A helium compressor adaptive energy-saving control method for a cryogenic pump system

Abstract

The present application relates to the technical field of compressor control, and specifically discloses a helium compressor adaptive energy-saving control method for a cryopump system, which comprises the following steps: collecting in real time suction pressure P in , exhaust pressure P out , first-stage cold head temperature T1, second-stage cold head temperature T2, vacuum cavity gas load estimation value L gas , compressor motor current I motor and historical operation data set D hist ; constructing a state feature vector and predicting the refrigeration load demand Q pred and optimal pressure difference ΔP opt of the next cycle by using a physical information neural network combined with a long short-term memory network; adopting a model predictive control with constraints to optimize compressor target rotating speed ω ref and bypass valve opening V bypass ; executing control and online fine-tuning the model; and triggering a protection mode in an abnormal situation. The method realizes significant energy saving and stable control through physical information constraints, time sequence prediction and dynamic optimization, and is suitable for various cryopump systems.

Description

Technical Field

[0001] This invention relates to the field of compressor control technology, and in particular to an adaptive energy-saving control method for a helium compressor used in a cryogenic pump system. Background Technology

[0002] Existing cryogenic pump systems are widely used in semiconductor manufacturing, particle accelerators, space simulation devices, and superconducting equipment cooling, with their core cooling power provided by helium compressors. Traditional helium compressor control mainly employs fixed-speed operation or simple PID pressure / temperature feedback control. This type of control method has the following drawbacks:

[0003] 1. The inability to accurately sense and predict dynamic gas load changes in the vacuum chamber causes the compressor to frequently operate at overpower, resulting in high energy consumption;

[0004] 2. The control precision of the cold head temperature is limited, and the temperature of the second-stage cold head is prone to large fluctuations (usually ±2K or more), which affects the pumping speed stability and ultimate vacuum of the cryogenic pump.

[0005] 3. Poor adaptability: When the system is in different operating conditions such as startup, steady-state air extraction, regeneration cycle switching or multiple pumps in parallel, it cannot dynamically balance cooling demand and energy consumption, resulting in low long-term operating efficiency.

[0006] 4. The lack of an effective online learning mechanism leads to a gradual deterioration in control performance as equipment ages, ambient temperature changes, or helium purity fluctuates slightly, requiring frequent manual intervention and parameter readjustment.

[0007] Although some advanced compressors use variable frequency drive or basic fuzzy control, they still rely mainly on single parameter feedback and lack deep prediction and optimization methods that integrate physical mechanisms and data-driven approaches. In low-temperature environments with strong nonlinearity and small sample sizes, the prediction accuracy and robustness are insufficient, resulting in high overall system energy consumption and poor stability, which cannot meet the high-efficiency and reliable operation requirements of high-end vacuum applications. Summary of the Invention

[0008] To address the problems of high energy consumption, large temperature fluctuations, and sluggish system response caused by the use of fixed speed or single PID feedback in existing technologies, which cannot dynamically adapt to varying vacuum loads and cold head heat loads, this invention provides an adaptive energy-saving control method for helium compressors in cryogenic pump systems.

[0009] The technical solution adopted in this invention is:

[0010] An adaptive energy-saving control method for a helium compressor in a cryogenic pump system includes the following steps:

[0011] S1. Real-time acquisition of operating parameters of the helium compressor system, including suction pressure P.in Exhaust pressure P out First-stage cold head temperature T1, second-stage cold head temperature T2, estimated gas load L in the vacuum chamber gas Compressor motor current I motor and historical running dataset D hist ;

[0012] S2. Construct a state feature vector X based on the acquired parameters. t =[P in ,P out ,T1,T2,L gas ,I motor ,ΔP t ], where ΔP t =P out -P in This represents the current pressure difference;

[0013] S3. Based on the state feature vector, use a physical information neural network combined with a long short-term memory network to determine the cooling load demand Q for the next cycle. pred and the optimal pressure difference ΔP opt Conduct joint forecasting;

[0014] S4, according to Q pred and ΔP opt A constrained model predictive control optimizer is used to calculate the target speed ω of the compressor inverter. ref and bypass valve opening V bypass ;

[0015] S5, ω ref and V bypass Apply to the actuator and update the historical running dataset D. hist And periodically perform online fine-tuning;

[0016] S6. When the temperature of the second-stage cold head T2 exceeds the preset threshold T... th When the vacuum level deteriorates, the protection mode is triggered and the system switches to conservative control parameters.

