A method and system for optimizing superheat based on machine learning to improve refrigeration efficiency

Abstract

The present application relates to the technical field of intelligent control of refrigeration systems, in particular to a method and system for optimizing superheat based on machine learning; containing state mapping, time lag perception, risk assessment and dual-mode control module; the system generates a saturation distance vector in the pressure-enthalpy coordinate system by receiving real-time data and determines the thermal response lag time window using gradient-based adaptive adjustment logic; the core is to input historical data into a deep neural network, constrained by asymmetric penalties, and predict the best energy efficiency opening and liquid strike risk probability in parallel; according to the comparison of risk value and safety threshold, the system dynamically outputs the best opening or triggers the safety fallback mechanism; the present application introduces physical context constraints, effectively solves the problem of poor model generalization ability, and improves the system adaptability and safety under complex variable working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for refrigeration systems, specifically a method and system for improving refrigeration efficiency by optimizing superheat based on machine learning. Background Technology

[0002] With the increasing complexity of refrigeration system applications, the dual demands for energy efficiency optimization and operational safety have significantly increased. This complexity poses a severe challenge to traditional control strategies, especially in handling variable operating conditions and multivariable coupling.

[0003] Currently, PID control strategies or conventional artificial intelligence algorithms are commonly used for regulation. However, traditional PID control is difficult to effectively cope with nonlinear disturbances under varying operating conditions, while conventional artificial intelligence control lacks a clear understanding of thermodynamic safety boundaries. Existing technologies mainly face the following technical bottlenecks: relying solely on raw sensor values ​​lacks physical context, resulting in poor generalization ability of the model under extreme conditions and difficulty in learning implicit thermodynamic laws; thermal inertia nonlinear drift caused by changes in refrigerant flow rate causes misalignment between control signals and temperature responses on the time axis, making it easy to generate false data correlations; the error symmetry assumption in traditional algorithms fails to distinguish between the risk of liquid slugging and energy efficiency loss in terms of weight, and the system is difficult to adaptively compensate for physical parameter drift caused by equipment aging.

[0004] Therefore, how to solve the problems of dynamic time delay and parameter drift on the basis of constructing a clear physical safety boundary, so as to achieve simultaneous improvement of cooling efficiency and system safety, has become an urgent problem to be solved in this field. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for improving cooling efficiency by optimizing superheat based on machine learning. Specifically, the technical solution of this invention is as follows:

[0006] A method for optimizing superheat based on machine learning to improve cooling efficiency includes:

[0007] Step 1, Thermodynamic State Space Mapping: Receive real-time suction pressure data and real-time suction temperature data of the refrigeration system, retrieve the refrigerant property parameter library, determine the position of the gas-liquid saturation line in the pressure-enthalpy logical coordinate system, calculate the thermodynamic distance between the current suction state point and the gas-liquid saturation line, and generate a saturation distance vector.

[0008] Step 2, Dynamic Time Lag Perception: Monitor the rate of change of the saturation distance vector, calculate the thermal response lag time window using dynamic attention logic, and extract the historical state data sequence based on the thermal response lag time window;

[0009] Step 3, Asymmetric Risk Assessment: Input the historical state data sequence into the time-series prediction deep neural network model, use the asymmetric penalty loss function to constrain the time-series prediction deep neural network model, and output the best energy efficiency opening prediction value and liquid hammer risk probability value in parallel.

[0010] Step 4, Dual-modal constraint execution: Compare the liquid slugging risk probability value with the safety threshold. If the liquid slugging risk probability value is lower than the safety threshold, output the optimal energy efficiency opening prediction value to the electronic expansion valve. If the liquid slugging risk probability value is higher than the safety threshold, trigger the safety backoff mechanism to correct the opening of the electronic expansion valve.

[0011] Step 5, Closed-loop feedback: Collect the actual superheat change after the execution command; transmit the actual superheat change back to Step 2 to correct the calculation parameters of the thermal response hysteresis window.

[0012] Preferably, the saturation distance vector is a physical state descriptor; the saturation distance vector represents the enthalpy or entropy margin of the current refrigerant state from the point of liquefaction; the process of generating the saturation distance vector in step 1 includes: generating the saturation distance vector and passing it as a unique physical feature to step 2 to solve the problem of lack of physical context in single sensor data.

