] The optimization problem is solved using Deep Reinforcement Learning (DRL) algorithm, and the upper-level energy management decision-maker is trained offline. The solution process for this optimization problem needs to satisfy the constraints of the electric compressed air station.

[0066] The inputs to the upper-level energy management decision-maker can include the current status information of the electric compressed air station, air consumption forecast information, and electricity price forecast information. The outputs are the air compressor unit operation strategy and target range information of the electric compressed air station.

[0067] The current status information of the electric compressed air station can include the current system status and unit status. The system status includes the pressure of each key node, total air consumption, air tank pressure and air tank filling level. The unit status includes the operating status, flow rate, power, unit energy consumption, start-stop count, cumulative running time, start-stop interval and fault / alarm flags of each air compressor unit.

[0068] Gas consumption forecast information refers to the predicted gas consumption over a specified future period.

[0069] Electricity price forecast information refers to the predicted electricity price curve for a specified future period.

[0070] The target range information may include the target load range or target frequency range for each air compressor unit.

[0071] Furthermore, supervised learning and imitation learning algorithms can be used to train the mid-level strategy generator offline. By learning expert control strategies, the operational details of each air compressor unit can be refined. The input to the mid-level strategy generator includes the air compressor unit operation strategy, target range information, and air consumption prediction information obtained from the trained upper-level energy management decision-maker. The output is a control strategy, which may include the frequency trajectory of each inverter, the loading / unloading switching time of each air compressor unit, and the target value in the target range information. The target value may include the target pressure or the target frequency.

[0072] After step S3 is executed, once the upper-level energy management decision-maker is trained, the current status information, air consumption prediction information, and electricity price prediction information of the electric-driven compressed air station are input into the trained upper-level energy management decision-maker. The trained upper-level energy management decision-maker can then be used to determine the operating strategy of the air compressor unit of the electric-driven compressed air station and the target range information.

[0073] Finally, in step S4, after the mid-level strategy generator is trained, the air consumption prediction information of the electric compressed air station, the air compressor unit operation strategy, and the target range information are input into the trained mid-level strategy generator. The trained mid-level strategy generator can be used to constrain and optimize the air compressor unit operation strategy and generate a control strategy.

[0074] Then, the control strategy is sent to the lower-level execution controller through the execution command queue, enabling the lower-level execution controller to apply the control strategy and control the electric-driven compressed air station. The lower-level execution controller can precisely control each air compressor unit in the electric-driven compressed air station using PID or MPC.

[0075] After that, the execution results of each air compressor unit in the electric compressed air station can be collected. If execution failure or timeout response is found, a rollback strategy is triggered, such as switching to local automatic mode.

[0076] The multi-unit control method for electric compressed air stations provided in this invention constructs a digital twin model, integrating multiple optimization objectives such as unit compressed air energy consumption, comprehensive operating cost, and equipment lifespan and health. It performs hierarchical offline training of the upper-level energy management decision-maker and the middle-level strategy generator. Then, using the trained upper-level energy management decision-maker, middle-level strategy generator, and lower-level execution controller, the electric compressed air station is jointly controlled. This significantly reduces reliance on manual experience-based scheduling and improves adaptability. Furthermore, it fully utilizes air consumption and electricity price forecasts, balancing unit air consumption energy consumption, comprehensive operating cost, equipment lifespan, and carbon emission intensity. This improves the rationality of start-up and shutdown of each air compressor and load allocation, enhancing the stability and robustness of the entire electric compressed air station operation. It achieves simultaneous optimization of energy saving, health, and reliability. This method can be flexibly applied to different sites or load scenarios, possessing good industrial implementation and self-evolution capabilities.

[0077] Based on the above embodiments, the step of applying a deep reinforcement learning algorithm to train the upper-layer energy management decision-maker offline in the offline training environment, based on each of the optimization objectives, includes: In the offline training environment, a reward function is constructed based on each of the optimization objectives, and a state space is constructed based on the current state information, the gas consumption prediction information, and the electricity price prediction information. An action space is constructed based on the air compressor unit operation strategy. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the electric compressed air station. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline.