[0017] Preferably, the estimated gas load value L in the vacuum chamber gas Through formula L gas =k1⋅(dP vac / dt)+k2⋅∫(T1-T 1ref )dt+k3⋅I motor Calculate dP, where k1, k2, and k2 are empirical calibration coefficients. vac / dt is the rate of change of vacuum chamber pressure, T 1ref This is the target temperature for the first-stage cold head.

[0018] Preferably, the physical information neural network consists of an input layer, three hidden layers, and an output layer: the input layer receives the normalized state feature vector X. t The dimension is 7, and the preprocessing method is min-max normalization to [0,1]. The first hidden layer is a fully connected layer with 128 neurons, activated by ReLU. The second hidden layer is a long short-term memory layer with 64 units, handling time series dependencies. The third hidden layer is a fully connected layer with 32 neurons, activated by ReLU, and incorporates physical residual connections to embed the thermodynamic equation residuals into the loss. The output layer contains 2 neurons, activated by linear activation, and outputs Q. pred and ΔP opt ;

[0019] The training process of the physical information neural network adopts a hybrid loss function L. total =L data +α⋅L phys , where L data For mean square error, L phys The physical residual loss is α, the balance coefficient is α, the Adam optimizer is used for training, the learning rate is 0.001, the number of training epochs is 200, and the mini-batch size is 32.

[0020] Preferably, dense connections are added between the third hidden layer and the output layer of the physical information neural network, and the output of the second hidden layer is concatenated with the output of the first hidden layer and then input into the third hidden layer.

[0021] Preferably, the joint prediction process satisfies physical constraint Q. pred ≈m·C p ·(T in -T out )+λ·L gas Where m is the helium mass flow rate, and C p For the specific heat capacity at constant pressure, T in T represents the inlet temperature of the cold head. out λ is the outlet temperature of the cold head, and λ is the load factor.

[0022] Preferably, the objective function of the model predictive control optimizer is: Where N is the prediction time domain length, w1, w2, and w3 are adaptive weights, T2(k) is the second-stage cold head temperature predicted in the k-th step, and T 2ref Let ω be the target temperature of the second-stage cold head, Power(ω) be the compressor power model, ∆P(k) be the pressure difference predicted at the k-th step, and ρ be the penalty coefficient.

[0023] Preferably, the adaptive weight adjustment rule in the model predictive control optimizer is w2=w 2base ⋅(1+β⋅(1-L gas / Lmax ), where w 2base =1, β=0.8 is the adjustment factor, L max This is an estimate of the maximum gas load in the vacuum chamber; constraints include ω. min ≤ω ref ≤ω max ΔP min ≤ΔP(k)≤ΔP max And T2(k)≤T th ω ref To optimize the output of the compressor target speed, ω min ω is the minimum permissible speed of the compressor. max ΔP is the maximum allowable speed of the compressor. min For the minimum safe pressure difference, ΔP max This represents the maximum safe pressure difference.

[0024] Preferably, the model predictive control optimizer has prediction time domain N and control time domain M satisfying N=5 and M=2, and uses a quadratic programming solver to solve for the minimum value of J in real time.

[0025] Preferably, the historical running dataset D hist The online fine-tuning uses an experience replay buffer. After every 10 control cycles, a small batch of samples is extracted to perform single-step gradient updates on the physical information neural network, with the learning rate decaying to 0.0001.

[0026] Preferably, the protection mode employs a conservative rotational speed ω. con =0.6⋅ω max At the same time, the auxiliary cooling circuit is activated until T2 returns to normal.

[0027] The beneficial effects of this invention are:

[0028] 1. Significantly reduce energy consumption: By combining physical information neural networks with long short-term memory networks, the cooling load demand Q for the next cycle is predicted. pred and the optimal pressure difference ΔP opt The adaptive weighted model predictive control optimizer dynamically optimizes the rotational speed ω. ref This enables energy supply on demand.

