A machine learning-based refrigeration system energy consumption optimization method
By constructing a thermodynamic benchmark and physical perturbation model based on the inverse Carnot cycle, a virtual fault state vector is generated, which solves the problem of insufficient robustness of fault diagnosis in refrigeration systems when real fault samples are lacking, and improves the accuracy of early fault identification and energy consumption optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHUNDE SANSHENG ELECTRICAL MFG CO LTD
- Filing Date
- 2025-12-04
- Publication Date
- 2026-05-15
AI Technical Summary
In the absence of real fault samples, existing refrigeration systems rely on historical data, resulting in insufficient robustness in fault diagnosis. This makes it difficult to accurately identify faults in their early stages, and energy consumption optimization strategies fail.
A thermodynamic ideal benchmark is constructed based on the inverse Carnot cycle principle and equipment design parameters to generate an ideal state vector. A virtual fault state vector is generated by superimposing physical disturbances. The fault type is identified by the projection similarity of the feature residual vector, and a closed-loop correction strategy is executed.
It enables accurate identification and early detection of various fault modes in the absence of real fault samples, improves fault diagnosis sensitivity and energy consumption optimization accuracy, and ensures the safe operation of the system under abnormal conditions.
Smart Images
Figure CN121383368B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control and fault diagnosis technology for HVAC systems, specifically a machine learning-based method for optimizing energy consumption in refrigeration systems. Background Technology
[0002] In the operation, maintenance and energy management of refrigeration systems, equipment needs to operate for a long time under varying ambient temperatures and load conditions, and its performance status evolves dynamically over time.
[0003] To achieve energy conservation and consumption reduction in systems, existing solutions generally adopt pure data-driven machine learning models based on historical operating data. Although this approach can utilize past records for training, the lack of sufficient negative fault samples in historical datasets limits the model's generalization ability. Furthermore, relying solely on statistical regularities makes it difficult to decouple uncontrollable environmental variables from system response variables, making the algorithm prone to misjudging long-term minor faults or equipment physical performance degradation as normal feature evolution. This over-reliance on historical data makes it difficult for the system to accurately identify faults in their early stages, leading to inaccurate energy consumption baselines, insufficient diagnostic robustness, and ineffective optimization strategies. Therefore, overcoming the limitations of relying solely on historical data and improving fault diagnosis sensitivity and energy consumption optimization accuracy in the absence of real fault samples has become an urgent technical problem to be solved. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a machine learning-based method for optimizing the energy consumption of a refrigeration system. Specifically, the technical solution of this invention includes:
[0005] Step 1: Collect the wet-bulb temperature, outlet water set temperature and real-time cooling load of the refrigeration system; calculate the ideal condensing temperature and ideal evaporating temperature based on the reverse Carnot cycle principle; calculate the zero entropy loss reference power using the isentropic efficiency of the equipment design; combine the ideal condensing pressure, ideal evaporating pressure and zero entropy loss reference power to generate the ideal state vector.
[0006] Step 2: Construct a fault characteristic sensitivity matrix using the refrigeration system mechanism model, and use the first-order Taylor series expansion principle to superimpose physical disturbances on the ideal state vector to generate a virtual fault state vector.
[0007] Step 3: Collect real operating data, calculate the difference ratio between real operating data and ideal state vector to generate real-time feature residual vector, calculate the difference ratio between virtual fault state vector and ideal state vector to generate virtual feature residual vector, calculate the projection similarity of real-time feature residual vector in the direction of virtual feature residual vector, and identify fault type based on projection similarity.
[0008] Step 4: In response to the fault type, invoke the preset closed-loop correction strategy to adjust the control parameters of the refrigeration system or execute an alarm operation.
[0009] Preferably, the steps in step one for calculating the ideal condensation temperature and the ideal evaporation temperature include:
[0010] Obtain the design cooling approach temperature and design evaporator temperature difference from the equipment's manufacturer's performance specifications.
[0011] The sum of the wet-bulb temperature and the design cooling approach temperature is defined as the ideal condensation temperature.
[0012] The difference between the set water temperature and the design evaporation temperature is calculated and defined as the ideal evaporation temperature.
