Artificial intelligence-based learning compressor energy-saving control method and system

Abstract

This invention provides an AI-based learning-based energy-saving control method and system for compressors, relating to the field of compressor control technology. By constructing a digital twin model of the compressor system, a more realistic offline training environment can be provided for deep reinforcement learning algorithms. An error correction module is introduced into the digital twin model, which can continuously correct the performance of each unit using measured data, solving the problem that relying solely on factory characteristic curves is insufficient to accurately reflect long-term operation and complex working conditions. By transforming the energy-saving operation task of the compressor system into multiple optimization objectives, task decomposition is achieved, reducing the difficulty of determining the target strategy controller and lowering the cost of energy-saving control for the compressor system. This method introduces a hierarchical structure in the initial strategy controller to achieve layer-by-layer control, reducing the difficulty of energy-saving control of the compressor system. Furthermore, through the upper-level energy management controller, it addresses the problem of unreasonable online / offline operation and unit number switching, avoiding energy waste and carbon emission pressure.

Description

Technical Field

[0001] This invention relates to the field of compressor control technology, and in particular to an artificial intelligence-based learning-based compressor energy-saving control method and system. Background Technology

[0002] Air compressors, also known simply as air compressors, are widely used in industrial production such as ethylene plants, public works stations, data centers, and compressed air stations, especially in compressed air supply and the driving of pneumatic tools and equipment.

[0003] Air compressor systems often have the following characteristics: 1) The system includes multiple air compressors connected in parallel, and each air compressor will frequently start and stop and operate under overload; 2) The adjustment strategy of each air compressor in the system is usually PID control or simple rule control, which not only easily leads to the system operating under low load and reducing operating efficiency, but also has the problem of unreasonable connection or unloading and switching of the number of units, resulting in greater energy waste and carbon emission pressure. Summary of the Invention

[0004] This invention provides an artificial intelligence-based learning-based compressor energy-saving control method and system to address the deficiencies in related technologies.

[0005] This invention provides an artificial intelligence-based learning-based energy-saving control method for compressors, comprising: A digital twin model of a compressor system is constructed; the digital twin model includes a mechanistic model of the compressor system and an error correction module built based on machine learning. Based on the digital twin model, an offline training environment is built, and the energy-saving operation task of the compressor system is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. Based on the target strategy controller, the compressor system is subjected to online closed-loop control; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

[0006] According to the present invention, an artificial intelligence-based learning-based compressor energy-saving control method is provided, wherein, based on the optimization objectives, a deep reinforcement learning algorithm is applied to perform offline training on an initial policy controller in an offline training environment to obtain a target policy controller, comprising: 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 measured data and prediction information. An action space is constructed based on the unit operation information of the compressor system. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the compressor system. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the initial policy controller offline to obtain the target policy controller.

[0007] According to the present invention, an artificial intelligence-based learning-based compressor energy-saving control method is provided, wherein the compressor system is subjected to online closed-loop control based on the target strategy controller, comprising: An online efficiency evaluation is performed on the compressor system to obtain the evaluation results, and the target strategy controller is optimized based on the evaluation results.

[0008] According to the present invention, an artificial intelligence-based learning-based compressor energy-saving control method is provided, wherein the reward function is calculated based on the following formula: ; in, Let t be the reward value of the compressor system at time t. Let be the total energy consumption of the compressor system at time t. Let be the operating cost of the compressor system at time t. Let t represent the degree of constraint violation of the compressor system. Let t be the number of start-stop cycles or adjustment parameters of each unit in the compressor system at time t. These are the weights.

[0009] According to the present invention, an artificial intelligence-based learning compressor energy-saving control method is provided, wherein the unit operation information includes the start-stop information, operating parameters, and system-level operation mode of each unit in the compressor system; The system-level operating modes include multi-unit parallel high-voltage mode, high-efficiency unit priority mode, single-unit deep frequency conversion mode, and shutdown pressure preservation mode.

[0010] According to the present invention, an artificial intelligence-based learning compressor energy-saving control method is provided, wherein the measured data includes system-level status data and unit-level status data of the compressor system, and the data types of the system-level status data and the unit-level status data both include pressure, temperature and flow rate.

[0011] According to the present invention, an artificial intelligence-based learning compressor energy-saving control method is provided, wherein the optimization objectives include safety and process constraints, energy and economic objectives, and flexibility objectives.