[0013] Preferably, the process of calculating the thermal response lag time window in step 2 includes: monitoring the change gradient of superheat represented by the saturation distance vector; if the rate of change accelerates, determining that the phase transition of the refrigeration system is drastic and the lag time is shortened; narrowing the thermal response lag time window; if the rate of change slows down, widening the thermal response lag time window; step 2 uses the thermal response lag time window to determine the range of historical data to be read in subsequent steps, so as to ensure that the historical state data sequence is strictly aligned in terms of time causality.

[0014] Preferably, the time-series prediction deep neural network model employs a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU); the asymmetric penalty loss function assigns uneven weights to the prediction error; if the state predicted by the time-series prediction deep neural network model shows that the superheat is trending towards falling below zero and entering the liquid phase region, the asymmetric penalty loss function assigns a first penalty weight; if the prediction result shows that the superheat is too high, the asymmetric penalty loss function assigns a second penalty weight; the value of the first penalty weight is greater than the value of the second penalty weight.

[0015] Preferably, the signal output in step 3 is defined as follows:

[0016] The optimal energy efficiency opening prediction value is the target opening of the electronic expansion valve that can achieve the highest energy efficiency ratio, inferred based on the characteristics of the current operating conditions.

[0017] The liquid slugging risk probability value is a confidence score that quantifies the degree of danger of liquid slugging occurring when the current state approaches the gas-liquid saturation line; the confidence score ranges from 0 to 1.

[0018] Preferably, the closed-loop feedback process in step 5 includes: using the actual response as a feedback signal; the actual response being the change in superheat actually measured by the sensor after the command is executed; comparing the time difference between the predicted response corresponding to the optimal energy efficiency opening value and the actual response; and correcting the time window calculation parameters within the dynamic time delay sensing step online based on the time difference to adapt to the drift of physical parameters caused by equipment aging.

[0019] A system for optimizing superheat based on machine learning to improve cooling efficiency includes: a thermodynamic state-space mapping unit, a dynamic time delay sensing unit, an asymmetric risk assessment unit, and a bimodal constraint execution unit;

[0020] The thermodynamic state space mapping unit is configured to receive sensor data, retrieve refrigerant property parameter library, reconstruct physical characteristics, and output saturation distance vector to the dynamic time delay sensing unit.

[0021] The dynamic time delay sensing unit is configured to monitor the rate of change of the saturation distance vector, calculate the thermal response hysteresis time window, and transmit the aligned historical state data sequence to the asymmetric risk assessment unit.

[0022] The asymmetric risk assessment unit has a built-in time-series prediction deep neural network model and is configured to output the best energy efficiency opening prediction value and liquid hammer risk probability value to the dual-modal constraint execution unit based on the asymmetric penalty loss function.

[0023] The dual-modal constraint execution unit is configured to execute valve actions based on the comparison result between the liquid hammer risk probability value and the safety threshold, and to feed back the actual response to the dynamic time delay sensing unit to form a closed-loop control.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention transforms the original intake pressure and temperature data into a saturation distance vector that represents the thermodynamic distance from the gas-liquid saturation line through thermodynamic state-space mapping, giving the control data a clear physical context. This mechanism effectively solves the problem in the prior art that the model has poor generalization ability under extreme conditions due to the lack of physical constraints caused by simply relying on the original sensor values. It ensures that the control strategy is established on a clear thermodynamic safety boundary and improves the system's adaptability to complex and variable operating conditions.

[0026] 2. This invention introduces a dynamic time-delay sensing mechanism. By monitoring the rate of change of the saturation distance vector, an adaptive algorithm is used to dynamically adjust the thermal response lag time window, thereby accurately capturing historical state data sequences that are strictly aligned in terms of physical causality. This technique effectively overcomes the problem of thermal inertia nonlinear drift caused by changes in refrigerant flow rate, eliminates the misalignment of control signals and temperature responses on the time axis, avoids false data correlation, and thus prevents the control oscillation phenomenon commonly seen in variable time-delay systems.