[0013] Preferably, the step of generating the ideal state vector in step one further includes:
[0014] Calculate the Carnot theoretical energy efficiency coefficient based on ideal condensation temperature and ideal evaporation temperature;
[0015] The compression ratio is determined based on the ideal condensing temperature and the ideal evaporating temperature. The compressor performance curve is then retrieved based on the compression ratio to obtain the design isentropic efficiency.
[0016] The zero-entropy loss baseline power is calculated using real-time cooling load, Carnot theoretical energy efficiency coefficient, and design isentropic efficiency.
[0017] The saturation pressure corresponding to the ideal condensation temperature is obtained as the ideal condensation pressure.
[0018] The saturation pressure corresponding to the ideal evaporation temperature is obtained as the ideal evaporation pressure.
[0019] An ideal state vector is generated by combining the ideal condensation pressure, the ideal evaporation pressure, and the zero entropy loss reference power.
[0020] Preferably, step two includes:
[0021] A fault characteristic sensitivity matrix is established based on the numerical solution of the differential equations of the refrigeration system mechanism.
[0022] The partial derivatives of physical parameters with respect to fault variables are determined using the fault characteristic sensitivity matrix, and a sensitivity Jacobian vector is constructed.
[0023] Among them, the sensitivity Jacobian vector represents the absolute change in physical parameters caused by a unit fault intensity;
[0024] An additive perturbation model is used to multiply the sensitivity Jacobian vector by the preset fault evolution intensity and superimpose it onto the ideal state vector to generate a virtual fault state vector.
[0025] Preferably, the fault evolution intensity is a numerical sequence generated by discretizing within a preset interval according to a step size;
[0026] The numerical sequence covers the entire range of faults, from minor to severe.
[0027] The virtual fault state vector corresponds to the evolution state under different fault mechanisms, where the evolution state includes the condenser fouling state or the abnormal refrigerant charge state.
[0028] Preferably, step three, which generates the real-time feature residual vector and the virtual feature residual vector, includes:
[0029] Calculate the difference between the actual running data and the ideal state vector, divide the difference by the corresponding benchmark value in the ideal state vector, and generate a real-time feature residual vector;
[0030] Calculate the difference between the virtual fault state vector and the ideal state vector, divide the difference by the corresponding reference value in the ideal state vector, and generate a virtual feature residual vector.
[0031] Among them, the real-time feature residual vector and the virtual feature residual vector convert the power deviation and pressure deviation into relative percentage change rates.
[0032] Preferably, step three, which involves calculating the projection similarity, includes:
[0033] The projection similarity of the real-time feature residual vector onto the virtual feature residual vector is calculated using the cosine similarity formula.
[0034] Among them, projection similarity represents the directional consistency between real-world degradation features and theoretical fault fingerprints.
[0035] Preferably, the step of identifying the fault type in step three includes:
[0036] The maximum value among the projection similarities is selected as the maximum projection similarity.
[0037] The maximum projection similarity is compared with a preset confidence threshold.
[0038] If the maximum projection similarity is greater than the confidence threshold, the refrigeration system is determined to have experienced the fault type corresponding to the maximum projection similarity.
[0039] If the maximum projection similarity is less than or equal to the confidence threshold, the refrigeration system is determined to be in an unstructured environmental noise state and no fault identification is triggered.
[0040] Preferably, step four includes:
[0041] If the fault type is condenser fouling, execute the flow compensation strategy, calculate the additional cooling water flow to compensate for the decrease in heat transfer coefficient, increase the cooling water flow and keep the chiller setpoint unchanged;
[0042] If the fault type is sensor deviation, execute the parameter locking strategy, lock the current control parameters and disable the automatic adjustment logic, and issue a calibration alarm signal.
[0043] Compared with the prior art, the present invention has the following beneficial effects:
[0044] 1. This method replaces the traditional modeling approach that relies solely on historical data by constructing a thermodynamic ideal benchmark based on the reverse Carnot cycle principle and equipment design parameters. This method can decouple uncontrollable environmental variables from system response variables, ensuring the purity of the energy consumption benchmark and effectively avoiding the problem of misjudging equipment aging or minor faults caused by long-term operation as normal operating conditions due to over-reliance on historical statistical data, thus improving the accuracy of benchmark establishment.