[0012] This invention also provides an artificial intelligence-based learning-type compressor energy-saving control device, comprising: A model building module is used to build a digital twin model of the compressor system; the digital twin model includes a mechanistic model of the compressor system and an error correction module built based on machine learning; The offline training module is used to build an offline training environment based on the digital twin model, and to transform the energy-saving operation task of the compressor system into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. A closed-loop control module is used to perform online closed-loop control of the compressor system based on the target strategy controller; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

[0013] 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 artificial intelligence-based learning compressor energy-saving control method as described above.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based learning compressor energy-saving control method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the artificial intelligence-based learning compressor energy-saving control method as described above.

[0016] This invention provides an AI-based learning-based compressor energy-saving control method and system. By constructing a digital twin model of the compressor system, it provides a more realistic offline training environment for deep reinforcement learning algorithms. Introducing an error correction module into the digital twin model allows for continuous correction of each unit's performance using measured data, addressing the problem that factory characteristic curves alone cannot accurately reflect long-term operation and complex working conditions. By transforming the compressor system's energy-saving operation task into multiple optimization objectives, task decomposition reduces the difficulty of determining the target strategy controller, thereby lowering the cost of energy-saving control for the compressor system. The method introduces a hierarchical structure in the initial strategy controller to achieve layer-by-layer control, reducing the difficulty of energy-saving control of the compressor system. Through an upper-level energy management controller, it addresses issues of unreasonable online / offline operation and unit number switching, avoiding energy waste and carbon emission pressure. This method not only reduces the compressor system's energy consumption and avoids frequent start-stop or overload operation of compressors, but also prevents the compressor system from operating under low load, improving its operating efficiency. Furthermore, the method incorporates predictive information from the compressor system, enabling it to adapt to multi-dimensional dynamic changes in predictive information, achieving greater energy-saving potential while ensuring safety and interpretability. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart illustrating the AI-based learning-based compressor energy-saving control method provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of the AI-based learning compressor energy-saving control device provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] 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.

[0022] Figure 1 This is a flowchart illustrating an artificial intelligence-based learning-based compressor energy-saving control method 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 compressor system; the digital twin model includes a mechanistic model of the compressor system and an error correction module built based on machine learning; S2. Based on the digital twin model, an offline training environment is built, and the energy-saving operation task of the compressor system is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. S3, Based on the target strategy controller, perform online closed-loop control on the compressor system; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

[0023] Specifically, the AI-based learning compressor energy-saving control method provided in this embodiment of the invention is executed by an AI-based learning compressor energy-saving control device, which can be configured in a computer. The computer can be a local computer or a cloud computer. The local computer can be a computer, tablet, etc., and no specific limitation is made here.

[0024] First, perform step S1 to construct a digital twin model of the compressor system. This compressor system can be a physical compressor station, and its energy efficiency targets can include isentropic work per unit compressed air, comprehensive power usage effectiveness (PUE) metrics, etc. If the gas source is process gas, such as cracked gas, the energy efficiency targets can also additionally consider the coupling between process constraints and downstream energy consumption.

[0025] The compressor system may include multiple units, each of which includes a compressor, a drive unit, and a cooling unit. The drive unit, acting as the driving end, is connected to the compressor via a coupling and located on the same shaft system, forming a drive-compressor unit. After compressing the intake air, the compressor discharges high-temperature, high-pressure air from the unit outlet. The discharged high-temperature, high-pressure air is cooled sequentially by an intercooler and / or an aftercooler before flowing into a header and an air storage tank. A cooling water system is connected to the other side of the intercooler: the cooling water is circulated by a cooling water pump, exchanges heat with the environment via a cooling tower fan, and then returns to the source.

[0026] From the perspective of energy flow, the drive equipment is located upstream of the compressor shaft system, and the cooling water pump / cooling tower fan is located in the heat exchange circuit on the compressor exhaust side.

[0027] The compressor can be a centrifugal air compressor or a screw air compressor, and the compressor can include multiple compression stages. Therefore, the compressor can be a multi-stage centrifugal air compressor or a multi-stage screw air compressor.

[0028] The drive equipment may include at least one of a steam turbine and an electric motor, and the cooling equipment may include a heat exchanger, a cooling water pump, and a cooling tower fan. In this embodiment of the invention, the compressor system may be a Steam Turbine – Multi-Stage Compressor System (SDMSCS).

