PSA oxygen production and nitrogen recovery cooperative control method and system

Abstract

The invention discloses a PSA oxygen production and nitrogen recovery cooperative control method and system, and relates to the technical field of gas separation. The method comprises the following steps: establishing a dynamic multi-target evaluation function comprising a plurality of performance indexes such as air compressor energy consumption, vacuum pump energy consumption, oxygen purity and nitrogen recovery rate, and collecting pressure, temperature, gas components and flow data in an adsorption tower in real time; in the adsorption stage, based on the real-time data and the current oxygen demand, a first machine learning model is used for generating and executing a dynamic pressure adjusting strategy for the air compressor; in a desorption stage, predicting an optimal vacuum degree curve based on historical desorption data by utilizing a second machine learning model, generating a start-stop and pumping rate control strategy of the vacuum pump according to the optimal vacuum degree curve, and determining the recovery rate of the vacuum pump according to a comparison result of the real-time nitrogen concentration and a preset recovery concentration threshold value. And desorption airflow is controlled to enter the nitrogen recovery pipeline or be directly discharged through the valve.

Description

Technical Field

[0001] This invention relates to the field of gas separation technology, and in particular to a method and system for the coordinated control of PSA oxygen production and nitrogen recovery. Background Technology

[0002] Pressure Swing Adsorption (PSA) is a mature and widely used gas separation technology, particularly important in industrial oxygen production. Its core principle utilizes the difference in adsorption capacity of porous adsorbents such as molecular sieves for different components (e.g., nitrogen and oxygen in air) in a gas mixture under varying pressures. By periodically performing a series of steps including pressure adsorption and depressurization desorption, the target gas (e.g., oxygen) is separated from impurity gases (e.g., nitrogen). In many applications, the high-concentration nitrogen desorbed from the adsorbent also has significant recycling value.

[0003] Currently, the control strategy of traditional PSA oxygen generation systems mainly relies on fixed programs pre-set based on experience or theoretical calculations. This control method typically employs fixed timing logic, pre-setting the duration of each process stage such as adsorption, pressure equalization, desorption, and purging, the timing of valve switching, and the operating parameters (such as outlet pressure and pumping rate) of key equipment (such as air compressors and vacuum pumps). For example, in the adsorption stage, the air compressor is usually set to operate at a constant high pressure to meet the maximum design oxygen production load; in the desorption stage, the vacuum pump operates according to fixed start-stop times and pumping power to complete the regeneration of the molecular sieve.

[0004] However, this traditional control method based on fixed parameters has revealed several inherent defects in practical applications: First, energy consumption is high and efficiency is low. Because the control strategy is static, the system cannot dynamically respond to actual operating conditions (such as real-time fluctuations in product oxygen demand). This means that even during periods of low oxygen demand, the air compressor and vacuum pump still operate at their maximum load settings, resulting in significant wasted energy consumption and a substantial increase in operating costs. Second, the system has poor adaptability and robustness. When external environmental conditions (such as ambient temperature, atmospheric pressure, and raw material air humidity) or internal system conditions (such as performance degradation of the molecular sieve due to long-term use) change, the fixed control program cannot adaptively adjust. This can lead to fluctuations in the purity and recovery rate of the product oxygen, or even failure to meet standards, making it difficult to maintain the entire system in its optimal, efficient, and economical operating range over the long term. Third, the nitrogen recovery process is poorly controlled. Traditional control methods often recover all gases generated during the desorption stage indiscriminately, ignoring the dynamic changes in gas component concentrations at the beginning and end of desorption. This approach may result in some low-purity gases mixing into the recovered nitrogen, reducing the overall quality and economic value of the recovered nitrogen and causing a decrease in resource utilization.

[0005] In summary, existing control methods for PSA oxygen production and nitrogen recovery systems generally suffer from high energy consumption, inability to adapt to changing operating conditions, and low resource utilization rates. Therefore, there is an urgent need in this field for a more advanced control method to achieve dynamic, intelligent, and collaborative optimization of the PSA system's operation, thereby effectively reducing energy consumption and improving overall economic efficiency while ensuring stable and reliable product quality. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for coordinated control of PSA oxygen production and nitrogen recovery, aiming to solve the technical problems of static control strategies, high energy consumption, inability to adapt to changes in operating conditions, and low comprehensive resource utilization rate in existing PSA systems.

[0007] In a first aspect, embodiments of the present invention propose a method for coordinated control of PSA oxygen production and nitrogen recovery, applied to a PSA oxygen production and nitrogen recovery coordinated control system comprising at least one adsorption tower, an air compressor connected to the inlet of the adsorption tower, a vacuum pump connected to the outlet of the adsorption tower, and a nitrogen recovery pipeline. The adsorption tower is filled with a molecular sieve for selective adsorption of nitrogen. The method includes: Establish a dynamic multi-objective evaluation function that includes multiple preset performance indicators, including at least the energy consumption of the air compressor, the energy consumption of the vacuum pump, the oxygen purity, and the nitrogen recovery rate. Multiple sensors deployed in the PSA oxygen production and nitrogen recovery co-control system are used to collect real-time data on pressure, temperature, gas composition, and flow rate inside the adsorption tower. During the adsorption stage, based on real-time collected data on pressure, temperature, gas composition and flow rate in the adsorption tower and the current oxygen demand, a dynamic pressure regulation strategy for the air compressor is generated and executed using the first machine learning model to dynamically optimize the multi-objective evaluation function. During the desorption phase, a second machine learning model is used to predict the optimal vacuum curve for the current desorption process based on historical desorption data. Based on the optimal vacuum curve, a start-up and shutdown control strategy for the vacuum pump and the pumping rate are generated and executed to dynamically optimize the multi-objective evaluation function. During the desorption stage, based on the comparison between the nitrogen concentration in the real-time collected gas composition data and the preset recovery concentration threshold, the desorption gas flow is controlled by valves to either enter the nitrogen recovery pipeline or be directly discharged.