[0029] 2. Improved temperature control accuracy and system stability: Thanks to physical information residual embedding and long short-term memory network time-series modeling, the model can accurately capture dynamic changes in the load and adjust the control quantity in advance. The model predictive control optimizer's multi-step rolling optimization further eliminates overshoot and lag, ensuring rapid stabilization of the cold head temperature, thereby maintaining a stable pumping speed and ultimate vacuum of the cryogenic pump.

[0030] 3. Enhanced environmental adaptability and long-term robustness: Online experience playback and fine-tuning combined with periodic knowledge distillation enable the model to continuously adapt to changes such as equipment aging and environmental drift, solving the problem of gradual deterioration of existing fixed parameter control and reducing the frequency and cost of manual maintenance.

[0031] 4. Improve system safety and availability: Quickly switch to protection mode under abnormal operating conditions to effectively prevent the cold head from freezing or overheating, thus improving system availability. At the same time, physical constraints and multi-constraint model predictive control optimizers ensure that all control actions comply with thermodynamic safety boundaries, avoiding infeasible decisions that may occur with purely data-driven methods. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the adaptive energy-saving control method for a helium compressor used in a cryogenic pump system, as described in an embodiment of the present invention. Detailed Implementation

[0033] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] Example

[0035] As shown in the figure, an adaptive energy-saving control method for a helium compressor in a cryogenic pump system is characterized by the following steps:

[0036] S1. Real-time acquisition of operating parameters of the helium compressor system, including suction pressure P. in Exhaust pressure P out First-stage cold head temperature T1, second-stage cold head temperature T2, estimated gas load L in the vacuum chamber gas Compressor motor current I motor and historical running dataset D hist ;

[0037] S2. Construct a state feature vector X based on the acquired parameters. t =[P in ,P out ,T1,T2,L gas ,I motor ,ΔP t ], where ΔP t =P out -P in This represents the current pressure difference;

[0038] S3. Based on the state feature vector, use a physical information neural network combined with a long short-term memory network to determine the cooling load demand Q for the next cycle (Δt=30 seconds). pred and the optimal pressure difference ΔP opt Conduct joint forecasting;

[0039] S4, according to Q pred and ΔP opt A constrained model predictive control optimizer is used to calculate the target speed ω of the compressor inverter. ref and bypass valve opening V bypass ;

[0040] S5, ω ref and V bypass Apply to the actuator and update the historical running dataset D. hist And periodically perform online fine-tuning;

[0041] S6. When the temperature of the second-stage cold head T2 exceeds the preset threshold T... th When the vacuum level deteriorates, the protection mode is triggered and the system switches to conservative control parameters.

[0042] Existing technologies often rely on current-moment feedback, resulting in significant lag. In contrast, this technique, steps S1-S2, collect multi-dimensional parameters and construct feature vectors. Step S3 utilizes a physical information neural network combined with a long short-term memory network for 30-second look-ahead joint prediction, enabling the early capture of the gas load's L... gas Dynamic trends. Long Short-Term Memory (LSTM) networks on historical running datasets D hist The memory capacity of Q is combined with the physical constraints of the physical information neural network (such as energy conservation and expansion equations) to enable Q to... pred and ΔP opt This approach closely aligns with real-world data and thermodynamic principles, avoiding the potential for physically infeasible predictions that can occur with purely data-driven models. Subsequently, the S4 model predictive control optimizer performs rolling optimization based on the prediction results, adjusting ω in advance. ref This effectively suppresses the temperature fluctuation of the second-stage cold head to within ±0.4K, which is far superior to the fluctuation of more than ±2K in traditional methods, ensuring stable pumping speed and reliable ultimate vacuum of the cryogenic pump.