[0045] 2. This method utilizes a hybrid modeling scheme driven by physical mechanisms to generate virtual fault state vectors by superimposing physical perturbations on ideal state vectors. This method of simulating fault occurrence in digital space can generate virtual fault samples covering the entire spectrum from minor to severe in the absence of real fault negative samples. It solves the technical defect of traditional data-driven models that have poor generalization ability due to insufficient fault samples, and realizes the pre-simulation and coverage of various fault modes.
[0046] 3. This method uses the projection similarity of feature residual vectors as the fault identification criterion, converts power deviation and pressure deviation into relative percentage change rates, and uses cosine similarity to focus on the directional consistency of fault features. This approach not only eliminates the calculation error caused by the inconsistency of physical dimensions, but also eliminates the interference of fault severity on the judgment, enabling the system to identify faults in the early stage of their occurrence through accurate matching of feature directions, which significantly improves the sensitivity of fault diagnosis and the ability to capture early minor faults.
[0047] 4. This method constructs a closed-loop correction strategy for different fault types, which can automatically perform flow compensation or parameter locking operations based on the identification results. For performance degradation faults such as condenser fouling, the system heat exchange efficiency is maintained by accurately calculating the additional flow that needs to be compensated. For sensing faults such as sensor deviation, the system oscillation is prevented by locking the control logic. This mechanism ensures the safe operation and energy efficiency optimization of the refrigeration system under abnormal operating conditions, and realizes closed-loop management of the entire process from fault diagnosis to control optimization. Attached Figure Description
[0048] The present invention will be further explained below with reference to the accompanying drawings and embodiments:
[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0050] 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.
[0051] Example 1:
[0052] Please see Figure 1 A machine learning-based method for optimizing the energy consumption of a refrigeration system, comprising the following steps:
[0053] Step 1: Collect the wet-bulb temperature, outlet water set temperature and real-time cooling load of the refrigeration system; calculate the ideal condensing temperature and ideal evaporating temperature based on the reverse Carnot cycle principle; calculate the zero entropy loss reference power using the isentropic efficiency of the equipment design; combine the ideal condensing pressure, ideal evaporating pressure and zero entropy loss reference power to generate the ideal state vector.
[0054] Step 2: Construct a fault characteristic sensitivity matrix using the refrigeration system mechanism model, and use the first-order Taylor series expansion principle to superimpose physical disturbances on the ideal state vector to generate a virtual fault state vector.
[0055] Step 3: Collect real operating data, calculate the difference ratio between real operating data and ideal state vector to generate real-time feature residual vector, calculate the difference ratio between virtual fault state vector and ideal state vector to generate virtual feature residual vector, calculate the projection similarity of real-time feature residual vector in the direction of virtual feature residual vector, and identify fault type based on projection similarity.
[0056] Step 4: In response to the fault type, invoke the preset closed-loop correction strategy to adjust the control parameters of the refrigeration system or execute an alarm operation.
[0057] This embodiment provides a machine learning-based method for optimizing the energy consumption of a refrigeration system. This method aims to address the problems in existing technologies where relying solely on historical data for training leads to the misclassification of long-term minor faults as normal features, and the poor generalization ability of the model due to a lack of negative fault samples. This embodiment employs a hybrid modeling scheme driven by physical mechanisms, specifically including the following steps:
[0058] Step 1: Construct a thermodynamic ideal baseline; collect real-time boundary conditions of the refrigeration system, including wet-bulb temperature, outlet water setpoint temperature, and real-time cooling load; based on the reverse Carnot cycle principle, combined with the equipment's factory design parameters and real-time boundary conditions, reconstruct the zero-entropy loss theoretical limit state under the current operating conditions; on this basis, calculate the ideal condensing temperature and ideal evaporating temperature using the equipment design parameters, and calculate the zero-entropy loss baseline power using the compressor's design characteristics; combine the ideal condensing temperature, ideal evaporating temperature, and zero-entropy loss baseline power to generate an ideal state vector representing a fault-free state;
[0059] Step 2: Generate virtual fault samples; construct a fault feature sensitivity matrix using the refrigeration system mechanism model; based on the first-order Taylor series expansion principle, superimpose physical disturbances of different types and intensities in the mathematical space with the ideal state vector as the origin; this process simulates the occurrence of faults in the digital space, thereby generating a virtual fault state vector containing multiple fault modes.