[0029] The digital twin model is a fusion architecture based on the mechanism model of the compressor system and using an error correction module built on machine learning to dynamically correct deviations in the mechanism model.

[0030] The mechanistic model of a compressor system is its physical model, which can include a compressor module, a drive unit module, and a cooling unit module. The compressor module can be represented by a multi-stage, multi-state thermodynamic model. The model parameters of the multi-stage, multi-state thermodynamic model can change with the operating state of the compressor, thereby covering multiple operating points. The operating state of the compressor can include speed, load rate, intake state, gas properties, etc.

[0031] Each compression stage in the compressor module corresponds to a pressure ratio, enthalpy rise, temperature change, mean polytropic index, isentropic index, and polytropic head. Variable efficiency Mechanical efficiency Parameters such as these are estimated using state equations such as the Soave-Redlich-Kwong or Peng-Robinson to determine the thermal properties of the compression medium. Furthermore, the various compression stages within the compressor module are coupled through relationships such as mass conservation, energy conservation, and flow continuity.

[0032] Furthermore, to address the loss terms in the compressor module that are difficult to accurately analyze and model, several loss terms with coefficients are introduced. These coefficients can be identified through data learning and can be understood as weights or correction factors for each loss term. For example, the pressure drop or power of a single-stage compressor can be expressed as: P actual =P ideal +k1·ΔP 流道损失 + k2·ΔP 密封泄漏 ; Where k1 and k2 are coefficients to be learned, which can be obtained through P actual With P ideal P is obtained by fitting the deviation. actual To measure voltage drop or power, P ideal ΔP is the voltage drop or power calculated through a mechanistic model. 流道损失 For flow channel loss term, ΔP 密封泄漏 This is for sealing leaks.

[0033] The drive unit module may include at least one of a steam turbine module and an electric motor module. The steam turbine module can be represented by the Willan's linear model as follows: ;

[0034] in, (This represents the shaft power or equivalent power output of the i-th steam turbine under current operating conditions, in kW). is the slope coefficient of the Willan's linear model of the i-th steam turbine, representing the ability to generate effective work per unit steam flow. Let be the inlet steam mass flow rate of the i-th steam turbine, in kg / s. Let be the enthalpy drop of the i-th steam turbine under isentropic expansion conditions, expressed in kJ / kg. This represents the no-load loss or constant loss term for the i-th steam turbine, such as mechanical loss, in kW.

[0035] The motor module can be simplified to an efficiency curve. It can be obtained by fitting the measured data of the motor.

[0036] If the compressor system has dual drive or switching between steam turbine and motor, the logic is similar to Bassam's CAES-TES energy flow allocation. The drive equipment module can be abstracted as a shaft power supply source. The cost is determined by the unit price of steam, the price of electricity, and the emission factor. That is, the cost of the steam part is calculated based on the unit price of steam and the steam enthalpy, and the cost of the motor part is calculated based on the price of electricity and the efficiency. The total cost is obtained by weighted summation of the cost of the steam part and the cost of the motor part.

[0037] The cooling equipment module uses a heat exchanger to achieve energy balance, supports the calculation of the compressor's actual and variable efficiency, and includes the power of the cooling water pump and the power of the cooling tower fan in the overall energy consumption.

[0038] The error correction module, built upon machine learning, can utilize supervised learning models, such as lightweight algorithms like XGBoost or Random Forest, or deep models like Multilayer Perceptron (MLP) or Long Short-Term Memory (LSTM). The module can update model parameters periodically or online through incremental learning to adapt to equipment aging and changes in operating conditions. Furthermore, the module performs concept drift processing on the measured data of the compressor system, setting correction upper limits or safety boundaries to ensure physical consistency.

[0039] The measured data for a compressor system can include system-level and unit-level status data. Both types of data include pressure, temperature, and flow rate. System-level status data may include the compressor system's main supply pressure. Gas tank pressure The data includes the temperature and flow rate of the cooling water inlet and outlet. Unit-level status data can include the inlet and outlet pressures, temperatures at each stage, flow rates, steam flow rates, power, speed, and valve positions of each unit in the compressor system. Furthermore, unit-level status data also includes the current variable efficiency deviation of each unit, the number of start-ups and shutdowns of each unit within a certain time window, the historical cumulative number of start-ups and shutdowns, and the cumulative operating time.