[0027] 3. This invention adopts an asymmetric risk assessment strategy and uses an asymmetric penalty loss function to constrain the time-series prediction deep neural network model, breaking the assumption of error symmetry in traditional algorithms. By assigning a weight to liquid slugging risk that is much higher than energy efficiency loss, the model has an inherent risk aversion characteristic during inference. While maximizing the cooling energy efficiency ratio, it maintains extremely high sensitivity to dangerous states approaching the liquid phase region, thereby achieving effective protection of compressor hardware at the algorithm level and solving the contradiction between energy efficiency optimization and operational safety.

[0028] 4. This invention constructs a dual-modal constraint execution and closed-loop feedback mechanism, which can not only dynamically switch between optimal energy efficiency control and safety backoff mechanism based on risk probability value, but also correct the time window calculation parameters online by comparing the time difference between the predicted response and the actual response. This closed-loop mechanism gives the system the ability to adaptively compensate for the drift of physical parameters such as equipment aging and heat exchanger fouling, and can ensure the stability of control accuracy throughout the entire life cycle of the equipment without manual intervention and recalibration. Attached Figure Description

[0029] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0030] Figure 1 This is a flowchart of the method of the present invention;

[0031] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Example 1:

[0034] Please see Figure 1 A method for optimizing superheat based on machine learning to improve cooling efficiency includes:

[0035] Step 1, Thermodynamic State Space Mapping: Receive real-time suction pressure and temperature data of the refrigeration system, retrieve refrigerant property parameter library, determine the position of the gas-liquid saturation line in the pressure-enthalpy logical coordinate system, calculate the thermodynamic distance between the current suction state point and the gas-liquid saturation line, and generate a saturation distance vector.

[0036] Step 2, Dynamic Time Lag Sensing: Monitor the rate of change of the saturation distance vector, calculate the thermal response lag time window using a gradient-based adaptive sliding window algorithm, and extract historical state data sequences based on the thermal response lag time window;

[0037] Step 3, Asymmetric Risk Assessment: Input the historical state data sequence into the time series prediction deep neural network model, use the asymmetric penalty loss function to constrain the time series prediction deep neural network model, and output the best energy efficiency opening prediction value and liquid hammer risk probability value in parallel.

[0038] Step 4, Dual-modal constraint execution: Compare the liquid slugging risk probability value with the safety threshold. If the liquid slugging risk probability value is lower than the safety threshold, output the optimal energy efficiency opening prediction value to the electronic expansion valve. If the liquid slugging risk probability value is higher than the safety threshold, trigger the safety backoff mechanism to correct the opening of the electronic expansion valve.

[0039] Step 5, Closed-loop feedback: Collect the actual superheat change after the command is executed; send the actual superheat change back to Step 2 to correct the calculation parameters of the thermal response hysteresis window.

[0040] In the technical solution detailed in this embodiment, a method for improving cooling efficiency by optimizing superheat based on machine learning solves the dual technical problems of traditional PID control being unable to cope with changing operating conditions and conventional AI control lacking safety boundary awareness by constructing a control logic that integrates physical constraints and data-driven approaches.

[0041] In the thermodynamic state-space mapping stage of this method, the system collects suction pressure and suction temperature data in real time through a high-frequency sensor interface. Unlike existing technologies that only standardize the raw values, this embodiment introduces a thermodynamic mechanism preprocessing mechanism: it calls a pre-set refrigerant property parameter library and constructs a pressure-enthalpy logical coordinate system in the calculation logic. Within this coordinate system, the critical enthalpy value of the gas-liquid saturation line is accurately located based on the real-time pressure value using an interpolation algorithm or a lookup table method. Then, the difference between the current suction state point and this critical position is calculated to generate a saturation distance vector. At the same time, the current electronic expansion valve opening data and compressor operating frequency data are collected and transmitted synchronously with the saturation distance vector to subsequent steps. This step establishes the physical reference of the control system, ensuring that subsequent data processing is based on a clear thermodynamic safety boundary.

[0042] Following the extraction of the aforementioned physical features, the dynamic time-delay sensing step addresses the thermal inertia nonlinear drift problem caused by changes in refrigerant flow rate. The monitoring logic tracks the first derivative of the saturation distance vector in real time, i.e., the rate of change, and uses a gradient-based adaptive horizon adjustment algorithm to deduce the effective response period under the current operating condition, calculating the thermal response lag time window. This time window is not a fixed value, but rather serves as the length parameter of a sliding window, used to extract a historical state data sequence that is strictly aligned in terms of physical causality from the time-series database. This processing eliminates spurious correlations caused by misalignment between the control signal and the temperature response signal on the time axis, providing accurate input samples for the neural network.