[0043] The model predictive control optimizer in S4 comprehensively considers temperature deviation, power consumption, and pressure tracking terms, forming a closed loop through continuous execution and data updates in S5. Under low load, the optimizer tends to reduce ω. ref and appropriately enable V bypass This reduces unnecessary compression work; under high load, it prioritizes cooling power. This on-demand energy supply mechanism avoids the waste of continuously high-power operation found in traditional fixed-speed or simple variable-frequency control, thus reducing average power consumption. Simultaneously, S1 collects I... motor and L gas Provide real-time feedback for S3-S4 to create positive incentives and further amplify the energy-saving effect.

[0044] S6 is the final step. When T2 exceeds the limit or the vacuum level deteriorates, the conservative control parameters are switched immediately, and the model is continuously optimized in conjunction with the online update mechanism of S5.

[0045] Multi-parameter fusion and physical constraint prediction make the system more sensitive to anomalies. The multi-constraint conditions of the model predictive control optimizer ensure that the control actions under normal operating conditions are always within the safety boundary, while the rapid intervention of the protection mode prevents cascading failures such as cold head damage or helium leakage, significantly reducing the frequency of downtime maintenance and operational risks.

[0046] S5 for D hist Real-time updates provide a data foundation for online fine-tuning in subsequent cycles, enabling the entire S1-S6 closed loop to have self-improvement capabilities. As the running time increases, the model can learn new patterns such as equipment aging and environmental changes, and the prediction and optimization accuracy gradually improves, avoiding the problem of parameters deteriorating with use in existing technologies. After long-term operation, the overall performance degradation is minimal, exhibiting good engineering maintainability and economy.

[0047] In summary, steps S1-S6, through systematic innovation, deeply integrate advanced artificial intelligence technology with the physical mechanism of cryogenic refrigeration, forming an efficient, intelligent, and safe adaptive control method. This solves the pain points of existing technologies and provides a strong guarantee for the efficient and reliable operation of cryogenic pump systems.

[0048] Because direct measurement of gas load is difficult and time-sensitive, the estimated gas load L in the vacuum chamber is therefore... gas Through formula L gas =k1⋅(dP vac / dt)+k2⋅∫(T1-T 1ref )dt+k3⋅I motor Calculate dP, where k1, k2, and k2 are empirical calibration coefficients. vac / dt is the rate of change of vacuum chamber pressure, T 1ref This is the target temperature for the first-stage cold head.

[0049] By integrating pressure dynamics, temperature deviation, and motor current to indirectly estimate the load, load changes can be detected in advance, and the speed can be adjusted accordingly to avoid temperature overshoot. The pressure change rate reflects the instantaneous gas influx, the temperature integral term captures the accumulated thermal load, and the motor current reflects mechanical power consumption. The linear combination of these three factors, after calibration, is highly correlated with the actual adsorption capacity, thereby improving foresight and reducing energy waste caused by overcompression.

[0050] The physical information neural network consists of an input layer, three hidden layers, and an output layer: the input layer receives the normalized state feature vector X. tThe dimension is 7, and the preprocessing method is min-max normalization to [0,1]. The first hidden layer is a fully connected layer with 128 neurons, activated by ReLU. The second hidden layer is a long short-term memory layer with 64 units, handling time series dependencies. The third hidden layer is a fully connected layer with 32 neurons, activated by ReLU, and incorporates physical residual connections to embed the thermodynamic equation residuals into the loss. The output layer contains 2 neurons, activated by linear activation, and outputs Q. pred and ΔP opt ;

[0051] Because pure data-driven models are prone to overfitting or violating physical laws in low-temperature systems with small samples and strong nonlinear scenarios, the training process of the physical information neural network adopts a hybrid loss function L. total =L data +α⋅L phys , where L data For mean square error, L phys For physical residual loss (based on the helium isentropic expansion equation T) 2out =T 2in ·(P ratio ) (γ-1) / γ α is the balancing coefficient (initial value is 0.3, adaptively decays during training), training uses the Adam optimizer, learning rate is 0.001, training epochs are 200, and mini-batch size is 32.

[0052] Traditional neural networks may predict Q that violates the law of energy conservation. pred This led to temperature runaway after execution, and L phys By penalizing such deviations, the hidden layer is forced to learn the true physical mapping, thereby quickly stabilizing T2 during load abrupt changes, reducing unnecessary high-power operation of the compressor, resulting in more significant overall energy savings and higher system safety.