[0040] The error correction module takes measured data from the compressor system as input, constructs error terms or parameter corrections based on the difference between the measured data and the predicted data calculated by the mechanistic model, and uses these corrections as learning targets to train the coefficients in the compressor module, thereby achieving automatic adjustment of the strength of the loss terms. After training, the error correction module can output error terms or parameter corrections based on new measured data from the compressor system, which can be used to correct the mechanistic model in real time.

[0041] The digital twin model constructed through the mechanism model and error correction module can have the ability to quickly simulate the energy consumption and state of the compressor system based on given control actions, providing a high-fidelity virtual environment for subsequent deep reinforcement learning (DRL) algorithms.

[0042] Next, step S2 is executed to build an offline training environment using a digital twin model. This offline training environment is the virtual environment required by DRL.

[0043] The energy-saving operation task of a compressor system refers to ensuring that the compressor system operates in a high-efficiency and energy-saving mode. This energy-saving operation task can be transformed into multiple optimization objectives.

[0044] The optimization objectives include safety and process constraints, energy and economic objectives, and flexibility objectives. Safety and process constraints are hard constraints, including the pressure range of the gas supply header, anti-surge operation boundaries for each unit, anti-overload operation boundaries, anti-overheating operation boundaries, unit start-up and shutdown frequency constraints, minimum dwell time constraints for mode switching, and start-up and shutdown cooling time constraints. Specifically, the unit start-up and shutdown frequency constraint specifies the maximum number of start-ups and shutdowns allowed for a single unit within a given time window. The minimum dwell time constraint for mode switching specifies that once a certain operating mode is activated, it must be maintained for at least a certain period before switching to other modes to prevent excessively frequent mode changes. The start-up and shutdown cooling time constraint specifies the minimum cooling time that the unit must meet between shutdown and restart to prevent equipment damage due to frequent start-ups and shutdowns at high temperatures.

[0045] In addition, hard constraints can include rule-based protection, such as forcibly activating a safety margin mode when pressure is detected to be approaching the lower limit. The safety margin mode can include activating all high-efficiency units and pre-setting valve openings.

[0046] The energy and economic objectives are soft constraints, including minimizing overall energy consumption, minimizing operating costs, and minimizing emissions. Overall energy consumption includes electricity consumption (kWh), steam consumption (tons), and cooling water pump / fan power consumption. Flexibility objectives include rapid response to load fluctuations and coordinated operation of the power grid / steam system.

[0047] Each optimization objective can be constructed as a weighted multi-objective function or a hierarchical objective. The hierarchical objective prioritizes safety and then optimizes the economy.

[0048] Furthermore, reward functions in deep reinforcement learning algorithms can be constructed using various optimization objectives. In an offline training environment, measured data and prediction information from the compressor system are randomly sampled, and uncertainty and noise can be injected to improve generalization. Subsequently, a deep reinforcement learning algorithm is used to interact the initial policy controller with the offline training environment, allowing for offline training of the initial policy controller to obtain the target policy controller. The initial policy controller can be a hierarchical structure, including an upper-layer energy management controller, a middle-layer prediction rule controller, and a lower-layer execution controller.

[0049] The upper-level energy management controller is used to make decisions regarding the compressor system's unit operation information based on measured data and forecast information. Here, the compressor system's forecast information can include load forecast curves for the next 1–4 hours, future electricity prices, future steam prices, and future carbon prices. Load forecast curves can be generated based on historical gas load data of the compressor system, process production plans, holiday / shift information, etc., using time series forecasting models, machine learning models, or rule-based forecasting methods. Future electricity prices can be predicted by connecting to the day-ahead / intra-day price curves or contract electricity prices published by the power grid. If the price curves or contract electricity prices fluctuate, short-term forecasts can also be made based on historical electricity price time series using statistical or machine learning methods. Future steam prices can be predicted by connecting to the day-ahead / intra-day price curves or contract steam prices published by the steam system. If the price curves or contract steam prices fluctuate, short-term forecasts can also be made based on historical steam price time series using statistical or machine learning methods.

[0050] Future carbon prices can be determined based on publicly available carbon market data and policy scenarios, or by using similar time series forecasting or scenario analysis methods.