[0078] Specifically, during the offline training of the upper-level energy management decision-maker, a reward function is first constructed using various optimization objectives. This reward function is used in reinforcement learning to evaluate the quality of actions taken by the upper-level energy management decision-maker, guiding it to learn towards the expected optimization objective. The reward function can include the overall energy consumption, operating cost, constraint violation degree, and start-stop regulation information of the electric-driven compressed air station. The constraint violation degree reflects the extent to which the electric-driven compressed air station exceeds pressure constraints, unit physical constraints, and safety constraint boundaries. The start-stop regulation information records or reflects the start-stop actions and frequency of the air compressor unit, and is used in the reward function to penalize frequent start-stops that affect equipment lifespan.

[0079] The reward function can be set as a weighted sum of its components. For example, the reward function can be calculated using the following formula: ; in, Let t be the reward value of the electric-driven compressed air station. The degree of constraint violation of the electric-driven compressed air station at time t. The normalized start-stop count of each air compressor unit in the electric compressed air station at time t, i.e., the start-stop regulation information, can be obtained by normalizing the maximum allowable start-stop count. and All are dimensionless. The dimensionless weighting coefficients are used to constrain the degree of violation.

[0080] It can be calculated using the following formula: ; in, For time t, the pressure constraint violation term is normalized. For time t, the hard constraint violation term is normalized. These are the weight coefficients for the corresponding items.

[0081] It can be calculated using the following formula: ; in, For the set of key nodes, For the number of critical nodes, Let be the predicted pressure at the j-th critical node at time t, in MPa or bar. This is the minimum pressure value. This is the maximum pressure value. The reference pressure for the j-th critical node can be the rated gas supply pressure, the design pressure, or the midpoint of the pressure range.

[0082] It can be calculated using the following formula: ; Where K represents the number of air compressor units. Let t be the duration of operation of the k-th air compressor unit within the statistical time period from the start of the time period to time t. This represents the minimum operating time of the k-th air compressor unit. This represents the shutdown duration of the k-th air compressor unit from the start of the statistical time period to time t within the period containing time t. This represents the minimum shutdown time for the k-th air compressor unit. This is an indicator function.

[0083] In addition, the reward function can also include a power change penalty term, which can explicitly punish excessive power fluctuations and guide the air compressor unit's operating strategy towards more stable operation.

[0084] A state space can be constructed using the current status information, air consumption forecast information, and electricity price forecast information of the electric compressed air station. This state space is a collection of environmental information that the upper-level energy management decision-maker can perceive at any given time, reflecting the current overall operating condition of the electric compressed air station and its future predictions. Furthermore, the state space can also include current time characteristics, such as shift, weekday, or weekend information.

[0085] By utilizing the air compressor unit's operating strategy, an action space can be constructed. This action space is the set of all possible actions that the upper-level energy management decision-maker can take at any given time, representing the macro-level scheduling instructions for the air compressor units. The action space can include discrete and continuous actions. Offline actions can include control actions for maintaining / starting / stopping each air compressor unit, while continuous actions can include setting the target frequency / load rate for each inverter and fine-tuning the upper and lower limits of the target range information. The upper-level energy management decision-maker can output both types of actions using a parameterized action space method or a multi-head action network.

[0086] After constructing the reward function, state space, and action space, a deep reinforcement learning algorithm is applied to allow the upper-level energy management decision-maker to continuously interact with the constructed digital twin model in an offline training environment. The network parameters of the upper-level energy management decision-maker are updated by maximizing the cumulative reward, thereby enabling offline training of the upper-level energy management decision-maker.

[0087] During training, the policy network and value network can be randomly initialized and run in an offline training environment to accumulate experience and update network parameters until they can stably meet the constraints and significantly reduce energy consumption in various typical day / time period patterns. The typical day / time period patterns can be obtained by clustering historical gas consumption curves of the electric compressed air station using clustering algorithms such as K-Means and spectral clustering.

[0088] Before outputting an action, a safety filter layer can be added: if the action violates hard constraints such as the minimum start-stop interval, then from the set of possible actions that satisfy the safety constraints, the set of actions closest to the current action is found to replace the previous action. Hard rules can also be injected: if the pressure at a critical node is significantly lower than the lower limit, the most effective high-efficiency inverter that can replenish the pressure at the critical node and is most suitable for starting in the current state must be started and operated in a rapid frequency ramp-up mode.

[0089] In the initial stage of the upper-level energy management decision-making system, it operates in parallel with suggestion mode and manual scheduling, recording human-machine differences; it feeds back operational data from the real environment to fine-tune the strategy; and it switches to semi-automatic / fully automatic mode when conditions are ripe.