[0060] Step 3: Vector Projection Decision; Collect real-world operating data; Calculate the difference ratio between the real-world operating data and the ideal state vector to generate a real-time feature residual vector; Simultaneously, calculate the difference ratio between the virtual fault state vector and the ideal state vector to generate a virtual feature residual vector; Calculate the projection similarity of the real-time feature residual vector in the direction of the virtual feature residual vector using the cosine similarity algorithm; Based on the magnitude of the projection similarity, identify whether the current system has experienced a fault and what type of fault it has experienced.
[0061] Step 4: Closed-loop feedback control; In response to the identified fault type, the system automatically calls the preset closed-loop correction strategy; If it is determined that the equipment performance is deteriorating, the control parameters of the refrigeration system are adjusted compensatorily; If it is determined that the sensing layer is faulty, an alarm operation is executed and the control logic is locked.
[0062] Through the above steps, this embodiment constructs an ideal state vector as an absolute benchmark, decoupling uncontrollable environmental variables from system response variables. At the same time, through virtual fault injection technology, a full spectrum of fault samples is generated without destroying physical entities, enabling the system to accurately identify faults in the early stages of their occurrence through projection analysis of residual vectors, thereby improving the robustness of energy consumption optimization and the sensitivity of fault diagnosis.
[0063] Example 2:
[0064] Step one involves calculating the ideal condensation temperature and the ideal evaporation temperature, including:
[0065] Obtain the design cooling approach temperature and design evaporator temperature difference from the equipment's manufacturer's performance specifications.
[0066] The sum of the wet-bulb temperature and the design cooling approach temperature is defined as the ideal condensation temperature.
[0067] The difference between the set water temperature and the design evaporation temperature is calculated and defined as the ideal evaporation temperature.
[0068] This embodiment is a further specification of step one in embodiment 1, and details the calculation logic of the ideal temperature boundary;
[0069] In step one, the specific process for calculating the ideal condensation temperature and the ideal evaporation temperature is as follows:
[0070] Obtain two key design parameters from the equipment's manufacturer's specifications: the design cooling approach temperature. and the design evaporator terminal difference Among them, the design cooling approach temperature Characterize the condenser's ability to approach the wet-bulb temperature at the condensing temperature under ideal heat exchange conditions; design the evaporator terminal temperature difference. Characterizes the evaporator's ability to approach the outlet water temperature at the limit under ideal heat exchange conditions;
[0071] The sum of the wet-bulb temperature and the design cooling approach temperature is defined as the ideal condensation temperature. The calculation formula is as follows:
[0072]
[0073] in, The outdoor wet-bulb temperature is collected in real time by the sensor, in degrees Celsius. These are preset constants derived from the equipment manufacturer's specifications, and the unit is degrees Celsius.
[0074] Simultaneously, the difference between the set water temperature and the design evaporation temperature is calculated and defined as the ideal evaporation temperature. The calculation formula is as follows:
[0075]
[0076] in, Set the outlet water temperature for the chiller unit, in degrees Celsius; These are preset constants derived from the equipment manufacturer's specifications, and the unit is degrees Celsius.
[0077] By introducing a design cooling approach temperature and the design evaporator terminal difference The ideal temperature reference constructed in this embodiment is based entirely on the thermodynamic design capabilities of the equipment, rather than on historical statistical data that may have experienced performance degradation; this ensures the purity of the reference and can effectively prevent the algorithm from misjudging equipment aging caused by long-term operation as normal operating conditions.
[0078] Example 3:
[0079] The steps in step one for generating the ideal state vector also include:
[0080] Calculate the Carnot theoretical energy efficiency coefficient based on ideal condensation temperature and ideal evaporation temperature;
[0081] The compression ratio is determined based on the ideal condensing temperature and the ideal evaporating temperature. The compressor performance curve is then retrieved based on the compression ratio to obtain the design isentropic efficiency.
[0082] The zero-entropy loss baseline power is calculated using real-time cooling load, Carnot theoretical energy efficiency coefficient, and design isentropic efficiency.
[0083] The saturation pressure corresponding to the ideal condensation temperature is obtained as the ideal condensation pressure.
[0084] The saturation pressure corresponding to the ideal evaporation temperature is obtained as the ideal evaporation pressure.
[0085] An ideal state vector is generated by combining the ideal condensation pressure, the ideal evaporation pressure, and the zero entropy loss reference power.