[0008] Preferably, the first machine learning model is a deep reinforcement learning model; the step of generating and executing a dynamic pressure regulation strategy for the air compressor includes: The real-time collected data and current oxygen demand are used as the state input of the deep reinforcement learning model, and the optimization result of the multi-objective evaluation function is used as the reward signal. The deep reinforcement learning model outputs an output pressure command to control the air compressor to operate with the lowest power consumption while meeting the oxygen demand.

[0009] Preferably, the second machine learning model is a long short-term memory network model; the step of predicting the optimal vacuum curve for the current desorption process based on historical desorption data includes: By using a long short-term memory network model to analyze the temporal relationship between historical desorption curve data and nitrogen desorption efficiency data, the minimum negative pressure intensity and duration required for the molecular sieve to reach the regeneration threshold can be predicted. The steps of generating and executing start-up and pumping rate control strategies for the vacuum pump based on the optimal vacuum curve include: determining the start-up timing, stop-up timing, and pumping rate of the vacuum pump based on the optimal vacuum curve.

[0010] Preferably, the step of controlling the desorption gas flow into the nitrogen recovery pipeline or directly discharging it via a valve includes: When the nitrogen concentration in the desorption gas flow is detected to be higher than the preset recovery concentration threshold, the control valve opens and directs the desorption gas flow to the recovery storage tank connected to the nitrogen recovery pipeline. When the nitrogen concentration in the desorption gas stream is detected to be lower than or equal to the preset recovery concentration threshold, the control valve closes and the desorption gas stream is switched to the direct emission path.

[0011] Preferably, the method further includes: The pressure inside the recovery storage tank is monitored in real time, and the depressurization rate of the buffer tank connected to the adsorption tower is dynamically adjusted based on the pressure feedback from the recovery storage tank to maintain the smooth operation of the nitrogen recovery process.

[0012] Preferably, the PSA oxygen production and nitrogen recovery coordinated control system is a multi-tower system comprising multiple adsorption towers, and the method further includes: Using a fuzzy logic controller, the switching cycle and nitrogen recovery sequence of multiple adsorption towers are optimized collaboratively based on external economic signals and system operating status. The external economic signals include at least time-of-use electricity price information, and the system operating status includes at least the liquid level or pressure of the oxygen storage tank.

[0013] Preferably, the steps of using a fuzzy logic controller to collaboratively optimize the switching cycle and nitrogen recovery sequence of multiple adsorption towers based on external economic signals and system operating status include: When the time-of-use electricity price information is during a low-electricity period, extend the adsorption time of the adsorption tower currently in the adsorption stage; when the time-of-use electricity price information is during a high-electricity period, shorten the vacuum maintenance time of the adsorption tower currently in the desorption stage.

[0014] Preferably, the method further includes: Continuously monitor and record at least one performance degradation index used to characterize the molecular sieve performance in multiple consecutive working cycles; Using a transfer learning model trained on historical performance degradation index data, the performance degradation status of molecular sieves in one or more future working cycles is predicted. Based on the predicted performance degradation state, compensation control parameters are generated, and the control strategies of the first and second machine learning models are actively modified using the compensation control parameters.

[0015] Preferably, the performance degradation metric is a performance vector containing multiple parameters, which are selected from: Under the condition of producing oxygen of the same purity, the duration of the adsorption stage, the average pressure drop in the adsorption tower, the energy consumption per unit of oxygen production, or any combination thereof; The compensation control parameters include the compensation value for the desorption negative pressure intensity during the desorption stage and the adjustment value for the preset recovery concentration threshold.

[0016] Secondly, embodiments of the present invention propose a PSA oxygen generation and nitrogen recovery coordinated control system, including at least one adsorption tower, an air compressor connected to the inlet of the adsorption tower, a vacuum pump connected to the outlet of the adsorption tower, a nitrogen recovery pipeline, valves, and multiple sensors for collecting pressure, temperature, gas composition, and flow data within the adsorption tower. The system also includes a controller, which is communicatively connected to the air compressor, vacuum pump, valves, and multiple sensors. The controller includes: The dynamic multi-objective evaluation module is used to calculate an evaluation function value that includes air compressor energy consumption, vacuum pump energy consumption, oxygen purity, and nitrogen recovery rate based on real-time data collected by multiple sensors. The adsorption control module is used to generate and output dynamic pressure regulation commands to the air compressor based on the evaluation function value and the current oxygen demand during the adsorption stage using a first machine learning model. The desorption prediction module is used to predict the optimal vacuum curve based on historical desorption data using a second machine learning model during the desorption stage, and to generate start / stop commands for the vacuum pump and pumping rate control commands. The nitrogen recovery decision module is used to generate valve opening and closing control commands based on the comparison results between the real-time nitrogen concentration and the preset recovery concentration threshold. The controller includes a self-compensation module, which is configured as follows: Receive and store molecular sieve performance degradation indicators calculated from data collected by multiple sensors in multiple consecutive working cycles; Using a pre-trained transfer learning model, the future performance degradation status of molecular sieves is predicted based on historical performance degradation index data. Based on the predicted performance degradation state, compensation control parameters are generated and output to the adsorption control module and the desorption prediction module to actively correct their control commands. Beneficial effects

[0017] This invention establishes a dynamic multi-objective evaluation function incorporating multiple dimensions such as energy consumption, purity, and recovery rate. Combined with real-time system operation data, it utilizes a machine learning model to achieve global synergistic optimization of the pressure swing adsorption (PSA) oxygen production and nitrogen recovery processes. This overcomes the limitations of traditional control methods, which rely on fixed parameters, resulting in high energy consumption, poor adaptability to operating conditions, and low resource utilization. Specifically, in the adsorption stage, a dynamic pressure regulation strategy based on current oxygen demand is introduced, preventing the air compressor from operating in a constant high-pressure mode, thus significantly reducing unnecessary energy consumption. In the desorption stage, precise prediction of the optimal vacuum curve enables refined control of the vacuum pump, ensuring effective regeneration of the molecular sieve while avoiding energy waste caused by excessive vacuuming. Furthermore, the intelligent valve control mechanism based on real-time nitrogen concentration designed in this invention can intelligently divert the desorption gas flow, ensuring the purity of the recovered nitrogen and improving its effective utilization rate. Ultimately, this achieves a significant improvement in overall system energy efficiency and economic benefits while ensuring product gas quality. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a method for coordinated control of PSA oxygen production and nitrogen recovery; Figure 2 This is a schematic diagram of the functional modules of a PSA oxygen production and nitrogen recovery coordinated control device. Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the invention, but does not constitute a limitation on the invention.