[0043] Entering the asymmetric risk assessment stage, the system inputs the time-aligned data sequence into the time-series prediction deep neural network model. The training and inference process of this model is strongly constrained by the asymmetric penalty loss function. The core of the design of this loss function is to break the symmetry assumption of the error, that is, to give the liquid slamming risk a much higher weight than the energy efficiency loss at the mathematical level. After deep inference of the model, the system outputs in parallel the optimal energy efficiency opening prediction value aimed at maximizing COP, as well as the liquid slamming risk probability value that quantifies the possibility of the current operating condition approaching the liquid phase region.

[0044] In the decision-making and execution phase, the dual-modal constraint execution steps dynamically switch control strategies based on the comparison between the risk probability value and the preset safety threshold. When the system determines that it is in a low-risk state, the execution unit directly adopts the optimal energy efficiency opening prediction value and drives the electronic expansion valve to operate close to the critical superheat. Once the risk probability value exceeds the safety threshold, the system immediately triggers the safety backoff mechanism and executes a forced correction action. Specifically, the optimal energy efficiency opening prediction value output by the neural network is immediately blocked, the control strategy is switched to the preset high-gain proportional-integral control mode, the superheat setpoint is forcibly increased to the safety value, and the opening of the electronic expansion valve is rapidly reduced in increments of 5% per control cycle until the risk probability value falls below 0.2, prioritizing ensuring that the compressor suction port is in a dry steam state.

[0045] In the closed-loop feedback step, the system continuously collects the actual superheat change trajectory after the action is executed and sends it back to the time delay sensing logic; by calculating the phase difference between the predicted response curve and the actual response curve, the system iteratively corrects the calculation parameters of the thermal response lag time window online; this closed-loop mechanism gives the system the ability to adaptively compensate for the drift of physical parameters such as equipment aging and heat exchanger fouling, ensuring the stability of control accuracy throughout the entire life cycle of the equipment.

[0046] Example 2:

[0047] The saturation distance vector is a physical state descriptor; the saturation distance vector represents the enthalpy or entropy margin of the current refrigerant state from the point of liquefaction; the process of generating the saturation distance vector in step 1 includes: generating the saturation distance vector and passing it as the only physical feature to step 2 to solve the problem of lack of physical context in single sensor data.

[0048] In the specific technical context of this embodiment, the saturation distance vector is strictly defined as a physical state descriptor characterizing the phase change potential of a refrigerant. The value of this descriptor not only reflects the temperature difference but, based on the thermodynamic equation of state, quantifies the enthalpy and entropy margins of the current refrigerant state point from entering the gas-liquid two-phase region, constructing a two-dimensional feature vector. Norm indicates that the data is standardized or normalized based on a statistical distribution. The mean and variance parameters used in the statistical distribution are derived from the accumulated moving window statistics in the system's historical operating database. This maps physical quantities of different dimensions to the same dimensionless interval to eliminate enthalpy differences. With entropy difference Calculation discrepancies due to different units; among which, For enthalpy difference, The entropy difference;

[0049] In the specific execution of step 1, the operation of generating the saturation distance vector aims to construct the physical context. In the prior art, neural networks usually directly receive temperature and pressure readings, which causes the model to consume a large number of parameters to learn the implicit thermodynamic laws and is prone to failure under extreme conditions not covered by training data. In this embodiment, the original sensor data is transformed into a saturation distance vector with clear physical meaning through pre-calculation of thermodynamics, and this vector is used as the core thermodynamic feature, along with the current opening degree of the electronic expansion valve and the operating frequency of the compressor, and then passed to step 2. This design reduces the dimensionality redundancy of the subsequent neural network while preserving the actuator state and system load boundary, and forces the model to focus on the core feature of how far away from the danger boundary. Thus, at the data level, through dimensionality reduction and feature reconstruction, the problem of poor model generalization ability caused by the lack of physical constraints in single sensor data is solved.