[0086] This embodiment is a further specification of Embodiment 2, detailing how to construct a complete ideal state vector based on an ideal temperature;
[0087] In step one, the ideal state vector is generated. The steps also include:
[0088] Based on the calculated ideal condensation and evaporation temperatures, a reverse Carnot cycle model is constructed using an absolute temperature scale, and the Carnot theoretical energy efficiency coefficient is calculated. The calculation formula is as follows:
[0089]
[0090] The constant 273.15 is used to convert Celsius to Kelvin absolute temperature scale;
[0091] The compression ratio is determined based on the ideal condensation temperature and the ideal evaporation temperature. Based on this compression ratio, the performance curve provided by the compressor manufacturer is retrieved to obtain the design isentropic efficiency under this operating condition. If a detailed performance mapping table provided by the manufacturer is unavailable, the following general binomial fitting model is used to calculate the design isentropic efficiency. :
[0092]
[0093] in, The compression ratio calculated in the preceding steps, The empirical fitting coefficients are preset based on the compressor type. For centrifugal compressors, a typical value can be set to... ;
[0094] Compression ratio The calculation formula is as follows:
[0095]
[0096] in, The refrigerant temperature-pressure saturation function is used in this step. The performance curve refers to the isentropic efficiency mapping table of the compressor at different compression ratios, which is pre-stored in the system database.
[0097] Using real-time cooling load, Carnot's theoretical energy efficiency coefficient, and design isentropic efficiency, the theoretical minimum power consumption under the current operating conditions, i.e., the zero-entropy loss reference power, is calculated. The calculation formula is as follows:
[0098]
[0099] in, The terminal cooling load is collected in real time, and the unit is kilowatts; The calculated theoretical minimum power consumption is expressed in kilowatts.
[0100] Obtain the ideal condensing temperature using the refrigerant saturation property table. The corresponding saturation pressure is taken as the ideal condensation pressure. To obtain the ideal evaporation temperature The corresponding saturation pressure is taken as the ideal evaporation pressure. The above parameters are combined to generate an ideal state vector. :
[0101]
[0102] This embodiment introduces a dynamic query mechanism for the compressor performance curve, enabling the calculation of the zero-entropy loss reference power. It is not a static theoretical value, but a dynamic benchmark that fits the actual physical characteristics of the equipment; it combines pressure and power into a multi-dimensional vector, providing a mathematical basis for the subsequent geometric projection of fault characteristics in multi-dimensional space.
[0103] Example 4:
[0104] Step two includes:
[0105] A fault characteristic sensitivity matrix is established based on the numerical solution of the differential equations of the refrigeration system mechanism.
[0106] The partial derivatives of physical parameters with respect to fault variables are determined using the fault characteristic sensitivity matrix, and a sensitivity Jacobian vector is constructed.
[0107] Among them, the sensitivity Jacobian vector represents the absolute change in physical parameters caused by a unit fault intensity;
[0108] An additive perturbation model is used to multiply the sensitivity Jacobian vector by the preset fault evolution intensity and superimpose it onto the ideal state vector to generate a virtual fault state vector.