[0020] It should be noted that in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying importance. Unless otherwise expressly specified and limited, the terms "connected" and "linked" should be interpreted broadly. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] For the first aspect, please refer to... Figure 1 This invention provides a method for coordinated control of PSA oxygen production and nitrogen recovery. This method aims to address the problems of static control strategies, high energy consumption, poor adaptability to operating conditions, and low resource utilization in existing pressure swing adsorption (PSA) systems. By introducing data-driven dynamic multi-objective optimization, machine learning prediction, and adaptive control, the method achieves global optimization of the system across multiple dimensions, including energy consumption, oxygen purity, and nitrogen recovery rate.

[0022] The method of this embodiment is applied to a PSA oxygen production and nitrogen recovery coordinated control system. The system includes at least: one or more adsorption towers internally filled with molecular sieves (e.g., 5A or LiLSX type zeolite molecular sieves) for selective nitrogen adsorption; an air compressor connected to the inlet flow path of the adsorption tower; a vacuum pump connected to the outlet flow path of the adsorption tower; and a nitrogen recovery pipeline for collecting high-purity nitrogen. The core of this method lies in the refined and intelligent coordinated control of the two key stages of adsorption and desorption in the PSA cycle through an integrated intelligent control system. The specific steps of the method of this embodiment will be described in detail below: S101: Establish a dynamic multi-objective evaluation function that includes multiple preset performance indicators, wherein the multiple performance indicators include at least the energy consumption of the air compressor, the energy consumption of the vacuum pump, oxygen purity, and nitrogen recovery rate.

[0023] To achieve a global, dynamic, and quantitative evaluation of the performance of the entire PSA system, this method first constructs a comprehensive dynamic multi-objective evaluation function. This function serves as the decision-making basis and objective guide for all subsequent optimization control strategies. Its purpose is to unify the economic (energy consumption) and technical (product quality, resource utilization) indicators of system operation into a quantifiable framework, thereby overcoming the problem of conflicting objectives and difficulty in coordinating optimization in traditional control methods.

[0024] S1011: Define performance metrics. The dynamic multi-objective evaluation function... It includes a set of preset performance indicators that comprehensively reflect the system's operating status. In this embodiment, these indicators are specifically defined as follows: Instantaneous energy consumption of air compressor This indicator is directly related to the energy cost of the adsorption stage. The air compressor is one of the main energy-consuming units in a PSA system. Its energy consumption can be directly measured based on real-time monitoring of motor input power, current, and voltage, or based on its operating pressure. and output flow Indirect calculations are performed using a thermodynamic model. A simplified model can be expressed as: , where the function This is the efficiency characteristic curve of the compressor, which has been pre-calibrated experimentally and stored in the controller. Instantaneous energy consumption of the vacuum pump. This indicator reflects the energy cost during the desorption phase. The energy consumption of the vacuum pump is related to the pumping rate. and the achieved vacuum level Closely related, and can also be quantified through real-time power monitoring or equipment performance models: Product oxygen purity This is a core indicator for measuring oxygen production performance, directly related to whether the product meets medical or industrial standards (for example, medical oxygen requires a purity of no less than 90%). This data is provided in real time by an online gas composition analyzer (such as a zirconia sensor) deployed on the product's oxygen outlet pipeline. Nitrogen recovery rate This indicator is used to evaluate the utilization efficiency of nitrogen resources during the desorption stage. It is defined as the ratio of the mass flow rate of pure nitrogen successfully recovered to the storage tank per unit time to the total mass flow rate of nitrogen contained in the air entering the system. The calculation formula is: in To recover pipeline flow, To recover nitrogen purity, This refers to the feed air flow rate. This represents the volume fraction of nitrogen in the air (usually taken as 0.78).

[0025] S1012: Construct a weighted evaluation function that integrates the above performance metrics into a single evaluation function through weighted summation. This facilitates the optimization algorithm's search for the best solution. The function takes the following form: ; in, , , , These are the weighting coefficients for each indicator, summing to 1, and all are positive values. These weights can be preset according to the priority of production tasks or dynamically adjusted online by operators. For example, in operating conditions highly sensitive to energy costs, the weights can be increased. and The weight of [the oxygen] should be increased; however, in scenarios requiring ultra-high purity oxygen, the weight should be increased. The weight.

[0026] This represents a normalization function, used to transform physical quantities with different dimensions to a unified range, thereby eliminating the influence of dimensional differences on the evaluation results. For example, for oxygen purity, a target purity can be set. and minimum acceptable purity The normalized value is For energy consumption indicators, a reasonable upper and lower limit can be set based on the equipment's rated power and historical operating data for normalization.

[0027] S102: The pressure, temperature, gas composition and flow rate data inside the adsorption tower are collected in real time by multiple sensors arranged in the PSA oxygen production and nitrogen recovery coordinated control system.

[0028] Real-time acquisition of pressure, temperature, gas composition, and flow rate data within the adsorption tower is fundamental to the intelligent control achieved by the method of this invention. To this end, various types of sensors are deployed at key locations in the PSA oxygen production and nitrogen recovery coordinated control system, forming a comprehensive real-time data acquisition network.