[0051] The upper-level energy management controller treats the unit as a hybrid energy node. Its decision-making can include which units in the compressor system should be started, the appropriate target pressure ratio / speed range for each unit, and pre-scheduling strategies for the next 0.5–4 hours. Pre-scheduling strategies can include unit combination and start-up / shutdown plans, unit output and pressure ratio settings, gas tank and system pressure target curves, and electric drive and steam drive power allocation. Unit combination and start-up / shutdown plans can include a list of units planned to start in the next 0.5–4 hours, units to be shut down and their shutdown times, and unit rotation plans. Unit output and pressure ratio settings can include target load rates or target output ranges for each unit; for multi-stage compressors, target pressure ratios or speed ranges are set. Gas tank and system pressure target curves can include target gas tank pressures for different time periods and a mains target pressure curve that can be adjusted according to electricity price / load changes.

[0052] In the case of hybrid drive, the power allocation between electric drive and steam drive can include the proportion of steam and electricity in the total shaft power in the future period, and the optimization of the usage period and proportion of steam / electricity by combining predicted electricity prices and steam costs.

[0053] The unit operation information obtained by the upper-level energy management controller can include start-up and shutdown information, operating parameters, and system-level operating modes of each unit in the compressor system. These system-level operating modes include multi-unit parallel high-pressure mode, high-efficiency unit priority mode, single-unit deep frequency conversion mode, and shutdown pressure maintenance mode. The multi-unit parallel high-pressure mode is suitable for high loads or high safety redundancy requirements, where multiple units operate simultaneously to maintain high system pressure and redundancy. The high-efficiency unit priority mode is suitable for medium load conditions, prioritizing the operation of the most efficient unit while other units remain on standby or at low loads to improve overall efficiency. The single-unit deep frequency conversion mode is suitable for lower load conditions, operating only one high-efficiency frequency conversion unit whenever possible, using frequency conversion regulation to meet demand and reduce start-up and shutdown frequency. The shutdown pressure maintenance mode is suitable for nighttime or extremely low load periods, minimizing unit shutdown and maintaining pressure through gas storage tanks / pipelines.

[0054] The mid-level predictive rule controller uses forecast information as control signals, converting it into load and price label level ranges, which can include high-range / low-range periods. Based on this, a Predictive Rule-Based Control (PRBC) strategy is introduced to constrain and optimize unit operating information, generating control commands. During control decision-making, not only the current state is used, but also forecast information for a future period is input into the controller, allowing for the pre-formulation and / or online adjustment of a series of rules. This achieves forward-looking scheduling optimization while ensuring the interpretability of the rule logic.

[0055] During periods of low electricity / steam prices, the storage inertia / volume buffering capacity of the gas storage tanks and pipelines is utilized to compress more air in advance and store it in the compressor system. During periods of high electricity / steam prices, or high carbon emissions, the previously stored compressed air is consumed to reduce the immediate compression power output during high-price periods, thereby minimizing compressor output or switching to a more efficient unit.

[0056] The underlying actuator controller can receive control commands and use them to control the compressor system. The underlying actuator controller can continue to use mature PID or traditional MPC to control the compressor system, adjusting parameters such as compressor speed, valve opening, and pump / fan frequency.

[0057] In this embodiment of the invention, the time scale of the upper-layer energy management controller is 0.5 to 4 hours, the time scale of the middle-layer prediction rule controller is several minutes to tens of minutes, and the time scale of the lower-layer execution controller is seconds to tens of seconds.

[0058] Finally, step S3 is executed, employing policy distillation and online fine-tuning to migrate the control strategy from the target policy controller to the actual compressor system. The control strategy from the target policy controller is then used to perform online closed-loop control of the compressor system's control variables. These control variables may include the start / stop status, speed, load factor, inlet guide vane (IGV), return valve, vent valve opening, steam turbine extraction parameters or motor parallel switching, intercooling temperature, and cooling water pump / cooling tower fan frequency of each compressor in the compressor system.

[0059] Intercooling temperature mainly refers to the target set value of the compressed air outlet temperature after passing through the intercooler from each compression stage. By adjusting the frequency of the cooling water pump and the frequency of the cooling tower fan, the cooling water flow rate and cooling intensity are controlled to make the actual intercooler outlet air temperature as close as possible to the set value.

[0060] Policy distillation is a technique that compresses complex or multiple control policies into a smaller neural network. It integrates multiple control policies trained in large-scale simulation scenarios into a more compact, efficient, and easily deployable policy network. After the policy is deployed to the real system, it undergoes online fine-tuning. With a small learning rate and strict parameter update increments, the control policy is updated in small steps using collected real-world operating data to adapt to real-world factors such as equipment aging and changes in operating conditions. Simultaneously, a safety protection mechanism is implemented to prevent policy updates from causing operational instability.