[0050] Example 3:

[0051] The process of calculating the thermal response lag time window in step 2 includes: monitoring the change gradient of superheat represented by the saturation distance vector; if the rate of change accelerates, it is determined that the phase change of the refrigeration system is drastic and the lag time is shortened; the thermal response lag time window is narrowed; if the rate of change slows down, the thermal response lag time window is widened; step 2 uses the thermal response lag time window to determine the range of historical data to be read in subsequent steps, so as to ensure that the historical state data sequence is strictly aligned in terms of time causality.

[0052] To clarify the implementation details of the dynamic time-delay sensing mechanism, this embodiment employs a gradient-based adaptive window adjustment strategy. The system continuously calculates the gradient of the input saturation distance vector over a unit of time. Specifically, the gradient calculation method involves calculating the two-dimensional feature vector. The rate of change of the Euclidean norm with time, i.e., the calculation This scalar result characterizes the rate of state point movement;

[0053] This gradient directly maps the intensity of the refrigerant phase change and its flow rate characteristics inside the evaporator. The specific parameter adjustment logic is as follows: When an increased rate of change is detected, the system determines, based on fluid dynamics principles, that the current refrigerant flow rate has increased, resulting in improved heat exchange efficiency and a shortened physical delay time between the control signal and the temperature sensor. In response, the algorithm module automatically reduces the length of the thermal response hysteresis window, for example, from the standard 60-second window to 30 seconds. The specific time window adjustment formula can be expressed as: ;in, The unit time constant is used to balance the dimensions and ensure that the input to the hyperbolic tangent function is a dimensionless value. As the baseline window length, For adjustment coefficients, The function is used to limit the boundary of the adjustment range; conversely, if a slowdown in the rate of change is detected and the absolute value of the gradient is lower than the preset steady-state threshold, it indicates that the system load has decreased or entered a steady state, and the thermal inertia has increased. The algorithm module then expands the thermal response lag time window, for example, to 120 seconds. This step uses the dynamically calculated time window to accurately define the start and end time points of reading data from the historical database. After reading the historical data, the system performs a time series resampling operation, using a linear interpolation algorithm to uniformly map the historical data sequences of different time lengths into fixed-length time steps to meet the fixed-length input requirements of the subsequent time series prediction deep neural network model. The input requirement is to concatenate the resampled saturation distance vector sequence with the corresponding historical opening sequence in the feature dimension to construct a tensor input of dimension .

[0054] This process ensures that each set of historical state data sequences input into the neural network corresponds strictly to the current control moment in terms of physical causality, effectively eliminating pseudo-causal data noise caused by fixed time steps and eliminating the control oscillation phenomenon commonly found in variable time delay systems.

[0055] Example 4:

[0056] The time-series prediction deep neural network model uses a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU); the asymmetric penalty loss function assigns uneven weights to the prediction error; if the state predicted by the time-series prediction deep neural network model shows that the superheat is likely to fall below zero and enter the liquid phase region, the asymmetric penalty loss function assigns a first penalty weight; if the prediction result shows that the superheat is too large, the asymmetric penalty loss function assigns a second penalty weight; the value of the first penalty weight is greater than the value of the second penalty weight.

[0057] In this embodiment, the time-series prediction deep neural network model uses an architecture with long sequence memory capability, preferably a long short-term memory network or a gated recurrent unit, to capture the thermal inertia dependency of the refrigeration system over a long period of time.

[0058] The core innovation of this model lies in its use of an asymmetric penalty loss function in its training objective function. This function is not merely a mathematical weighting, but also an algorithmic translation of refrigeration process safety regulations. Specifically, the asymmetric penalty loss function... The mathematical expression is as follows:

[0059]

[0060] in, This represents the actual overheat value. For predicted values, For indicator functions, Corresponding to the first penalty weight, Corresponding to the second penalty weight, and satisfying ; The total number of time steps for the historical state data sequence of the input model;

[0061] In practice, the function includes a conditional judgment logic: when the superheat state value predicted by the model is less than or equal to zero, that is, when the prediction shows a tendency to fall below zero and enter the liquid phase region, the loss function calls the first penalty weight with a very large value to amplify the error, forcing the model to adjust the parameters significantly during backpropagation to move away from the region; conversely, when the prediction result shows that the superheat is too large, the loss function calls the second penalty weight with a smaller value, allowing the model to have a certain energy efficiency optimization error.