[0029] S1021: Sensor Deployment; Pressure sensors are installed at the top, middle, and bottom of each adsorption tower, as well as at the air compressor outlet, vacuum pump inlet, product oxygen storage tank, and nitrogen recovery storage tank, to accurately monitor dynamic pressure changes at various points in the system and provide real-time pressure profile information for the control model. Temperature sensors are arranged alongside the pressure sensors to monitor the thermal effects during adsorption and desorption. Since adsorption is an exothermic process and desorption is an endothermic process, temperature changes significantly affect the adsorption equilibrium and mass transfer rate of the molecular sieve; therefore, real-time temperature data is crucial for accurate modeling and control.

[0030] Gas component sensors include: an oxygen concentration sensor, installed at the product oxygen outlet, to monitor oxygen purity in real time, serving as one of the core control targets; a nitrogen concentration sensor, installed at the desorption gas outlet, to determine nitrogen purity, a key basis for nitrogen recovery decisions; and flow sensors, installed in the feed air line, product oxygen line, nitrogen recovery line, and purge gas line, to monitor the flow rate of each gas, forming the basis for calculating energy consumption, recovery rate, and material balance.

[0031] S1022: Data Processing and Transmission; All analog or digital signals acquired by sensors are preprocessed through the Data Acquisition (DAQ) module, including filtering, amplification, and analog-to-digital conversion, and then transmitted at a high frequency (e.g., 10 samples per second) to the central controller via industrial Ethernet or fieldbus (such as Profibus). The central controller can be an industrial PC (IPC) or a high-performance programmable logic controller (PLC), responsible for storing historical data, executing subsequent machine learning model calculations, and controlling the logic.

[0032] S103: During the adsorption stage, based on the real-time collected data on pressure, temperature, gas composition, and flow rate within the adsorption tower, and the current oxygen demand, a first machine learning model is used to generate and execute a dynamic pressure regulation strategy for the air compressor to dynamically optimize the multi-objective evaluation function. Traditional PSA systems typically employ a constant high-pressure operation mode during the adsorption stage, which results in significant energy waste when oxygen demand is low. This invention introduces a first machine learning model to dynamically adjust the compressor output pressure according to real-time operating conditions, achieving on-demand pressure supply and energy saving.

[0033] In a preferred embodiment, the first machine learning model is a deep reinforcement learning (DRL) model. DRL is particularly suitable for solving problems that require continuous decision-making in dynamic and uncertain environments to achieve optimal long-term goals, which is highly compatible with the control requirements of the PSA adsorption process.

[0034] S1031: Constructing a deep reinforcement learning control framework; State Defined as at time The DRL agent observes system environmental information. It is a multi-dimensional vector containing all key real-time data collected in S102 (pressure, temperature, composition, flow rate), as well as the current oxygen demand input from the upper-level production scheduling system. .Right now Action Defined as the DRL agent observing the state Subsequently, the control decision is output. In this embodiment, the action is aimed at the output pressure setpoint of the air compressor. The action space can be continuous (e.g., pressure values ​​between 4.5 and 7 bar) or discrete (e.g., {increase by 0.1 bar, remain unchanged, decrease by 0.1 bar}). Reward. Defined as an action performed by an intelligent agent. Subsequently, the system environment provides immediate feedback. The design of the reward function is crucial, as it directly guides the agent's learning direction. In this embodiment, the reward... Directly related to the dynamic multi-objective evaluation function established in S101 Hook. When When the value increases (meaning an improvement in overall system performance), a positive reward is given; conversely, a negative reward (penalty) is given.

[0035] S1032: Model Training and Execution; The training process of the DRL model is a trial-and-error learning process. The agent continuously interacts with the environment in a precise PSA process simulation environment, or in the safe operation mode of the actual system. At each time step, it selects an action (stress adjustment) based on the current policy network (a deep neural network), observes environmental changes (new states), and receives a reward. Then, using the obtained (state-action-reward-new state) tuple, the weights of the policy network are updated through algorithms such as Q-learning or Actor-Critic, making it more likely to select actions that bring higher long-term cumulative rewards in the future. After training, the DRL model is deployed to the central controller. During the adsorption phase of actual operation, the controller collects the states in real time. When input into the trained DRL model, the model will output an optimal compressor pressure setpoint in real time. This setpoint is sent as an instruction to the frequency converter that controls the compressor speed, thereby achieving dynamic, precise, and energy-saving pressure control.

[0036] S104: During the desorption stage, the second machine learning model is used to predict the optimal vacuum curve of the current desorption process based on historical desorption data, and the start-stop and pumping rate control strategies for the vacuum pump are generated and executed according to the optimal vacuum curve to dynamically optimize the multi-objective evaluation function.

[0037] The goal of the desorption stage is to use low pressure or vacuum conditions to desorb the nitrogen adsorbed on the molecular sieve, thereby regenerating the molecular sieve. Traditional control methods usually set a fixed vacuum pump operating time and pumping speed, which can easily lead to "over-negative pressure," that is, the degree of vacuum pumping exceeds the threshold required for complete regeneration of the molecular sieve, resulting in unnecessary energy waste.

[0038] In a preferred embodiment, the second machine learning model is a Long Short-Term Memory (LSTM) network model. LSTM is a special type of recurrent neural network (RNN) that is very good at handling and predicting time-series related problems and can effectively capture the complex dynamic relationships of variables such as pressure and concentration changing over time during desorption.