[0109] This embodiment is a further specification of step two in embodiment 1, and explains in detail the generation mechanism of the virtual fault vector;
[0110] In step two, to simulate fault behavior in the absence of real fault data, a fault characteristic sensitivity matrix is established based on the numerical solution of the differential equations governing the refrigeration system mechanism. The differential equations governing the refrigeration system mechanism include the condenser heat transfer equation and the compressor polytropic compression equation. Specifically, the condenser heat transfer equation defines the condensing pressure. With condenser heat transfer coefficient The coupling relationship between them; where, This is a function relating refrigerant saturation temperature and pressure. Wet-bulb temperature, The heat exchange area of the condenser. For real-time cooling load, The compressor power is given by the following formula:
[0111]
[0112] The polytropic compression equation of a compressor defines the compressor power. The functional relationship between system pressure and isentropic efficiency is expressed as follows:
[0113]
[0114] in, This is the refrigerant mass flow rate. For evaporation pressure, For inhalation specific volume, For isentropic efficiency, It is a variable index, and its value is set as an adiabatic index constant based on the physical properties of the refrigerant, or obtained by real-time fitting based on the suction and discharge pressures;
[0115] Based on the above set of equations, the specific steps for establishing the fault characteristic sensitivity matrix are: to apply the above equations to the fault variable respectively. Find the partial derivatives to obtain the Jacobian vector. The various components;
[0116] This matrix is used to determine the partial derivatives of the physical parameters with respect to the fault variables, and a sensitivity Jacobian vector is constructed. ;
[0117] Sensitivity Jacobian Vector The mathematical expression for the absolute change in physical parameters caused by a unit fault intensity is:
[0118]
[0119] in, Indicates the first Physical variables related to fault types, such as the condenser heat transfer coefficient; Represents partial differential operations;
[0120] An additive perturbation model is used to multiply the sensitivity Jacobian vector by a preset fault evolution intensity and then superimpose it onto the ideal state vector to generate a virtual fault state vector. The calculation formula is as follows:
[0121]
[0122] in, The ideal state vector generated in step one; The intensity of fault evolution;
[0123] Determine the partial derivatives of the physical parameters with respect to the fault variables, specifically referring to the ideal state vector generated in step one. At the corresponding operating point, the fault characteristic sensitivity matrix is numerically differentiated.
[0124] This embodiment utilizes the Taylor series expansion principle to linearize the complex nonlinear fault mechanism into a Jacobian vector, greatly reducing computational complexity. Through this parameterized fault injection method, the system can instantly generate massive amounts of fault samples that conform to physical laws in the digital space.
[0125] Example 5:
[0126] The fault evolution intensity is a numerical sequence generated by discretizing within a preset interval according to a step size.
[0127] The numerical sequence covers the entire range of faults, from minor to severe.
[0128] The virtual fault state vector corresponds to the evolution state under different fault mechanisms, where the evolution state includes the condenser fouling state or the abnormal refrigerant charge state.
[0129] This embodiment is a further specification of embodiment 4, which limits the generation and coverage of fault evolution intensity;
[0130] In this embodiment, fault evolution intensity Within the preset range The numerical sequence is generated by discretization according to the step size; taking condenser scaling failure as an example. Representing fouling thermal resistance, this numerical sequence covers the entire range of faults from minor to severe; for example, the numerical sequence can be set from... to ;
[0131] Virtual Fault State Vector Corresponding to the evolution states under different failure mechanisms; the evolution states include the condenser fouling state, which corresponds to the condenser heat transfer coefficient. The decrease in efficiency; and the compressor's airtightness leakage state, corresponding to the compressor's isentropic efficiency. Deviation;
[0132] By generating a numerical sequence covering the entire range through discretization, this embodiment can simulate the full life cycle characteristics of a fault from its inception to its outbreak; this enables the algorithm not only to detect serious faults, but also to provide early warnings through feature matching when the fault is still in its minor stage.
[0133] Example 6:
[0134] Step three, which generates the real-time feature residual vector and the virtual feature residual vector, includes the following steps:
[0135] Calculate the difference between the actual running data and the ideal state vector, divide the difference by the corresponding benchmark value in the ideal state vector, and generate a real-time feature residual vector;
[0136] Calculate the difference between the virtual fault state vector and the ideal state vector, divide the difference by the corresponding reference value in the ideal state vector, and generate a virtual feature residual vector.
[0137] Among them, the real-time feature residual vector and the virtual feature residual vector convert the power deviation and pressure deviation into relative percentage change rates.
[0138] This embodiment is a further specification of step three in embodiment 1, focusing on the dimensionless processing of the feature residual vector;
[0139] In step three, a real-time feature residual vector is generated. and virtual feature residual vector The steps include:
[0140] Calculate real running data With the ideal state vector The difference is calculated and divided by the corresponding baseline value in the ideal state vector to generate the real-time feature residual vector. For the first vector... Each component is calculated using the following formula:
[0141]
[0142] Simultaneously, calculate the virtual fault state vector. With the ideal state vector The difference is also divided by the corresponding baseline value in the ideal state vector to generate a virtual feature residual vector. For the first vector... Each component is calculated using the following formula:
[0143]
[0144] in, The component index in the vector corresponds to condensation pressure, evaporation pressure, or power.