[0061] The AI-based learning-based compressor energy-saving control method provided in this invention constructs a digital twin model of the compressor system, providing a more realistic offline training environment for deep reinforcement learning algorithms. Introducing an error correction module into the digital twin model allows for continuous correction of each unit's performance using measured data, addressing the problem that factory characteristic curves alone cannot accurately reflect long-term operation and complex working conditions. By transforming the compressor system's energy-saving operation task into multiple optimization objectives, task decomposition reduces the difficulty of determining the target strategy controller, thereby lowering the cost of energy-saving control for the compressor system. This method introduces a hierarchical structure in the initial strategy controller to achieve layer-by-layer control, reducing the difficulty of energy-saving control of the compressor system. Through an upper-level energy management controller, it addresses issues of unreasonable online / offline operation and unit number switching, avoiding energy waste and carbon emission pressure. This method not only reduces the compressor system's energy consumption and prevents frequent start-stop or overload operation of compressors, but also avoids operation of the compressor system under low load, improving the compressor system's operating efficiency. Furthermore, this method incorporates predictive information from the compressor system, enabling the compressor system to adapt to multi-dimensional dynamic changes in predictive information, achieving greater energy-saving potential while ensuring safety and interpretability.

[0062] Based on the above embodiments, the step of applying a deep reinforcement learning algorithm to perform offline training on the initial policy controller in the offline training environment, based on each of the optimization objectives, to obtain the target policy controller, 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 measured data and prediction information. An action space is constructed based on the unit operation information of the compressor system. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the compressor system. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the initial policy controller offline to obtain the target policy controller.

[0063] Specifically, when training the initial policy controller offline, a reward function is first constructed in the offline training environment using various optimization objectives.

[0064] The reward function includes the compressor system's overall energy consumption, operating costs, constraint violation degree, and start-stop regulation information. The constraint violation degree can be defined as the weighted sum of the various constraint violation quantities, which may include the lower limit of the main pipe pressure P. min Violation amount, upper limit of main pipe pressure P max The amount of violation and the upper limit of temperature T max Violation quantity, lower limit of traffic Q min The amount of violation, etc., P minThe amount of violation can be represented as viol_P min (t)=max{0,(P min –P(t)) / P min}, P max The amount of violation can be represented as viol_P max (t) = max{0,(P(t)–P max ) / P max}

[0065] The reward function can be calculated using the following formula: ; in, Let t be the reward value of the compressor system at time t. Let be the total energy consumption of the compressor system at time t. Let be the operating cost of the compressor system at time t. Let t represent the degree of constraint violation of the compressor system. The start-stop count or regulation index of each unit in the compressor system at time t, i.e., start-stop regulation information. These are the weights.

[0066] Furthermore, to avoid the initial policy controller frequently encountering hard constraints during the early stages of training, a safety barrier function can be added or a safe reinforcement learning (Safe RL) framework can be adopted. The safety barrier function is a set of hard rules based on experience and process constraints. For example, when predicting that a critical variable is approaching or about to exceed the safety boundary in the next step or several steps in the future, the safety barrier function will trim, modify, or completely replace the actions output by the DRL with a safety-assured action, such as prohibiting the shutdown of critical units, forcing the activation of high-efficiency units, or limiting load reduction. The safe reinforcement learning framework employs constrained policy optimization methods, such as Constrained Policy Optimization, which explicitly adds constraint terms to the reward function and uses methods such as Lagrange multipliers to control the expected value of constraint violations from exceeding a threshold.

[0067] Subsequently, a state space is constructed using measured data and predictive information; that is, the state vectors in the state space contain both measured data and predictive information. Using the unit operating information of the compressor system, an action space can be constructed. This action space can include discrete actions and continuous actions. Discrete actions can include the start-up and shutdown information of each unit and the system-level operating mode of the compressor system, while continuous actions can include the operating parameters of the compressor system. Here, operating parameters can include the speed of each unit, load factor, valve opening degree, cooling water pump / fan frequency, etc.

[0068] Finally, using the reward function, state space, and action space, a deep reinforcement learning algorithm is applied to train the initial policy controller offline, resulting in the target policy controller.