[0039] S1041: Construction and Training of the Long Short-Term Memory (LSTM) Network Model; Training Data: Extensive historical operational data from the PSA system was used for training. Each training sample is a time series, containing pressure changes, gas component (especially nitrogen) concentration changes, vacuum pump energy consumption data, and the final molecular sieve regeneration efficiency (which can be indirectly evaluated through the performance of subsequent adsorption cycles) within a complete desorption cycle. Model Input: During training, the LSTM model is input with multivariate time series data from historical desorption cycles. Model Output / Prediction Target: The LSTM model needs to predict the "optimal vacuum curve" for the current desorption cycle. This curve is not a single value, but a pressure target sequence that changes over time. The characteristic of this curve is that it can regenerate the molecular sieve to the preset regeneration threshold with the lowest energy consumption (i.e., the shortest vacuum pump run time and the most suitable pumping speed).

[0040] S1042: Predictive and Control Execution; At the start of the desorption phase of actual operation, the central controller inputs the current and past desorption data from the previous few cycles into the trained LSTM model. Based on this latest historical information, the model predicts the optimal vacuum curve for the current operating conditions. Subsequently, a lower-level controller (Proportional-Integral-Derivative, PID) or fuzzy controller generates a precise start / stop and pumping rate control strategy for the vacuum pump based on this predicted pressure curve. For example, controlling the output frequency of the vacuum pump inverter dynamically changes its pumping rate, allowing the actual pressure curve inside the adsorption tower to accurately track the optimal vacuum curve predicted by the LSTM. When the actual pressure reaches the end of the curve (i.e., the regeneration threshold), the controller promptly stops the vacuum pump, thus avoiding energy waste.

[0041] S105: During the desorption stage, based on the comparison result between the nitrogen concentration in the real-time collected gas component data and the preset recovery concentration threshold, the desorption gas flow is controlled by a valve to enter the nitrogen recovery pipeline or be directly discharged.

[0042] The nitrogen concentration in the gas discharged during desorption is dynamic. Initially, the discharged gas may contain a higher residual oxygen content, while later it consists of high-purity nitrogen. Indiscriminately recovering all of it will reduce the overall quality of the recovered nitrogen.

[0043] In a preferred embodiment, the method uses a smart valve to dynamically divert the desorption gas flow.

[0044] S1051: Intelligent valve control; a fast-switching valve (e.g., a three-way solenoid valve) controlled by a central controller is installed on the outlet pipeline of the desorbed gas flow. The pipeline splits into two paths here: one leads to the nitrogen recovery storage tank, and the other leads directly to atmospheric emission.

[0045] S1052: Decision logic based on concentration threshold; the controller monitors the nitrogen concentration at the desorption gas outlet collected in S102 in real time. The system presets a recovery concentration threshold. (For example, 95%). When the nitrogen concentration in the desorbed gas stream is detected to be higher than the preset recovery concentration threshold, the controller determines that the current gas stream has recovery value and immediately drives the valve to switch the gas stream to the nitrogen recovery pipeline. When the nitrogen concentration in the desorbed gas stream is detected to be lower than or equal to the preset recovery concentration threshold, the controller considers the gas stream purity insufficient and drives the valve to switch it to the direct discharge path to avoid contaminating the recovered high-purity nitrogen.

[0046] In a preferred embodiment, a feedback regulation mechanism based on the pressure of the recovery tank is introduced to ensure the stability of the nitrogen recovery process.

[0047] S1053: Pressure feedback and buffer tank depressurization rate regulation; real-time monitoring of pressure in the nitrogen recovery storage tank. When the desorbed gas stream is directed to the recovery storage tank, if the depressurization rate of the adsorption tower is too fast, it may cause a sudden surge in pressure in the recovery pipeline and storage tank, affecting system stability. Therefore, this method uses a buffer tank connected to the adsorption tower, and is based on... The feedback signal dynamically controls the depressurization rate of the buffer tank through a proportional regulating valve. If the pressure in the recovery tank rises too quickly, the depressurization rate is appropriately reduced, and vice versa, thereby achieving flexible control of the pressure in the recovery process and maintaining the stability of the nitrogen recovery process.

[0048] In more complex application scenarios, the present invention can be further expanded to address issues such as multi-tower collaborative operation and equipment aging.

[0049] In a preferred embodiment, when the PSA oxygen production and nitrogen recovery coordinated control system is a multi-tower system containing multiple adsorption towers, a fuzzy logic controller (FLC) is introduced for upper-level coordinated optimization.

[0050] The goal of the FLC (Flexible Cell Controller) is to dynamically adjust the switching cycle and nitrogen recovery sequence of each adsorption tower based on external economic signals (such as time-of-use electricity prices) and the overall system operating status (such as total oxygen reserves), achieving a higher level of economic operation. An example of a fuzzy rule is as follows: If the electricity price is in a "valley period" and the oxygen reserve is "low," then the switching cycle is "extended," and the adsorption phase is "lengthened." (This rule aims to utilize low-priced electricity to produce more oxygen and replenish the inventory.) If the electricity price is in a "peak period" and the oxygen reserve is "sufficient," then the switching cycle is "shortened," and the desorption phase vacuum time is "reduced." (This rule aims to minimize power consumption and reduce operating costs when the electricity price is high.) The FLC's output (such as "extended" or "shortened") is used as setpoints and passed to the underlying DRL and LSTM models to guide the generation of their specific control strategies.

[0051] In a preferred embodiment, to address the issue of performance degradation of molecular sieves after long-term use, the method further includes a self-compensation step for molecular sieve performance degradation based on transfer learning.