[0062] Among them, the first penalty weight is significantly greater than the second penalty weight; this asymmetric weight allocation gives the trained neural network an inherent risk aversion characteristic during inference, and it will exhibit extremely high sensitivity when approaching the liquid impact risk zone, thus achieving safety protection for the compressor hardware at the algorithm level.

[0063] Example 5:

[0064] The signal output in step 3 is defined as follows:

[0065] Predicted optimal energy efficiency opening: The target opening of the electronic expansion valve that achieves the highest energy efficiency ratio, inferred based on the characteristics of the current operating conditions;

[0066] Liquid slugging risk probability value: a confidence score that quantifies the degree of danger of liquid slugging due to the current state approaching the gas-liquid saturation line; the confidence score ranges from 0 to 1.

[0067] In this embodiment, the key control parameters of the system output are clearly defined in engineering; the optimal energy efficiency opening prediction value refers to a specific target number of steps or opening percentage of the electronic expansion valve stepper motor after the model comprehensively considers the current saturation distance vector and historical time series characteristics. Its optimization objective function is to maximize the energy efficiency ratio of the system.

[0068] The liquid slugging risk probability value is a normalized confidence score, strictly limited to a range of 0 to 1. This value is not a physical measurement, but a statistical assessment of the model's safety under the current operating conditions. The closer the value is to 1, the closer the current refrigerant state point is to the gas-liquid saturation line in the pressure-enthalpy diagram, and the extremely high probability of liquid slugging. The closer the value is to 0, the more likely it is to be in the deep superheated safety zone. The calculation of this probability value is usually based on the activation value of the softmax function in the model output layer or the output of a specific risk assessment subnet. By clearly defining these two signals, the control objective and the constraint objective are decoupled, providing a clear decision-making basis for subsequent execution units.

[0069] Example 6:

[0070] The closed-loop feedback process in step 5 includes: using the actual response as the feedback signal; the actual response is the change in superheat actually measured by the sensor after the command is executed; comparing the time difference between the predicted response and the actual response corresponding to the optimal energy efficiency opening value; and correcting the time window calculation parameters inside the dynamic time delay sensing step online based on the time difference to adapt to the drift of physical parameters caused by equipment aging.

[0071] The correction process includes a logic gating mechanism: when the probability value of liquid slamming risk output by the dual-modal constraint execution unit is lower than the safety threshold and the execution action is not covered by the safety backoff mechanism, the system performs online correction in real time; when the system triggers the safety backoff mechanism, the current actual response data and prediction deviation are marked as samples to be corrected and stored in a buffer queue. After the system returns to steady state, the data in the buffer queue is used for offline parameter correction to prevent the introduction of erroneous feedback gain due to the inconsistency between the safety forced action and the model prediction action, while ensuring that the system can correct the fundamental parameter deviation that leads to high risk.

[0072] This embodiment details the system's adaptive calibration mechanism; the closed-loop feedback process aims to solve the model mismatch problem caused by component wear, dust accumulation, or minor refrigerant leakage during long-term operation of the refrigeration system;

[0073] The system records the moment after executing a valve opening command and continuously tracks the actual response from the sensor, i.e., the moment and magnitude of the actual change in superheat. The processing logic records the timestamp Tcmd of the valve opening command that produces a significant change and locks the timestamp of the significant change in the superheat value reported by the sensor. Calculate the actual response delay ;in, It is the timestamp when the sensor feedback value changes significantly; This is the timestamp of the valve opening command issued; the system has a pre-set standard response mapping model generated based on the standard operating condition test calibration before leaving the factory. This model is constructed as a two-dimensional lookup table with the valve opening change as the index key and the time response curve as the value. This model stores the theoretical superheat response curve and theoretical response delay time corresponding to different electronic expansion valve opening changes under standard operating conditions; the system calls this mapping model, inputs the best energy efficiency opening prediction value generated in step 3, and queries to obtain the corresponding theoretical response delay value and theoretical superheat change amplitude, which are used as the comparison benchmark;