[0053] To avoid feature loss due to gradient vanishing during deep network training, dense connections are added between the third hidden layer and the output layer of the physical information neural network. The output of the second hidden layer is concatenated with the output of the first hidden layer and then input into the third hidden layer.

[0054] The splicing mechanism preserves low-level physical features and high-level temporal features, alleviates gradient problems, enables physical information neural networks to converge faster with limited samples, and ultimately achieves more stable and energy-efficient control, improves the model's ability to express complex nonlinear mappings, and reduces joint prediction errors.

[0055] Purely data-driven models may output Q values ​​that do not conform to energy conservation under out-of-sample conditions. predFor example, excessively high or low cooling demand forecasts can cause the model predictive control optimizer to issue unreasonable speed commands, ultimately leading to temperature runaway or a surge in energy consumption. Therefore, the joint prediction process satisfies the physical constraint Q. pred ≈m·C p ·(T in -T out )+λ·L gas Where m is the helium mass flow rate, and C p For the specific heat capacity at constant pressure, T in T represents the inlet temperature of the cold head. out λ is the outlet temperature of the cold head, and λ is the load factor.

[0056] The thermodynamic equation (based on the specific heat capacity of helium at constant pressure C) p The temperature difference convective heat transfer term, and λ·L reflecting the effect of gas load. gas This term, as part of the network loss function or output constraint, forces the physical information neural network to always learn a mapping within the physically feasible region. Even with insufficient training data or sudden load changes, Q... pred It will not deviate from the actual physical process, thus providing reliable forward-looking information for subsequent S4 MPC optimization and reducing control deviation from the source.

[0057] Physical constraints serve as prior knowledge, reducing reliance on large amounts of labeled data and enabling the model to learn essential thermodynamic relationships even with limited samples. Simultaneously, the temporal features captured by the Long Short-Term Memory (LSTM) network are mutually regularized with the physical constraints, preventing gradient explosion or vanishing problems. The model's prediction errors for complex operating conditions such as startup, regeneration switching, and multiple pump parallel operation are significantly reduced, ensuring ΔP... opt The accuracy of this control ultimately ensures that the entire control closed loop remains stable during long-term operation.

[0058] Through explicit physical constraints, engineers can directly analyze m and C. p Parameters such as λ affect Q pred This contribution allows for rapid problem identification (such as whether the loading factor λ needs to be adjusted based on the actual gas composition) and targeted model optimization. It accelerates system debugging and iteration, providing a reliable foundation for subsequent online fine-tuning and knowledge distillation.

[0059] Physical constraints reduce the uncertainty space of prediction, enabling the objective function of the model predictive control optimizer to be more aggressively optimized towards energy saving, while the safety boundary is clearer. Ultimately, this achieves a significant reduction in energy consumption and a decrease in the frequency of abnormal triggering while ensuring the stability of the second-stage cold head temperature.

[0060] To achieve dynamic balancing of temperature stability and energy consumption based on real-time load, avoiding energy waste or unstable control performance, the objective function of the model predictive control optimizer is: Where N is the prediction time domain length, and w1, w2, and w3 are adaptive weights (based on L). gas (Dynamically adjusted), T2(k) is the predicted second-stage cold head temperature in the k-th step, T 2ref Let ω be the target temperature of the second-stage cold head, Power(ω) be the compressor power model, ∆P(k) be the pressure difference predicted at the k-th step, and ρ be the penalty coefficient.

[0061] The objective function of the model predictive control optimizer dynamically balances multiple objectives through adaptive weights, and combines look-ahead rolling optimization and constraint penalty mechanisms to achieve on-demand intelligent control of the helium compressor. Under low load, the weight w2 automatically increases, prompting the optimizer to prioritize reducing speed to decrease power consumption; under high load, it strengthens temperature and pressure tracking terms to ensure cooling performance. This mechanism fully utilizes the predictive information from S3, avoiding the short-sightedness of single-step greedy decision-making. Simultaneously, the penalty term forces the optimized solution to satisfy the safety boundary, thereby significantly reducing average energy consumption, suppressing temperature fluctuations, and improving the system's robustness and safety under varying operating conditions. This method makes the control logic transparent, facilitating engineering debugging and long-term maintenance, and achieving a balance between energy saving, stability, and reliability.