[0052] Performance degradation monitoring involves the system continuously recording and analyzing degradation indicators characterizing the molecular sieve performance. Preferably, the performance degradation indicator is a performance vector containing multiple parameters, selected from: the duration of the adsorption phase, the average pressure drop within the adsorption tower, and the energy consumption per unit of oxygen production, under the condition of producing the same oxygen purity, or any combination thereof. For example, if the system detects that the adsorption time needs to be extended by 10% year-on-year to reach the target purity, it determines that the molecular sieve performance has significantly degraded. Degradation state prediction utilizes a general prediction model pre-trained on a large amount of molecular sieve aging data as a foundation. Through transfer learning technology, the model is fine-tuned using a small amount of performance degradation data accumulated by the system itself, enabling it to accurately predict the future performance degradation trend of the molecular sieve in the system. Adaptive compensation, based on the predicted degradation state, automatically generates compensation control parameters. Preferably, the compensation control parameters include a compensation value for the desorption negative pressure intensity of the desorption phase and an adjustment value for the preset recovery concentration threshold. For example, if a decrease in adsorption capacity is predicted, the module will automatically increase the target vacuum degree of the desorption phase (enhancing regeneration intensity) or adjust the nitrogen recovery concentration threshold. These compensation parameters are updated in real time to the settings of the DRL and LSTM models, so that the system can still maintain near-optimal operating efficiency during the aging process of the equipment, realizing the self-correction of the control strategy and adaptability to the entire life cycle of the equipment.

[0053] In summary, this embodiment constructs a multi-layered, data-driven intelligent control architecture that organically integrates various advanced technologies such as dynamic multi-objective optimization, deep reinforcement learning, long short-term memory network prediction, fuzzy logic collaboration, and transfer learning self-compensation. This achieves unprecedentedly refined, adaptive, and globally collaborative control of the PSA oxygen production and nitrogen recovery process, thereby maximizing energy efficiency and resource utilization while ensuring product quality.

[0054] For the second aspect, please refer to... Figure 2 This invention also proposes a pressure swing adsorption (PSA) oxygen generation and nitrogen recovery coordinated control system. This system is adapted to the PSA oxygen generation and nitrogen recovery coordinated control method described in the foregoing embodiments, and can achieve efficient, intelligent, and integrated coordinated regulation of the PSA oxygen generation and nitrogen recovery processes. The system will be described in detail below with reference to specific embodiments.

[0055] The system comprises a physical equipment layer and an intelligent control layer. The physical equipment layer forms the foundation for gas separation and includes at least an adsorption tower, an air compressor connected to the tower inlet, a vacuum pump connected to the tower outlet, a nitrogen recovery pipeline, one or more valves for flow path switching, and a sensor network for comprehensive system status monitoring. The core of the intelligent control layer is a controller that communicates with all the aforementioned controllable physical equipment and sensors.

[0056] The adsorption tower is filled with molecular sieves, such as 5A or LiLSX type zeolite molecular sieves, for selective adsorption of nitrogen. An air compressor, preferably a screw compressor equipped with a variable frequency drive (VFD), is used to provide dynamically adjustable high-pressure air to the adsorption tower during the adsorption stage. A vacuum pump, also preferably equipped with a VFD, is used to evacuate the adsorption tower during the desorption stage to regenerate the molecular sieves. Valves, such as fast-response three-way solenoid valves or pneumatic regulating valves, are used to precisely switch the direction of the desorption gas flow according to control commands.

[0057] The aforementioned sensor network forms the cornerstone of data-driven control. Its specific deployment, as described in step S102 of the preceding embodiment, includes, but is not limited to, pressure sensors, temperature sensors, gas component sensors (such as zirconia oxygen concentration sensors and thermal conductivity nitrogen concentration sensors), and mass flow sensors deployed at key nodes such as the adsorption tower, compressor outlet, vacuum pump inlet, product outlet, and recovery pipeline. These sensors transmit high-frequency acquired real-time data to the controller via an industrial fieldbus (such as Modbus TCP / IP or PROFIBUS-DP).

[0058] The controller is the core of this system and can be a high-performance industrial PC (IPC) or a PLC integrating advanced computing modules. Multiple collaborative software modules are deployed within the controller to execute the complex control logic described in the preceding embodiments. Specifically, the controller includes: a dynamic multi-objective evaluation module, an adsorption control module, a desorption prediction module, a nitrogen recovery decision module, and a self-compensation module.

[0059] The dynamic multi-objective evaluation module calculates an evaluation function value that includes air compressor energy consumption, vacuum pump energy consumption, oxygen purity, and nitrogen recovery rate based on real-time data collected from multiple sensors. This module runs as a software entity on the controller processor, and its function corresponds to step S101 in the aforementioned embodiment. It receives multi-data streams from the sensor network in real time and calculates instantaneous energy consumption using built-in mathematical models (such as power-pressure-flow characteristic curve models for the compressor and vacuum pump). Simultaneously, it reads the oxygen concentration at the product outlet and the gas composition and flow data from the recovery pipeline to calculate oxygen purity and nitrogen recovery rate. Subsequently, the module normalizes these performance indicators of different dimensions and calculates a comprehensive dynamic multi-objective evaluation function value based on preset or dynamically adjusted weighting coefficients. This evaluation value, as a quantified performance score, is provided in real-time to the adsorption control module and the desorption prediction module as a reward or penalty signal for their optimization decisions, thereby guiding the entire system towards global optimization.

[0060] The adsorption control module, during the adsorption phase, generates and outputs dynamic pressure regulation commands for the air compressor based on the evaluation function value and current oxygen demand using a first machine learning model. This module's function corresponds to step S103 in the aforementioned embodiment. Its core is the deployment of a pre-trained offline DRL model. In each control cycle of the adsorption phase, this module uses real-time system status (including tower pressure, temperature, gas composition, and flow data) and current oxygen demand obtained from the sensor network and the upper-level production management system as state inputs to the DRL model. The DRL model outputs an optimal air compressor outlet pressure setpoint based on its policy network. This setpoint is then converted into a specific control signal (e.g., a 4-20mA analog signal or a digital communication command) and sent to the air compressor's variable frequency drive, thereby achieving real-time, adaptive, and energy-saving regulation of the adsorption pressure to dynamically maximize the multi-objective evaluation function value.