[0090] In this embodiment of the invention, by constructing a reward function that includes comprehensive energy consumption, operating costs, constraint violation degree, and start-up / shutdown adjustment information, and combining it with deep reinforcement learning training in a comprehensive state space and action space, the upper-level energy management decision-maker can learn a comprehensive optimal multi-machine scheduling strategy that takes into account energy saving, cost reduction, safety, and equipment lifespan. This enables the upper-level energy management decision-maker to support complex strategies such as off-peak operation, early start-up of high-efficiency units, and avoidance of frequent start-ups and shutdowns, rather than relying solely on fixed rules.

[0091] Based on the above embodiments, the steps for constructing the data-driven model of each air compressor unit include: For any air compressor unit, obtain the operating status parameters of the air compressor in the air compressor unit, the inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical characteristics of the air compressor unit, and determine the actual flow rate and actual power of the air compressor unit. The operating status parameters, inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical features are input into the data-driven model of any air compressor unit to obtain the predicted flow rate and predicted power output by the data-driven model of any air compressor unit. Based on the actual flow rate, the actual power, the predicted flow rate, and the predicted power, the data-driven model of any air compressor unit is trained. Once the training process is complete, the construction process of the data-driven model of any air compressor unit is finished.

[0092] Specifically, the construction of the data-driven model for each air compressor unit constitutes the training process. For any air compressor unit, the operating status of the air compressor in any air compressor unit, the inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical characteristics can be obtained as inputs to the data-driven model of any air compressor unit, so as to obtain the predicted flow rate and predicted power output by the data-driven model of any air compressor unit.

[0093] Given the actual flow rate and actual power of any air compressor unit, the loss function can be calculated using the difference between the predicted and actual flow rates, and the difference between the predicted and actual power. The formula for calculating the loss function can be expressed as: ; in, This represents the loss value corresponding to air compressor unit k. The model parameters are the data-driven model parameters for air compressor unit k. Let K be the actual flow rate of air compressor unit k at time t. Let K be the predicted flow rate of air compressor unit k at time t. For reference traffic, Let k be the actual power of air compressor unit k at time t. For reference power, Let be the predicted power of air compressor unit k at time t. This represents the weighting coefficient for the power error term.

[0094] Using this loss function, the model parameters of the data-driven model for any air compressor unit can be iteratively optimized until a preset number of iterations is reached or the loss function converges, at which point the training process ends. At this point, the construction process of the data-driven model for any air compressor unit is complete.

[0095] In this embodiment of the invention, by using multidimensional information as input to the data-driven model, a reliable underlying device response basis is provided for system-level digital twin simulation and upper-level optimization decision-making, thereby improving the accuracy of control strategy formulation.

[0096] Based on the above embodiments, the inputs of the pipeline end pressure response model include the predicted flow rate of each air compressor unit, the air compressor unit allocation structure information, the air storage tank pressure of the electric compressed air station, and the flow rate of the representative pipeline flow measurement point of the electric compressed air station; the flow rate of the representative pipeline flow measurement point reflects the change in the gas consumption of the pipeline of the electric compressed air station. The output of the pipeline end pressure response model includes the predicted pressure of each key node of the electric compressed air station.

[0097] Specifically, the air compressor unit allocation structure information indicates which air compressor units are in operation and their load distribution ratios among multiple air compressor units. Representative pipeline flow measurement points refer to measurement points in the pipeline network that meet specified conditions. These conditions may include load coverage, operating condition diversity, geographical / topological characteristics, and historical data characteristics. Load coverage means that the branch or node where the measurement point is located covers a high proportion of air consumption in the gas-using equipment or production line, contributing significantly to the total system load. Operating condition diversity means that the gas consumption characteristics at the measurement point are typical, reflecting the changing patterns of different types of loads. Geographical / topological characteristics mean that the measurement point is located in a critical position in the pipeline network, such as the middle section of the main line, the starting point of a major branch, or a sensitive distant end, and is representative of the pressure distribution. Historical data characteristics mean that in historical data, the flow rate changes at this measurement point have a strong correlation with the total system load or other nodes, which can be used for load estimation and model training.

[0098] The flow rate at representative pipeline flow measurement points can be measured by ultrasonic / electromagnetic flow meters, which can better reflect the gas consumption changes and load characteristics of the entire pipeline network.