[0069] In this embodiment of the invention, a system-level operating mode is introduced into the action space. Mode selection can be treated as an action. Through simulation, the control strategies that can be expected to be learned include: during periods of high electricity price and low load, prioritizing the operation of the most efficient unit or a very small number of units, suspending inefficient units, and increasing the load rate; during periods of low electricity price and high load, slightly increasing the pressure of the gas storage tank or system pressure in advance for "pre-charging" to transfer some of the compressed power to the low-price electricity period; when equipment deteriorates / efficiency decreases, automatically reducing the output of the deteriorating units and transferring the load to units in better condition; during seasonal changes, such as when cooling efficiency decreases in summer, considering cooling energy consumption, the strategy tends to maintain a medium pressure ratio and moderately distribute the load among multiple units to avoid excessive cooling burden caused by a single unit having a high pressure ratio.

[0070] Based on the above embodiments, the online closed-loop control of the compressor system based on the target strategy controller includes: An online efficiency evaluation is performed on the compressor system to obtain the evaluation results, and the target strategy controller is optimized based on the evaluation results.

[0071] Specifically, when performing online closed-loop control on a compressor system, online efficiency assessments can be conducted periodically to obtain the assessment results. The objects of online efficiency assessment can include various levels of polyvariable efficiency and mechanical efficiency; therefore, the assessment results can include the assessed values ​​of various levels of polyvariable efficiency and mechanical efficiency.

[0072] Furthermore, by utilizing the evaluation results and combining them with the theoretical efficiency values ​​determined by the digital twin model, the efficiency offset rate and shaft power loss ratio can be obtained. Based on the evaluation results, efficiency offset rate, and shaft power loss ratio, the target strategy controller can be optimized. For example, if a significant drop in the efficiency of a unit is detected, its load weight can be automatically reduced, favoring other high-efficiency units to undertake the main compression task; if a system-level operating mode is found to have consistently low efficiency, the state-action-reward trajectory under that system-level operating mode can be analyzed offline, and reward weights can be updated or additional constraints can be added.

[0073] In this embodiment of the invention, online efficiency evaluation enables the target strategy controller to perceive the health status and energy efficiency level of each unit in real time, thereby achieving integrated synergistic optimization of energy efficiency, health, and control.

[0074] like Figure 2 As shown, based on the above embodiments, this embodiment of the invention provides an artificial intelligence-based learning-type compressor energy-saving control device, comprising: The model building module 21 is used to build a digital twin model of the compressor system; the digital twin model includes a mechanism model of the compressor system and an error correction module built based on machine learning. The offline training module 22 is used to build an offline training environment based on the digital twin model, and to transform the energy-saving operation task of the compressor system into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. The closed-loop control module 23 is used to perform online closed-loop control of the compressor system based on the target strategy controller; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

[0075] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in this embodiment of the invention, wherein the offline training module is specifically used for: 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 measured data and prediction information. An action space is constructed based on the unit operation information of the compressor system. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the compressor system. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the initial policy controller offline to obtain the target policy controller.

[0076] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in this embodiment of the invention, wherein the closed-loop control module is specifically used for: An online efficiency evaluation is performed on the compressor system to obtain the evaluation results, and the target strategy controller is optimized based on the evaluation results.

[0077] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in this embodiment of the invention has a reward function calculated based on the following formula: ; in, Let t be the reward value of the compressor system at time t. Let be the total energy consumption of the compressor system at time t. Let be the operating cost of the compressor system at time t. Let t represent the degree of constraint violation of the compressor system. Let t be the number of start-stop cycles or adjustment parameters of each unit in the compressor system at time t. These are the weights.

[0078] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in this embodiment of the invention includes the start-stop information, operating parameters, and system-level operating mode of each unit in the compressor system in the unit operation information. The system-level operating modes include multi-unit parallel high-voltage mode, high-efficiency unit priority mode, single-unit deep frequency conversion mode, and shutdown pressure preservation mode.

[0079] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in this embodiment of the invention includes system-level status data and unit-level status data of the compressor system. The data types of the system-level status data and the unit-level status data both include pressure, temperature and flow rate.

[0080] Based on the above embodiments, the artificial intelligence-based learning compressor energy-saving control device provided in the embodiments of the present invention includes optimization objectives including safety and process constraints, energy and economic objectives, and flexibility objectives.