[0145] Through the above calculations, the real-time feature residual vector and the virtual feature residual vector convert the power deviation and pressure deviation into relative percentage change rates. This dimensionless processing corrects the calculation errors caused by the inconsistency of dimensions, avoids the large power deviation from masking the small pressure deviation, and improves the diagnostic accuracy of multi-parameter coupled faults.
[0146] Example 7:
[0147] Step three, which involves calculating projection similarity, includes:
[0148] The projection similarity of the real-time feature residual vector onto the virtual feature residual vector is calculated using the cosine similarity formula.
[0149] Among them, projection similarity represents the directional consistency between real-world degradation features and theoretical fault fingerprints.
[0150] This embodiment is a further specification of embodiment 6, and details the method for calculating projection similarity;
[0151] In step three, the real-time feature residual vector is calculated using the cosine similarity formula. In the virtual feature residual vector Projection similarity in direction The calculation formula is as follows:
[0152]
[0153] in, Represents the dot product operation of vectors; The Euclidean norm of a vector;
[0154] Projection similarity The method characterizes the consistency of the orientation between actual degradation features and theoretical fault fingerprints. This embodiment uses cosine similarity instead of Euclidean distance as the criterion because the orientation of the fault features represents the mechanism of the fault, that is, the proportional relationship of the changes of each parameter, while the magnitude only represents the severity of the fault. By focusing on orientation consistency, this method can ignore the interference of fault severity and can still accurately identify the fault type through orientation matching when the fault is very minor.
[0155] Example 8:
[0156] Step three, identifying the fault type, includes:
[0157] The maximum value among the projection similarities is selected as the maximum projection similarity.
[0158] The maximum projection similarity is compared with a preset confidence threshold.
[0159] If the maximum projection similarity is greater than the confidence threshold, the refrigeration system is determined to have experienced the fault type corresponding to the maximum projection similarity.
[0160] If the maximum projection similarity is less than or equal to the confidence threshold, the refrigeration system is determined to be in an unstructured environmental noise state and no fault identification is triggered.
[0161] This embodiment is a further specification of embodiment 7, and details the fault identification judgment logic;
[0162] Step three, identifying the fault type, includes:
[0163] The maximum value among the projection similarities is selected as the maximum projection similarity. And record the fault type corresponding to the maximum value; the judgment criteria are as follows:
[0164]
[0165] Maximum projection similarity Compared with the preset confidence threshold Comparison; in this embodiment, the confidence threshold Based on the statistical analysis of Gaussian white noise in the sensor, the following steps are taken: collect historical data of the system under normal operating conditions, calculate the cosine similarity distribution of its residual vector, and select the lower bound of the 95% confidence interval of this distribution as the threshold, for example, set to 0.90.
[0166] If the maximum projection similarity Greater than the confidence threshold Then, it is determined that the refrigeration system has experienced the fault type corresponding to the maximum projection similarity, for example, the first... Class of faults;
[0167] If the maximum projection similarity Less than or equal to the confidence threshold If the refrigeration system is in an unstructured environmental noise state, it is determined that the current parameter fluctuations are normal random disturbances and do not trigger fault identification.
[0168] Introducing a confidence threshold effectively constructs a noise firewall, which only diagnoses a fault when the actual operating characteristics highly match a certain theoretical fault fingerprint in the direction, thereby reducing the false alarm rate.
[0169] Example 9:
[0170] Step four includes:
[0171] If the fault type is condenser fouling, execute the flow compensation strategy, calculate the additional cooling water flow to compensate for the decrease in heat transfer coefficient, increase the cooling water flow and keep the chiller setpoint unchanged;
[0172] If the fault type is sensor deviation, execute the parameter locking strategy, lock the current control parameters and disable the automatic adjustment logic, and issue a calibration alarm signal.