[0099] When applying the pipeline end-of-line pressure response model, the predicted flow rates of each air compressor unit, the air compressor unit distribution structure information, the air tank pressure of the electric compressed air station, and the flow rates at representative pipeline flow measurement points of the electric compressed air station are used as input features. The pipeline end-of-line pressure response model captures the complex dynamic processes of the electric compressed air station and ultimately outputs the predicted pressures of each key node of the electric compressed air station.

[0100] In this embodiment of the invention, by taking the total flow rate of the integrated unit, the distribution structure, the pressure of the gas storage tank, and the representative flow characteristics as inputs, the pipeline end pressure response model can accurately predict the pressure changes of each key node under complex pipeline networks. This enables the pipeline end pressure response model to have the ability to predict the comprehensive impact of multi-unit scheduling on the pressure distribution of the entire pipeline network. It overcomes the shortcomings of traditional single-unit energy efficiency curves that cannot reflect the spatiotemporal distribution and buffering characteristics of the pipeline network, and effectively ensures the pressure stability of key gas consumption points at the far end of the workshop.

[0101] Based on the above embodiments, the steps for constructing the data-driven model of any air compressor unit further include: Obtain the theoretical power and physical constraints of pressure and flow rate for any of the air compressor units; The theoretical power and the physical constraints of pressure and flow are used as mechanistic priors to update the input or training loss of the data-driven model of any air compressor unit.

[0102] Specifically, in the process of constructing a data-driven model for any air compressor unit, in addition to using a purely data-driven approach, the theoretical power and pressure-flow physical constraints of any air compressor unit can also be obtained. The theoretical power is the ideal power value under specific operating conditions calculated based on the thermodynamic principles of the compressor. The pressure-flow physical constraints are the physical limiting relationships between pressure and flow rate that follow the laws of fluid mechanics or pipeline characteristics.

[0103] Theoretical power can be calculated using the following formula: ; in, Let be the theoretical power of the k-th air compressor unit at time t, and n be the polytropic index. The motor efficiency of the k-th air compressor unit can be obtained from its nameplate. Let be the inlet pressure of the k-th air compressor unit at time t. Let be the outlet pressure of the k-th air compressor unit at time t. Let be the inlet volumetric flow rate of the k-th air compressor unit at time t.

[0104] The physical constraints of pressure and flow rate of an air compressor unit can be expressed by the following formula: ; in, The outlet pressure of the k-th air compressor unit is... Minimum outlet volumetric flow rate at that time. The outlet pressure of the kth air compressor unit is The maximum outlet volumetric flow rate at time k can be obtained by fitting the performance curve of the kth air compressor unit or historical outlet volumetric flow rate data.

[0105] The physical constraints of pressure and flow in a pipeline network can be approximated by the following formula: ; in, For the pressure drop of the pipeline network, Constraint coefficients, determined by factors such as pipe diameter, length, and roughness, or obtained from historical data, are used to constrain the rationality of predicted pressure and total flow rate changes, where Q is the flow rate in the pipeline network.

[0106] Subsequently, theoretical power and pressure / flow physical constraints can be introduced as mechanistic priors into the physical mechanisms and laws of the data-driven model to guide and constrain its learning process. These mechanistic priors can be added as additional physical features to the input of any air compressor unit's data-driven model, or theoretical power can be added as a mechanistic penalty term to the loss function of the data-driven model to update the training loss. The updated loss function can be expressed as: ; in, Here, ReLU represents the updated loss value corresponding to air compressor unit k, and ReLU is the activation function. These are the penalty weights corresponding to the activation functions.

[0107] In this embodiment of the invention, by incorporating prior mechanisms such as theoretical power and physical constraints into the input or training loss of the data-driven model, a gray-box modeling mechanism can be formed. While maintaining the strong data fitting ability of the data-driven model, it enhances its physical rationality and interpretability, improves the prediction accuracy and generalization performance of the data-driven model in scenarios with changing operating conditions or sparse data, and provides a more solid mechanistic support for the digital twin model.

[0108] Based on the above embodiments, the prediction step for gas consumption prediction information includes: Obtain historical air consumption information, production plan characteristics, time characteristics, and environmental characteristics of the electric compressed air station; The historical gas consumption information, the production plan characteristics, the time characteristics, and the environmental characteristics are input into the gas consumption prediction model to obtain the gas consumption prediction information output by the gas consumption prediction model.