[0061] The desorption prediction module, during the desorption stage, uses a second machine learning model based on historical desorption data to predict the optimal vacuum curve and generate start / stop commands for the vacuum pump and its pumping rate control. This module corresponds to step S104 in the aforementioned embodiment. Its core is the deployment of a pre-trained LSTM model. At the beginning of each desorption stage, this module retrieves historical desorption cycle data (such as pressure and gas composition changes over time) stored in memory and inputs it into the LSTM model. Based on its learned time-series dependencies, the model predicts the "optimal vacuum curve" that allows for the regeneration of the molecular sieve with the lowest energy consumption under the current operating conditions. This curve, as a time-varying pressure setpoint sequence, is passed to a low-level PID controller. This PID controller then compares the actual pressure in the adsorption tower with the setpoint in real time, dynamically adjusting the output frequency of the vacuum pump's variable frequency drive to precisely control the pumping rate, ensuring that the actual pressure curve closely follows the predicted optimal trajectory, and promptly stopping the vacuum pump when the regeneration endpoint is reached, thus avoiding energy waste.

[0062] The nitrogen recovery decision module generates valve control commands based on a comparison between the real-time nitrogen concentration and a preset recovery concentration threshold. This module corresponds to step S105 in the aforementioned embodiment. It implements a clear decision logic: during the desorption phase, the module continuously monitors the data from the nitrogen concentration sensor at the desorption outlet. Once the nitrogen concentration exceeds the preset recovery threshold (e.g., 95%), the module immediately sends a "open recovery path" command to the valve controlling the desorption flow path; conversely, when the concentration is below or equal to the threshold, it sends a "switch to emission path" command. This intelligent diversion control based on real-time data ensures that only high-purity nitrogen is sent into the recovery pipeline, thereby guaranteeing the quality of the recovered product and improving resource utilization efficiency.

[0063] The controller also includes a self-compensation module, configured to: receive and store molecular sieve performance degradation indices calculated from data collected by multiple sensors over multiple consecutive working cycles; predict the future performance degradation state of the molecular sieve based on historical performance degradation index data using a pre-trained transfer learning model; and generate and output compensation control parameters to the adsorption control module and desorption prediction module based on the predicted performance degradation state to actively correct their control commands. This module's function corresponds to the preferred technical solution for molecular sieve performance degradation in the aforementioned embodiments. It integrates a data recording and analysis unit and a transfer learning-based prediction model. The data recording unit continuously tracks and calculates key performance indicators (KPIs) characterizing molecular sieve activity, such as "adsorption time at the same oxygen purity" or "energy consumption per unit of oxygen production." When these indicators show a continuous deterioration trend, the transfer learning model uses this locally accumulated degradation data to fine-tune a general molecular sieve aging baseline model, thereby accurately predicting the performance state of the molecular sieve in the system over a future period. Based on the prediction results, the module generates a set of compensation parameters. For example, if a decrease in adsorption capacity is predicted, it may generate a compensation instruction to "increase the target value of desorption vacuum by 5%" or "reduce the nitrogen recovery concentration threshold by 0.5%". These compensation parameters are automatically applied to the settings of the desorption prediction module and the nitrogen recovery decision module, enabling the entire control system to proactively adapt to and compensate for the performance degradation caused by equipment aging, thereby extending the effective service life of the molecular sieve and keeping the system in a high-efficiency operating state for a long time.

[0064] Please see Figure 3 The third aspect of this application provides a PSA oxygen generation and nitrogen recovery coordinated control device, including a memory and a processor connected in series and in communication. The memory is used to store a computer program, and the processor is used to read the computer program and execute the PSA oxygen generation and nitrogen recovery coordinated control method as described in the first aspect of the embodiment. Specifically, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out memory (FIFO), and / or last-in-first-out memory (FILO), etc.; the processor may be not limited to microprocessors of the STM32F105 series, ARM (Advanced RISC Machines), x86 architecture processors, or processors with integrated NPU (neural-network processing units). The working process, working details, and technical effects of the device provided in the third aspect of this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0065] This fourth aspect of the embodiment provides a computer-readable storage medium storing instructions for the PSA oxygen generation and nitrogen recovery coordinated control method of the first aspect of the embodiment. Specifically, the computer-readable storage medium stores instructions that, when executed on a computer, perform the PSA oxygen generation and nitrogen recovery coordinated control method as described in the first aspect. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The working process, details, and technical effects of the computer-readable storage medium provided in this fourth aspect of the embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0066] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the PSA oxygen generation and nitrogen recovery coordinated control method as described in the first aspect of this embodiment. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0067] The embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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.

[0068] 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 part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a repository code merging device to execute the methods of various embodiments or some parts of the embodiments.

[0069] Finally, it should be noted that the above are merely preferred embodiments of the invention and are not intended to limit the scope of protection of the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for coordinated control of PSA oxygen production and nitrogen recovery, applied to a PSA oxygen production and nitrogen recovery coordinated control system comprising at least one adsorption tower, an air compressor connected to the inlet of the adsorption tower, a vacuum pump connected to the outlet of the adsorption tower, and a nitrogen recovery pipeline, wherein the adsorption tower is filled with a molecular sieve for selective adsorption of nitrogen, characterized in that, The method includes: Establish a dynamic multi-objective evaluation function that includes multiple preset performance indicators, including at least the energy consumption of the air compressor, the energy consumption of the vacuum pump, oxygen purity, and nitrogen recovery rate. Multiple sensors arranged in the PSA oxygen production and nitrogen recovery coordinated control system are used to collect real-time data on pressure, temperature, gas composition and flow rate inside the adsorption tower. During the adsorption stage, based on the real-time collected data on pressure, temperature, gas composition and flow rate inside the adsorption tower and the current oxygen demand, a dynamic pressure regulation strategy for the air compressor is generated and executed using a first machine learning model to dynamically optimize the multi-objective evaluation function. During the desorption phase, a second machine learning model is used to predict the optimal vacuum curve for the current desorption process based on historical desorption data. Based on the optimal vacuum curve, a start-stop and pumping rate control strategy for the vacuum pump is generated and executed to dynamically optimize the multi-objective evaluation function. During the desorption stage, based on the comparison between the nitrogen concentration in the real-time collected gas component data and the preset recovery concentration threshold, the desorption gas flow is controlled by a valve to enter the nitrogen recovery pipeline or be directly discharged.