[0074] The system will calculate the actual response delay. Compare with the theoretical delay value set in the current thermal response hysteresis time window; the theoretical delay value is obtained based on a preset standard response mapping model, which stores the reference delay time corresponding to different valve opening changes under standard operating conditions; if it is found If the deviation from the theoretical delay value exceeds the preset tolerance range, for example, if the deviation exceeds 2 seconds, the system determines that the current physical parameters have drifted. At this time, the algorithm will correct the time window calculation parameters within the dynamic time delay perception step online based on the magnitude and direction of the time difference, such as adjusting the mapping coefficient between the gradient and the time window length. Through this continuous online iteration, the system can automatically adapt to the changes in physical characteristics caused by equipment aging without manual intervention to recalibrate the parameters. In addition, for the heat exchange efficiency decay caused by heat exchanger fouling, the closed-loop feedback step also includes a gain compensation mechanism: the system calculates the ratio of the actual superheat change to the theoretical expectation and generates an aging compensation coefficient. This coefficient is fed back to the execution unit to linearly correct the predicted value of the optimal energy efficiency opening, thereby compensating for the control gain mismatch caused by the drift of physical parameters while correcting the time delay.

[0075] Example 7:

[0076] Please see Figure 2A system for optimizing superheat based on machine learning to improve cooling efficiency includes: a thermodynamic state-space mapping unit, a dynamic time delay sensing unit, an asymmetric risk assessment unit, and a dual-modal constraint execution unit;

[0077] The thermodynamic state space mapping unit is configured to receive sensor data, retrieve refrigerant property parameter library, reconstruct physical characteristics, and output saturation distance vector to dynamic time delay sensing unit.

[0078] The dynamic time delay sensing unit is configured to monitor the rate of change of the saturation distance vector, calculate the thermal response lag time window, and transmit the aligned historical state data sequence to the asymmetric risk assessment unit.

[0079] The asymmetric risk assessment unit has a built-in time-series prediction deep neural network model, which is configured to output the best energy efficiency opening prediction value and liquid hammer risk probability value to the dual-modal constraint execution unit based on the asymmetric penalty loss function.

[0080] The dual-modal constraint execution unit is configured to execute valve actions based on the comparison result between the liquid hammer risk probability value and the safety threshold, and feed the actual response back to the dynamic time delay sensing unit to form a closed-loop control.

[0081] This embodiment constructs a modular and collaborative control system architecture; the thermodynamic state space mapping unit serves as the data acquisition and preprocessing center, integrating a high-precision A / D conversion interface and an embedded physical property database, and is responsible for converting discrete sensor voltage signals into saturation distance vectors with physical meaning;

[0082] The dynamic time delay sensing unit, as the core of time sequence alignment processing, has a built-in sliding window algorithm and gradient monitoring logic. It is responsible for dynamically adjusting the data truncation window according to the working conditions and providing the historical state data sequence of time causality alignment to the subsequent units.

[0083] The asymmetric risk assessment unit, as the core of the system's computational inference, is equipped with a pre-trained time-series prediction deep neural network model and has asymmetric penalty logic embedded in the inference engine. It is responsible for parallel computation and outputting the optimal energy efficiency opening prediction value and the probability value of liquid slugging risk. The dual-modal constraint execution unit, as the decision execution terminal, is equipped with a logic comparator and a PID drive circuit. Based on the logical relationship between the received risk probability value and the preset safety threshold, this unit switches between high-efficiency control mode and safety protection mode in milliseconds. At the same time, it is responsible for collecting and transmitting the actual response data after execution, forming a closed-loop control data link. All units are interconnected through a high-speed bus to jointly realize an intelligent refrigeration control system that combines high efficiency and high safety.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for optimizing superheat based on machine learning to improve cooling efficiency, characterized in that, include: Step 1, Thermodynamic State Space Mapping: Receive real-time suction pressure data and real-time suction temperature data of the refrigeration system, retrieve the refrigerant property parameter library, determine the position of the gas-liquid saturation line in the pressure-enthalpy logical coordinate system, calculate the thermodynamic distance between the current suction state point and the gas-liquid saturation line, and generate a saturation distance vector. Step 2, Dynamic Time Lag Sensing: Monitor the rate of change of the saturation distance vector, calculate the thermal response lag time window using gradient-based adaptive adjustment logic, and extract historical state data sequences based on the thermal response lag time window; Step 3, Asymmetric Risk Assessment: Input the historical state data sequence into the time-series prediction deep neural network model, use the asymmetric penalty loss function to constrain the time-series prediction deep neural network model, and output the best energy efficiency opening prediction value and liquid hammer risk probability value in parallel. Step 4, Dual-modal constraint execution: Compare the liquid slugging risk probability value with the safety threshold. If the liquid slugging risk probability value is lower than the safety threshold, output the optimal energy efficiency opening prediction value to the electronic expansion valve. If the liquid slugging risk probability value is higher than the safety threshold, trigger the safety backoff mechanism to correct the opening of the electronic expansion valve. Step 5, Closed-loop feedback: Collect the actual superheat change after the execution command; transmit the actual superheat change back to Step 2 to correct the calculation parameters of the thermal response hysteresis window.