[0109] Specifically, air consumption forecasting information can be obtained through an air consumption forecasting model. The forecasting process involves first acquiring historical air consumption information, production plan characteristics, time characteristics, and environmental characteristics of the electrically driven compressed air station. Production plan characteristics reflect the company's production scheduling, such as planned daily output and production line start-up / shutdown schedules, indirectly determining the demand for compressed air. Time characteristics identify different time periods, such as time slots, shifts, weekdays, or weekends, helping to identify periodic air consumption patterns. Environmental characteristics may include ambient temperature and humidity.

[0110] Subsequently, the collected historical gas consumption information, production plan characteristics, time characteristics, and environmental characteristics are uniformly input into the gas consumption prediction model. The gas consumption prediction model can process the multi-dimensional input features, learn their evolution patterns, and ultimately obtain the gas consumption prediction information output by the model. The gas consumption prediction model can be LSTM / GRU or a lightweight Transformer such as Informer or Temporal-Fusion-Transformer.

[0111] In this embodiment of the invention, gas consumption prediction is made by comprehensively utilizing multi-dimensional features such as historical gas consumption, production plans, time and environment. This can accurately capture the long-term trend and short-term fluctuations of gas consumption behavior, significantly improve the accuracy of multi-step gas consumption prediction, and provide a prerequisite for the intelligent scheduling system to achieve proactive peak shifting and optimized start-up and shutdown.

[0112] Based on the above embodiments, the step of applying the control strategy to control the electrically driven compressed air station based on the underlying execution controller then includes: Calculate the instantaneous efficiency of each air compressor unit, and calculate the relative efficiency deviation based on the instantaneous efficiency; Based on the relative efficiency deviation, the health score of each air compressor unit is calculated, and the health score of each air compressor unit is statistically analyzed within a preset time period. If the health score of the target air compressor unit is lower than the preset score within the preset time period, the priority of the target air compressor unit will be reduced in the decision-making process of the upper-level energy management decision-maker after training.

[0113] Specifically, after controlling the electrically driven compressed air station, the instantaneous efficiency of each air compressor unit can be calculated. This instantaneous efficiency is the operating efficiency of the air compressor unit at the current moment, calculated based on real-time operating data such as power and flow rate, or estimated using a model.

[0114] Instantaneous efficiency can be calculated using the following formula: ; in, Let be the instantaneous efficiency of the k-th air compressor unit at time t.

[0115] The relative efficiency deviation can be calculated from the instantaneous efficiency. This relative efficiency deviation is the degree to which the instantaneous efficiency of the air compressor unit deviates from its design reference efficiency.

[0116] The relative efficiency deviation can be calculated using the following formula: ; in, Let be the relative efficiency deviation of the k-th air compressor unit at time t. This is the reference unit energy consumption of the k-th air compressor unit under design or historical high-efficiency conditions, in units of... .

[0117] Subsequently, the health score of each air compressor unit was calculated using the relative efficiency deviation. This health score is a quantitative indicator calculated using the relative efficiency deviation and other operating characteristics, used to reflect the health status and degree of degradation of the equipment.

[0118] The health score of each air compressor unit can be calculated using the following formula: ; in, Let the health score of the k-th air compressor unit at time t be denoted as . This is a clipping function used to restrict 'a' to between 'b' and 'c'. Let be the maximum efficiency deviation of the k-th air compressor unit. Let be the outlet exhaust temperature of the k-th air compressor unit at time t. This represents the upper limit of the outlet exhaust temperature of the k-th air compressor unit. This is the normalized reference value for the temperature exceeding the limit of the k-th air compressor unit. Let be the number of failures of the k-th air compressor unit from the start of the statistical time period to time t within the time period containing time t. This serves as a reference value for the number of failures, and can be taken as the upper limit of the allowed number of failures for the statistical period, the historical average, or an empirical threshold. These are the weight coefficients for the corresponding items.

[0119] In addition, instantaneous efficiency and instantaneous efficiency deviation can be input into the anomaly detection model to identify sudden efficiency degradation or failures in each air compressor unit. The anomaly detection model can be an isolated forest, a one-class SVM, or similar model.

[0120] The health score sequence of each air compressor unit is continuously statistically analyzed within a preset time period. If the average health score of a target air compressor unit within the preset time period is found to be lower than the preset score, it indicates that the target air compressor unit may have efficiency degradation or potential failure. In the next multi-machine scheduling decision made by the upper-level energy management decision-maker after training, the scheduling priority of the target air compressor unit is actively reduced to reduce its operating load or avoid its start-up.