2. The method according to claim 1, characterized in that, The first machine learning model is a deep reinforcement learning model; the step of generating and executing a dynamic pressure regulation strategy for the air compressor includes: The real-time collected data and current oxygen demand are used as the state input of the deep reinforcement learning model, and the optimization result of the multi-objective evaluation function is used as the reward signal. The deep reinforcement learning model outputs an output pressure command to control the air compressor to operate with the lowest power consumption while meeting the oxygen demand.

3. The method according to claim 1 or 2, characterized in that, The second machine learning model is a long short-term memory network model; the step of predicting the optimal vacuum curve for the current desorption process based on historical desorption data includes: The time-series relationship between historical desorption curve data and nitrogen desorption efficiency data is analyzed using the long short-term memory network model to predict the minimum negative pressure intensity and duration required for the molecular sieve to reach the regeneration threshold. The step of generating and executing a start-up and pumping rate control strategy for the vacuum pump based on the optimal vacuum curve includes: determining the start-up timing, stop-up timing, and pumping rate of the vacuum pump based on the optimal vacuum curve.

4. The method according to claim 1, characterized in that, The step of controlling the desorption gas flow through a valve to enter the nitrogen recovery pipeline or to discharge it directly includes: When the nitrogen concentration in the desorption gas flow is detected to be higher than the preset recovery concentration threshold, the valve is controlled to open, and the desorption gas flow is directed to the recovery storage tank connected to the nitrogen recovery pipeline. When the nitrogen concentration in the desorption gas flow is detected to be lower than or equal to the preset recovery concentration threshold, the valve is controlled to close and the desorption gas flow is switched to the direct emission path.

5. The method according to claim 4, characterized in that, The method further includes: The pressure inside the recovery tank is monitored in real time, and the depressurization rate of the buffer tank connected to the adsorption tower is dynamically adjusted based on the pressure feedback inside the recovery tank to maintain the smooth operation of the nitrogen recovery process.

6. The method according to claim 1, characterized in that, The PSA oxygen production and nitrogen recovery coordinated control system is a multi-tower system comprising multiple adsorption towers, and the method further includes: Using a fuzzy logic controller, the switching cycle and nitrogen recovery sequence of multiple adsorption towers are optimized collaboratively based on external economic signals and system operating status; the external economic signals include at least time-of-use electricity price information, and the system operating status includes at least the liquid level or pressure of the oxygen storage tank.

7. The method according to claim 6, characterized in that, The step of using a fuzzy logic controller to collaboratively optimize the switching cycle and nitrogen recovery sequence of multiple adsorption towers based on external economic signals and system operating status includes: When the time-of-use electricity price information is during a low-electricity period, the adsorption time of the adsorption tower currently in the adsorption stage is extended; when the time-of-use electricity price information is during a high-electricity period, the vacuum maintenance time of the adsorption tower currently in the desorption stage is shortened.

8. The method according to claim 1, characterized in that, The method further includes: Continuously monitor and record at least one performance degradation index used to characterize the performance of the molecular sieve during multiple consecutive working cycles; Using a transfer learning model trained on historical performance degradation index data, the performance degradation status of the molecular sieve in one or more future working cycles is predicted. Based on the predicted performance degradation state, compensation control parameters are generated, and the control strategies of the first machine learning model and the second machine learning model are actively corrected using the compensation control parameters.

9. The method according to claim 8, characterized in that, The performance degradation metric is a performance vector containing multiple parameters, which are selected from: Under the condition of producing oxygen of the same purity, the duration of the adsorption stage, the average pressure drop in the adsorption tower, the energy consumption per unit of oxygen production, or any combination thereof; The compensation control parameters include a compensation value for the desorption negative pressure intensity during the desorption stage and an adjustment value for the preset recovery concentration threshold.

10. A PSA oxygen generation and nitrogen recovery coordinated control system, comprising at least one adsorption tower, an air compressor connected to the inlet of the adsorption tower, a vacuum pump connected to the outlet of the adsorption tower, a nitrogen recovery pipeline, valves, and multiple sensors for collecting pressure, temperature, gas composition, and flow rate data within the adsorption tower, characterized in that, The system further includes a controller, which is communicatively connected to the air compressor, the vacuum pump, the valves, and the plurality of sensors. The controller includes: The dynamic multi-objective evaluation module is used to calculate an evaluation function value that includes air compressor energy consumption, vacuum pump energy consumption, oxygen purity and nitrogen recovery rate based on real-time data collected by the multiple sensors. An adsorption control module is used to generate and output dynamic pressure regulation commands for the air compressor based on the evaluation function value and the current oxygen demand during the adsorption stage using a first machine learning model. The desorption prediction module is used to predict the optimal vacuum curve based on historical desorption data using a second machine learning model during the desorption stage, and to generate start / stop and pumping rate control commands for the vacuum pump. The nitrogen recovery decision module is used to generate on / off control commands for the valve based on the comparison results between the real-time nitrogen concentration and the preset recovery concentration threshold. The controller includes a self-compensation module, which is used for: Receive and store molecular sieve performance degradation indicators calculated from data collected by the multiple sensors in multiple consecutive working cycles; Using a pre-trained transfer learning model, the future performance degradation status of molecular sieves is predicted based on historical performance degradation index data. Based on the predicted performance degradation state, compensation control parameters are generated and output to the adsorption control module and the desorption prediction module to actively correct their control commands.

Images