2. The method for improving cooling efficiency by optimizing superheat based on machine learning according to claim 1, characterized in that, The saturation distance vector is a physical state descriptor; the saturation distance vector represents the enthalpy or entropy margin of the current refrigerant state from the point of liquefaction; the process of generating the saturation distance vector in step 1 includes: generating the saturation distance vector and passing it as a unique physical feature to step 2 to solve the problem of lack of physical context in single sensor data.

3. The method for improving cooling efficiency by optimizing superheat based on machine learning according to claim 1, characterized in that, The process of calculating the thermal response lag time window in step 2 includes: monitoring the change gradient of superheat represented by the saturation distance vector; if the rate of change accelerates, it is determined that the phase change of the refrigeration system is drastic and the lag time is shortened; the thermal response lag time window is narrowed; if the rate of change slows down, the thermal response lag time window is widened; step 2 uses the thermal response lag time window to determine the range of historical data to be read in subsequent steps, so as to ensure that the historical state data sequence is strictly aligned in terms of time causality.

4. The method for improving cooling efficiency by optimizing superheat based on machine learning according to claim 1, characterized in that, The time-series prediction deep neural network model employs a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU); the asymmetric penalty loss function assigns uneven weights to the prediction error; if the state predicted by the time-series prediction deep neural network model shows that the superheat is trending towards falling below zero and entering the liquid phase region, the asymmetric penalty loss function assigns a first penalty weight; if the prediction result shows that the superheat is too large, the asymmetric penalty loss function assigns a second penalty weight. The first penalty weight value is greater than the second penalty weight value.

5. The method for improving cooling efficiency by optimizing superheat based on machine learning according to claim 1, characterized in that, The signal output in step 3 is defined as follows: The optimal energy efficiency opening prediction value is the target opening of the electronic expansion valve that can achieve the highest energy efficiency ratio, inferred based on the characteristics of the current operating conditions. The liquid slugging risk probability value is a confidence score that quantifies the degree of danger of liquid slugging occurring when the current state approaches the gas-liquid saturation line; the confidence score ranges from 0 to 1.

6. The method for optimizing superheat based on machine learning to improve cooling efficiency according to claim 1, characterized in that, The closed-loop feedback process in step 5 includes: using the actual response as a feedback signal; the actual response is the change in superheat actually measured by the sensor after the command is executed; comparing the time difference between the predicted response corresponding to the optimal energy efficiency opening value and the actual response; and correcting the time window calculation parameters inside the dynamic time delay sensing step online based on the time difference to adapt to the drift of physical parameters caused by equipment aging.

7. A system for optimizing superheat based on machine learning to improve refrigeration efficiency, applied to the method for optimizing superheat based on machine learning to improve refrigeration efficiency as described in any one of claims 1 to 6, characterized in that, include: Thermodynamic state-space mapping unit, dynamic time-delay sensing unit, asymmetric risk assessment unit, and bimodal constraint execution unit; The thermodynamic state space mapping unit is configured to receive sensor data, retrieve refrigerant property parameter library, reconstruct physical characteristics, and output saturation distance vector to the dynamic time delay sensing unit. The dynamic time delay sensing unit is configured to monitor the rate of change of the saturation distance vector, calculate the thermal response hysteresis time window, and transmit the aligned historical state data sequence to the asymmetric risk assessment unit. The asymmetric risk assessment unit has a built-in time-series prediction deep neural network model and is configured to output the best energy efficiency opening prediction value and liquid hammer risk probability value to the dual-modal constraint execution unit based on the asymmetric penalty loss function. The dual-modal constraint execution unit is configured to execute valve actions based on the comparison result between the liquid hammer risk probability value and the safety threshold, and to feed back the actual response to the dynamic time delay sensing unit to form a closed-loop control.

Images