[0121] In this embodiment of the invention, by calculating the instantaneous efficiency and the instantaneous efficiency deviation and evaluating the health score, the performance degradation of each air compressor unit can be detected in a timely manner, and the priority of sub-healthy units can be automatically reduced in the upper-level decision-making, so as to realize the linkage between operation optimization and equipment health management, effectively avoid the increase in energy consumption and the deterioration of faults caused by operation with defects, and extend the overall service life of system equipment.

[0122] Based on the above embodiments, the offline training of the mid-level policy generator based on supervised learning algorithms and imitation learning algorithms includes: The system obtains the air compressor unit operation strategy sample and target interval information sample obtained by the upper-level energy management decision-maker after training, and obtains the gas consumption change trend sample, the expert control strategy of each air compressor unit, and the working mode label. The air compressor unit operation strategy sample, the target interval information sample, and the air consumption change trend sample are input into the middle-level strategy generator to obtain the control strategy sample output by the middle-level strategy generator. Based on the underlying execution controller, the control strategy sample is applied to control the electric compressed air station, and the unit compressed air energy consumption and overall operating cost of the electric compressed air station are detected. The intermediate-level strategy generator is trained offline based on the expert control strategy, the control strategy samples, the unit compressed air energy consumption under different operating mode labels, and the comprehensive operating cost.

[0123] Specifically, during the offline training of the middle-level strategy generator, samples of air compressor unit operation strategies and target range information obtained by the upper-level energy management decision-maker after training can be obtained first, along with samples of air consumption change trends, expert control strategies of each air compressor unit, and working mode labels.

[0124] Expert control strategies are derived from historical data, which is used by experienced operators to control the system during days of excellent operation, obtaining actual frequency trajectories, load switching times, etc.

[0125] Operating mode labels are category labels used to identify the current operating mode of an electrically driven compressed air station. Operating modes can include: typical daily / typical time period patterns obtained from load clustering; operating modes formed by different control strategies under these load scenarios; traditional local automatic mode and AI control mode, etc.

[0126] Subsequently, the air compressor unit operation strategy sample, target range information sample, and air consumption change trend sample are used as input features and fed into the intermediate strategy generator to obtain the control strategy sample containing continuous actions output by the intermediate strategy generator. The control strategy sample may include the frequency trajectory of each air compressor unit, load switching time, etc.

[0127] Then, in the offline training environment, the underlying execution controller applies the control strategy sample to control the electric compressed air station, and detects and calculates the unit compressed air energy consumption and overall operating cost of the electric compressed air station under the simulation operation.

[0128] Finally, the mean square error between the expert control strategy and the control strategy samples is calculated, and an energy consumption regularization term is constructed using the unit compressed air energy consumption and the comprehensive operating cost. A loss function is constructed based on the mean square error and the energy consumption regularization term, and the mid-level policy generator is trained offline using this loss function.

[0129] During offline training, samples corresponding to operating modes with low unit compressed air energy consumption and low overall operating costs can be selected, and the weight of these samples can be increased during offline training. Simultaneously, starting with a policy network, reinforcement learning fine-tuning can be incorporated into the offline training environment to explore energy-saving potential beyond human experience.

[0130] In this embodiment of the invention, the expert control strategy is combined with imitation learning, and working mode labels are introduced for enhanced fine-tuning. This enables the mid-level strategy generator to inherit reliable human experience and explore better energy-saving space in simulation, achieving a smooth transition from human-machine collaboration to data-driven operation. This ensures the safety of the mid-level fine-grained control strategy and the energy-saving effect of superhuman experience.

[0131] like Figure 2 As shown, based on the above embodiments, this embodiment of the invention provides a multi-unit control device for an electrically driven compressed air station, comprising: Model building module 21 is used to build a digital twin model of the electric-driven compressed air station; the digital twin model includes the data-driven model of each air compressor unit in the electric-driven compressed air station and the pressure response model at the end of the pipeline network; The offline training module 22 is used to build an offline training environment based on the digital twin model, and to transform the multi-machine control task of the electric compressed air station into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and a supervised learning algorithm and an imitation learning algorithm are used to train the middle-level policy generator offline. Each optimization objective includes a unit compressed air energy consumption objective, a comprehensive operating cost objective, and an equipment lifespan and health objective. The upper-level decision module 23 is used to make decisions on the air compressor unit operation strategy and target range information of the electric-driven compressed air station based on the trained upper-level energy management decision-maker and by applying the current status information, air consumption prediction information and electricity price prediction information of the electric-driven compressed air station. The control module 24 is used to optimize the operation strategy of the air compressor unit based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information, generate a control strategy, and control the electric-driven compressed air station based on the bottom-level execution controller and the control strategy.