[0062] To address the issue that fixed weights cannot balance energy-saving priorities under different operating conditions, the adaptive weight adjustment rule in the model predictive control optimizer is w2=w 2base ⋅(1+β⋅(1-L gas / L max ), where w 2base =1, β=0.8 is the adjustment factor, L max The maximum gas load estimate for the vacuum chamber is determined based on the design capacity, historical operating data, or engineering calibration reference value, thereby enabling adaptive weight adjustment to normalize the current load; constraints include ω. min ≤ω ref ≤ω max ΔP min ≤ΔP(k)≤ΔP max And T2(k)≤T th ω ref To optimize the output of the compressor target speed, ω min ω is the minimum permissible speed of the compressor. max ΔP is the maximum allowable speed of the compressor. min For the minimum safe pressure difference, ΔP max This represents the maximum safe pressure difference.

[0063] Load L gasWhen the load is low, w2 is automatically increased, prompting the optimizer to select a lower speed. By dynamically tilting the objective function through the load normalization factor, the waste of high power that traditional model predictive control optimizers maintain under light load is avoided. Combined with constraints, the minimum cooling capacity is guaranteed, realizing on-demand energy supply, which reduces electricity and thermal management costs, and extends the service life of the frequency converter and motor.

[0064] Since multi-step optimization has high computational complexity and is difficult to deploy in real time, the model predictive control optimizer ensures that the prediction time domain N and control time domain M satisfy N=5 and M=2, and uses a quadratic programming solver to solve for the minimum value of J in real time.

[0065] With N=5, sufficient forward coverage is achieved; with M=2, recent actions are focused on; and quadratic programming ensures global optimality, avoiding suboptimal solutions from greedy algorithms. This allows for long-term energy savings while maintaining real-time performance.

[0066] To address the issue of offline-trained models struggling to adapt to long-term drift, the historical running dataset D is described. hist The online fine-tuning uses an experience replay buffer (capacity of 1000 samples). After every 10 control cycles, a small batch of samples is extracted to perform single-step gradient updates on the physical information neural network, with the learning rate decaying to 0.0001.

[0067] The buffer avoids catastrophic forgetting caused by time-related samples, and the computational cost of single-step updates is small, making it suitable for real-time execution by embedded controllers. This allows the control strategy to be continuously optimized as the device state evolves, ensuring long-term stable energy saving and reducing manual intervention and maintenance costs.

[0068] Under abnormal operating conditions, the system is prone to loss of control and equipment damage. Therefore, the protection mode adopts a conservative rotational speed ω. con =0.6⋅ω max At the same time, the auxiliary cooling circuit is activated until T2 returns to normal.

[0069] The 0.6 times maximum speed provides sufficient cooling while significantly reducing power consumption and vibration. The auxiliary circuit accelerates thermal equilibrium, avoiding production interruptions caused by complete shutdowns in traditional methods, thus achieving a balance between safety and continuous operation.

[0070] In addition, due to performance degradation caused by model drift during long-term operation, the method also includes periodically (every 24 hours) globally fine-tuning the weights of the entire physical information neural network and the model predictive control optimizer, using D... hist The most recent 5000 data points are used to transfer the soft labels of the teacher model (high-precision offline physical information neural network) to the online model through knowledge distillation.

[0071] The distillation mechanism enables the online lightweight model to inherit the rich knowledge of the offline model, while adapting to changes in real-time data distribution, ensuring that the control strategy remains in the optimal state, resulting in considerable long-term economic benefits and improved equipment reliability.

[0072] All parameter acquisition, prediction, optimization, and execution steps are performed periodically in a closed-loop manner by an embedded industrial controller, with a cycle time of 5 seconds.