[0081] Specifically, the functions of each module in the AI-based learning compressor energy-saving control device provided in this embodiment correspond one-to-one with the operation flow of each step in the above-mentioned method embodiment, and the achieved effects are also the same. Please refer to the above embodiments for details, and this will not be repeated in this embodiment.

[0082] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As 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 artificial intelligence-based learning compressor energy-saving control method provided in the above embodiments.

[0083] 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.

[0084] 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 artificial intelligence-based learning compressor energy-saving control method provided in the above embodiments.

[0085] 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 artificial intelligence-based learning compressor energy-saving control method 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 is not specifically limited herein.

[0086] 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.

[0087] 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.

[0088] 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. An energy-saving control method for a learning-based compressor based on artificial intelligence, characterized in that, include: Construct a digital twin model of the compressor system; The digital twin model includes a mechanistic model of the compressor system and an error correction module built based on machine learning; Based on the digital twin model, an offline training environment is built, and the energy-saving operation task of the compressor system is transformed into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. Based on the target strategy controller, the compressor system is subjected to online closed-loop control; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

2. The artificial intelligence-based learning compressor energy-saving control method according to claim 1, characterized in that, Based on the aforementioned optimization objectives, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller, 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 measured data and prediction information. An action space is constructed based on the unit operation information of the compressor system. The reward function includes the comprehensive energy consumption, operating cost, constraint violation degree, and start-stop adjustment information of the compressor system. Based on the reward function, the state space, and the action space, a deep reinforcement learning algorithm is applied to train the initial policy controller offline to obtain the target policy controller.

3. The artificial intelligence-based learning compressor energy-saving control method according to claim 2, characterized in that, The online closed-loop control of the compressor system based on the target strategy controller includes: An online efficiency evaluation is performed on the compressor system to obtain the evaluation results, and the target strategy controller is optimized based on the evaluation results.

4. The artificial intelligence-based learning compressor energy-saving control method according to claim 2, characterized in that, The reward function is calculated based on the following formula: ； in, Let t be the reward value of the compressor system at time t. Let be the total energy consumption of the compressor system at time t. Let be the operating cost of the compressor system at time t. Let t represent the degree of constraint violation of the compressor system. Let t be the number of start-stop cycles or adjustment parameters of each unit in the compressor system at time t. These are the weights.

5. The artificial intelligence-based learning compressor energy-saving control method according to any one of claims 1-4, characterized in that, The unit operation information includes the start-up and shutdown information, operating parameters, and system-level operation mode of each unit in the compressor system. The system-level operating modes include multi-unit parallel high-voltage mode, high-efficiency unit priority mode, single-unit deep frequency conversion mode, and shutdown pressure preservation mode.

6. The artificial intelligence-based learning compressor energy-saving control method according to any one of claims 1-4, characterized in that, The measured data includes system-level status data and unit-level status data of the compressor system. The data types of both the system-level status data and the unit-level status data include pressure, temperature, and flow rate.

7. The artificial intelligence-based learning compressor energy-saving control method according to any one of claims 1-4, characterized in that, The optimization objectives mentioned include safety and process constraints, energy and economic objectives, and flexibility objectives.

8. An artificial intelligence-based learning-type compressor energy-saving control device, characterized in that, include: The model building module is used to build a digital twin model of the compressor system; The digital twin model includes a mechanistic model of the compressor system and an error correction module built based on machine learning; The offline training module is used to build an offline training environment based on the digital twin model, and to transform the energy-saving operation task of the compressor system into multiple optimization objectives. Based on each optimization objective, a deep reinforcement learning algorithm is applied to train the initial policy controller offline in the offline training environment to obtain the target policy controller. A closed-loop control module is used to perform online closed-loop control of the compressor system based on the target strategy controller; The error correction module is used to dynamically correct the deviation of the mechanism model based on the measured data of the compressor system. The initial strategy controller includes an upper-level energy management controller, a middle-level prediction rule controller, and a lower-level execution controller. The upper-level energy management controller is used to determine the unit operation information of the compressor system based on the measured data and prediction information. The middle-level prediction rule controller is used to apply prediction rule control strategies based on the prediction information to constrain and optimize the unit operation information and generate control commands. The lower-level execution controller is used to receive the control commands and control the compressor system based on the control commands.

9. 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 artificial intelligence-based learning compressor energy-saving control method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the artificial intelligence-based learning compressor energy-saving control method as described in any one of claims 1-7.

Images