[0132] Specifically, the functions of each module in the multi-machine control device of the electric-driven compressed air station provided in the embodiments of the present invention correspond one-to-one with the operation flow of each step in the above-mentioned method embodiments, and the achieved effects are also the same. For details, please refer to the above embodiments, and this will not be repeated in the embodiments of the present invention.

[0133] Figure 3An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the multi-machine control method of the electrically driven compressed air station provided in the above embodiments.

[0134] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the multi-machine control method of the electrically driven compressed air station provided in the above embodiments.

[0136] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the multi-machine control method for the electrically driven compressed air station provided in the above embodiments. This computer-readable storage medium can be either a non-transitory computer-readable storage medium or a transient computer-readable storage medium, and no specific limitation is made herein.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0138] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-unit control method for an electrically driven compressed air station, characterized in that, include: Construct a digital twin model of an electrically driven compressed air station; The digital twin model includes a cascaded coupling structure of data-driven models for each air compressor unit in the electrically driven compressed air station and pressure response models at the end of the pipeline network. Based on the digital twin model, an offline training environment is built, and the multi-machine control task of the electric compressed air station is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and the middle-level policy generator is trained offline based on supervised learning and imitation learning algorithms. The optimization objectives mentioned include unit compressed air energy consumption target, overall operating cost target, and equipment life and health target; Based on the trained upper-level energy management decision-maker, the current status information, air consumption prediction information, and electricity price prediction information of the electric-driven compressed air station are used to determine the air compressor unit operation strategy and target range information of the electric-driven compressed air station. Based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information are applied to constrain and optimize the air compressor unit operation strategy, generate a control strategy, and based on the bottom-level execution controller, the control strategy is applied to control the electric-driven compressed air station. The steps for constructing the data-driven model for each air compressor unit include: For any air compressor unit, obtain the operating status parameters of the air compressor in the air compressor unit, the inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical characteristics of the air compressor unit, and determine the actual flow rate and actual power of the air compressor unit. The operating status parameters, inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical features are input into the data-driven model of any air compressor unit to obtain the predicted flow rate and predicted power output by the data-driven model of any air compressor unit. Based on the actual flow rate, the actual power, the predicted flow rate, and the predicted power, the data-driven model of any air compressor unit is trained. Once the training process is complete, the construction process of the data-driven model of any air compressor unit is finished. The pipeline terminal pressure response model is a system-level time series model. The inputs of the pipeline terminal pressure response model include the predicted flow rate of each air compressor unit, the air compressor unit allocation structure information, the air storage tank pressure of the electric compressed air station, and the flow rate of the representative pipeline flow measurement point of the electric compressed air station. The flow rate of the representative pipeline flow measurement point reflects the change in the gas consumption of the pipeline of the electric compressed air station. The output of the pipeline end pressure response model includes the predicted pressure of each key node of the electric compressed air station.

2. The multi-unit control method for an electrically driven compressed air station according to claim 1, characterized in that, Based on the aforementioned optimization objectives, a deep reinforcement learning algorithm is applied to train the upper-layer energy management decision-maker offline in the offline training environment, including: In the offline training environment, a reward function is constructed based on each of the optimization objectives, and a state space is constructed based on the current state information, the gas consumption prediction information, and the electricity price prediction information. An action space is constructed based on the air compressor unit operation strategy. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the electric compressed air station. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline.

3. The multi-unit control method for an electrically driven compressed air station according to claim 1, characterized in that, The steps for constructing the data-driven model of any air compressor unit also include: Obtain the theoretical power and physical constraints of pressure and flow rate for any of the air compressor units; The theoretical power and the physical constraints of pressure and flow are used as mechanistic priors to update the input or training loss of the data-driven model of any air compressor unit.