[0073] Short-cycle, high-frequency feedback corrects deviations in a timely manner. Combined with the aforementioned prediction and optimization modules, it forms a closed-loop positive feedback, achieving a balance between energy saving, stability, and safety, and possessing outstanding industrial practical value.

[0074] The following is an experimental comparison between the adaptive energy-saving control method of the present invention and the traditional PID control method:

[0075] I. Experimental Platform:

[0076] A standard cryogenic pump system (including GM or pulse tube cold head, helium compressor, vacuum chamber, temperature / pressure sensor), with compressor equipped with frequency converter driver and switchable control modes.

[0077] II. Comparison Methods:

[0078] Traditional method: Classical PID control (pressure / temperature dual closed loop, parameters are fixed after Ziegler-Nichols tuning).

[0079] The method of this invention consists of complete steps S1-S6.

[0080] III. Environmental conditions: constant temperature laboratory (25±2℃), helium purity ≥99.999%, initial vacuum degree 10⁻³Pa.

[0081] IV. Test conditions (each group was repeated 3 times, and the average value was taken):

[0082] 1. Steady-state low load (gas load L) gas Low, simulating normal air extraction.

[0083] 2. Steady-state high load (sudden introduction of gas pulse, simulating process load).

[0084] 3. Variable operating condition cycle (low-high-regeneration-low, lasting 4 hours).

[0085] 4. Abnormal operating conditions (simulated sensor noise or cold head temperature disturbance).

[0086] V. Experimental Procedure:

[0087] 1. Pre-cool the system to a stable state (the temperature of the second-stage cold head reaches the standard).

[0088] 2. Load the traditional PID and the control algorithm of this invention respectively (on the same hardware, switched by software).

[0089] 3. Operate under the above conditions and record all data throughout the process (sampling frequency 1Hz).

[0090] 4. Repeat the experiment to eliminate random errors.

[0091] 5. Data processing and statistical analysis (using t-test to verify significant differences).

[0092] VI. Evaluation Indicators:

[0093] 1. Energy consumption indicators: average compressor power (kW), total energy consumption in 4 hours (kWh).

[0094] 2. Temperature stability: Temperature fluctuation range (±K) and standard deviation of the second-stage cold head.

[0095] 3. Control accuracy: Temperature tracking error (compared to T) 2ref The root mean square error).

[0096] 4. Response performance: Temperature recovery time (seconds) after a sudden load change.

[0097] 5. Robustness: System availability percentage and maximum temperature overshoot under abnormal operating conditions.

[0098] 6. Other: Helium consumption, compressor speed fluctuations (qualitative).

[0099] VII. Experimental Results:

[0100] The experimental results are shown in Table 1 below:

[0101] Operating conditions method Average power Temperature fluctuations Recovery time Total energy consumption in 4 hours steady-state low load Traditional PID ～1.8 ±2.1 - ～7.2 steady-state low load This invention ～1.1 ±0.4 - ～4.4 High load mutation Traditional PID ～2.5 ±2.8 180 - High load mutation This invention ～2.0 ±0.6 45 -

[0102] Table 1. Comparison of experimental results between adaptive energy-saving control method and traditional PID control method

[0103] As shown in Table 1, the method of the present invention reduces the average power by 25%-40% and the temperature fluctuation by more than 70% under various operating conditions, improves the response speed by more than 3 times, and exhibits better robustness under abnormal operating conditions, proving the effectiveness of innovations such as physical constraint prediction and adaptive MPC.

[0104] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An adaptive energy-saving control method for a helium compressor in a cryogenic pump system, characterized in that, Includes the following steps: S1. Real-time acquisition of operating parameters of the helium compressor system, including suction pressure P. in Exhaust pressure P out First-stage cold head temperature T1, second-stage cold head temperature T2, estimated gas load L in the vacuum chamber gas Compressor motor current I motor and historical running dataset D hist ; S2. Construct a state feature vector X based on the acquired parameters. t =[P in ,P out ,T1,T2,L gas ,I motor ,ΔP t ], where ΔP t =P out -P in This represents the current pressure difference; S3. Based on the state feature vector, use a physical information neural network combined with a long short-term memory network to determine the cooling load demand Q for the next cycle. pred and the optimal pressure difference ΔP opt Conduct joint forecasting; S4, according to Q pred and ΔP opt A constrained model predictive control optimizer is used to calculate the target speed ω of the compressor inverter. ref and bypass valve opening V bypass ; S5, ω ref and V bypass Apply to the actuator and update the historical running dataset D. hist And periodically perform online fine-tuning; S6. When the temperature of the second-stage cold head T2 exceeds the preset threshold T... th When the vacuum level deteriorates, the protection mode is triggered and the system switches to conservative control parameters.

2. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The estimated gas load value L in the vacuum chamber gas Through formula L gas =k1⋅(dP vac / dt)+k2⋅∫(T1-T 1ref )dt+k3⋅I motor Calculate dP, where k1, k2, and k2 are empirical calibration coefficients. vac / dt is the rate of change of vacuum chamber pressure, T 1ref This is the target temperature for the first-stage cold head.

3. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The physical information neural network consists of an input layer, three hidden layers, and an output layer: the input layer receives the normalized state feature vector X. t The dimension is 7, and the preprocessing method is min-max normalization to [0,1]. The first hidden layer is a fully connected layer with 128 neurons, activated by ReLU. The second hidden layer is a long short-term memory layer with 64 units, which handles time series dependencies. The third hidden layer is a fully connected layer with 32 neurons, activated by ReLU, and physical residual connections are added to embed the thermodynamic equation residuals into the loss. The output layer contains two neurons, using linear activation, and outputs Q. pred and ΔP opt ; The training process of the physical information neural network adopts a hybrid loss function L. total =L data +α⋅L phys , where L data For mean square error, L phys The physical residual loss is α, the balance coefficient is α, the Adam optimizer is used for training, the learning rate is 0.001, the number of training epochs is 200, and the mini-batch size is 32.

4. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 3, characterized in that, Dense connections are added between the third hidden layer and the output layer of the physical information neural network, and the output of the second hidden layer is concatenated with the output of the first hidden layer and then input into the third hidden layer.

5. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The joint prediction process satisfies the physical constraint Q. pred ≈m·C p ·(T in -T out )+λ·L gas Where m is the helium mass flow rate, and C p For the specific heat capacity at constant pressure, T in T represents the inlet temperature of the cold head. out λ is the outlet temperature of the cold head, and λ is the load factor.

6. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The objective function of the model predictive control optimizer is: Where N is the prediction time domain length, w1, w2, and w3 are adaptive weights, T2(k) is the second-stage cold head temperature predicted in the k-th step, and T 2ref Let ω be the target temperature of the second-stage cold head, Power(ω) be the compressor power model, ∆P(k) be the pressure difference predicted at the k-th step, and ρ be the penalty coefficient.

7. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 6, characterized in that, The adaptive weight adjustment rule in the model predictive control optimizer is w2=w 2base ⋅(1+β⋅(1-L gas / L max ), where w 2base =1, β=0.8 is the adjustment factor, L max This is an estimate of the maximum gas load in the vacuum chamber; constraints include ω. min ≤ω ref ≤ω max ΔP min ≤ΔP(k)≤ΔP max And T2(k)≤T th ω ref To optimize the output of the compressor target speed, ω min ω is the minimum permissible speed of the compressor. max ΔP is the maximum allowable speed of the compressor. min For the minimum safe pressure difference, ΔP max This represents the maximum safe pressure difference.

8. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 7, characterized in that, The model predictive control optimizer predicts the time domain N and controls the time domain M, satisfying N=5 and M=2, and uses a quadratic programming solver to solve for the minimum value of J in real time.

9. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The historical running dataset D hist The online fine-tuning uses an experience replay buffer. After every 10 control cycles, a small batch of samples is extracted to perform single-step gradient updates on the physical information neural network, with the learning rate decaying to 0.0001.

10. The adaptive energy-saving control method for a helium compressor in a cryogenic pump system according to claim 1, characterized in that, The protection mode employs a conservative rotational speed ω con =0.6⋅ω max At the same time, the auxiliary cooling circuit is activated until T2 returns to normal.

Images