4. The multi-unit control method for an electrically driven compressed air station according to any one of claims 1-3, characterized in that, The prediction steps for the gas consumption prediction information include: Obtain historical air consumption information, production plan characteristics, time characteristics, and environmental characteristics of the electric compressed air station; The historical gas consumption information, the production plan characteristics, the time characteristics, and the environmental characteristics are input into the gas consumption prediction model to obtain the gas consumption prediction information output by the gas consumption prediction model.

5. The multi-unit control method for an electrically driven compressed air station according to any one of claims 1-3, characterized in that, The process of controlling the electrically driven compressed air station based on the underlying execution controller and applying the control strategy includes: Calculate the instantaneous efficiency of each air compressor unit, and calculate the relative efficiency deviation based on the instantaneous efficiency; Based on the relative efficiency deviation, the health score of each air compressor unit is calculated, and the health score of each air compressor unit is statistically analyzed within a preset time period. If the health score of the target air compressor unit is lower than the preset score within the preset time period, the priority of the target air compressor unit will be reduced in the decision-making process of the upper-level energy management decision-maker after training.

6. The multi-unit control method for an electrically driven compressed air station according to any one of claims 1-3, characterized in that, The offline training of the mid-level policy generator based on supervised learning and imitation learning algorithms includes: The system obtains the air compressor unit operation strategy sample and target interval information sample obtained by the upper-level energy management decision-maker after training, and obtains the gas consumption change trend sample, the expert control strategy of each air compressor unit, and the working mode label. The air compressor unit operation strategy sample, the target interval information sample, and the air consumption change trend sample are input into the middle-level strategy generator to obtain the control strategy sample output by the middle-level strategy generator. Based on the underlying execution controller, the control strategy sample is applied to control the electric compressed air station, and the unit compressed air energy consumption and overall operating cost of the electric compressed air station are detected. The intermediate-level strategy generator is trained offline based on the expert control strategy, the control strategy samples, the unit compressed air energy consumption under different operating mode labels, and the comprehensive operating cost.

7. A multi-unit control device for an electrically driven compressed air station, characterized in that, include: The model building module is used to build a digital twin model of an electrically driven compressed air station; The digital twin model includes a cascaded coupling structure of data-driven models for each air compressor unit in the electrically driven compressed air station and pressure response models at the end of the pipeline network. The offline training module is used to build an offline training environment based on the digital twin model, and to transform the multi-machine control task of the electric compressed air station into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the upper-level energy management decision-maker offline in the offline training environment, and a supervised learning algorithm and an imitation learning algorithm are used to train the middle-level policy generator offline. Each optimization objective includes a unit compressed air energy consumption target, a comprehensive operating cost target, and an equipment lifespan and health target. The upper-level decision module is used to determine the operating strategy and target range information of the air compressor unit of the electric-driven compressed air station based on the trained upper-level energy management decision-maker and by applying the current status information, air consumption prediction information and electricity price prediction information of the electric-driven compressed air station. The control module is used to optimize the operation strategy of the air compressor unit based on the trained mid-level strategy generator, the air consumption prediction information and the target interval information, generate a control strategy, and control the electric-driven compressed air station based on the bottom-level execution controller by applying the control strategy. The steps for constructing the data-driven model for each air compressor unit include: For any air compressor unit, obtain the operating status parameters of the air compressor in the air compressor unit, the inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical characteristics of the air compressor unit, and determine the actual flow rate and actual power of the air compressor unit. The operating status parameters, inlet and outlet pressures, inlet and outlet temperatures, frequency parameters, current operating mode codes, and historical statistical features are input into the data-driven model of any air compressor unit to obtain the predicted flow rate and predicted power output by the data-driven model of any air compressor unit. Based on the actual flow rate, the actual power, the predicted flow rate, and the predicted power, the data-driven model of any air compressor unit is trained. Once the training process is complete, the construction process of the data-driven model of any air compressor unit is finished. The pipeline terminal pressure response model is a system-level time series model. The inputs of the pipeline terminal pressure response model include the predicted flow rate of each air compressor unit, the air compressor unit allocation structure information, the air storage tank pressure of the electric compressed air station, and the flow rate of the representative pipeline flow measurement point of the electric compressed air station. The flow rate of the representative pipeline flow measurement point reflects the change in the gas consumption of the pipeline of the electric compressed air station. The output of the pipeline end pressure response model includes the predicted pressure of each key node of the electric compressed air station.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-machine control method for an electrically driven compressed air station as described in any one of claims 1-6.

Images