[0173] This embodiment is a further specification of step four in embodiment 1, detailing the closed-loop correction strategy for different fault types;
[0174] In step four, the following strategy is executed based on the diagnosed fault type:
[0175] If the fault type is determined to be condenser fouling, the flow compensation calculation stage begins. The specific estimation process is as follows: using the magnitude of the real-time feature residual vector corresponding to the maximum projection similarity obtained in step three. Characterizes the severity of the fault, combined with the preset fouling thermal resistance coefficient. Calculate the current fouling thermal resistance. ;in, It is a preset proportional coefficient obtained based on regression analysis of historical dirt sample data; the heat load to be compensated is calculated based on the thermal resistance series principle. The calculation formula is as follows:
[0176]
[0177] in, To measure the compressor power, To design the heat transfer coefficient; The fouling thermal resistance needs to be divided by the condenser heat exchange area. It is converted into an absolute thermal resistance value for calculation; this formula quantifies the heat transfer attenuation caused by the increase in fouling thermal resistance;
[0178] If the fault type is condenser fouling, i.e., a decrease in condenser heat transfer efficiency, then a flow compensation strategy is implemented; the system calculates the compensation for the heat transfer coefficient based on the principle of heat balance. The reduced additional cooling water flow rate is converted into a volumetric flow rate and added to the current setpoint; the corrected new flow rate. The calculation formula is as follows:
[0179]
[0180] in, This is the current cooling water volumetric flow rate, expressed in cubic meters per second. The heat load to be compensated is expressed in kilowatts, and its value is estimated by the system based on the current fouling thermal resistance. The density of the cooling water is a constant. kilograms per cubic meter; Let be the specific heat capacity of water, and take a value of . kilojoules per kilogram of Kelvin; This represents the temperature difference between the inlet and outlet cooling water, expressed in degrees Celsius or Kelvin.
[0181] Before performing traffic adjustments, the system needs to verify... Is it less than the maximum allowed flow rate by the system? ;like Then adjust the cooling water flow rate to And trigger a system performance limitation alert;
[0182] The system accordingly increases the cooling water flow while keeping the chiller setpoint constant; this strategy is achieved by introducing a density parameter. This ensures the consistency of the physical dimensions of the flow control command, effectively utilizes the enhanced heat exchange effect brought about by increasing the flow rate, and avoids unnecessary increases in compressor power consumption caused by blindly lowering the chiller set point.
[0183] If the fault type is sensor deviation, such as temperature probe drift, a parameter locking strategy is executed; the system locks the current control parameters and disables the automatic adjustment logic, while issuing a calibration alarm signal; this strategy prevents the control system from oscillating due to false data, ensuring the underlying security of the system.
[0184] 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 machine learning-based method for optimizing the energy consumption of a refrigeration system, characterized in that, The specific steps include: Step 1: Collect the wet-bulb temperature, outlet water set temperature and real-time cooling load of the refrigeration system; calculate the ideal condensing temperature and ideal evaporating temperature based on the reverse Carnot cycle principle; calculate the zero entropy loss reference power using the isentropic efficiency of the equipment design; combine the ideal condensing pressure, ideal evaporating pressure and zero entropy loss reference power to generate the ideal state vector. Step 2: Construct a fault characteristic sensitivity matrix using the refrigeration system mechanism model, and use the first-order Taylor series expansion principle to superimpose physical disturbances on the ideal state vector to generate a virtual fault state vector. Step 3: Collect real operating data, calculate the difference ratio between real operating data and ideal state vector to generate real-time feature residual vector, calculate the difference ratio between virtual fault state vector and ideal state vector to generate virtual feature residual vector, calculate the projection similarity of real-time feature residual vector in the direction of virtual feature residual vector, and identify fault type based on projection similarity. Step 4: In response to the fault type, invoke the preset closed-loop correction strategy to adjust the control parameters of the refrigeration system or execute an alarm operation. Step one involves calculating the ideal condensation temperature and the ideal evaporation temperature, including: Obtain the design cooling approach temperature and design evaporator temperature difference from the equipment's manufacturer's performance specifications. The sum of the wet-bulb temperature and the design cooling approach temperature is defined as the ideal condensation temperature. The difference between the set water temperature and the design evaporation temperature is calculated and defined as the ideal evaporation temperature. The steps in step one for generating the ideal state vector also include: Calculate the Carnot theoretical energy efficiency coefficient based on ideal condensation temperature and ideal evaporation temperature; The compression ratio is determined based on the ideal condensing temperature and the ideal evaporating temperature. The compressor performance curve is then retrieved based on the compression ratio to obtain the design isentropic efficiency. The zero-entropy loss baseline power is calculated using real-time cooling load, Carnot theoretical energy efficiency coefficient, and design isentropic efficiency. The saturation pressure corresponding to the ideal condensation temperature is obtained as the ideal condensation pressure. The saturation pressure corresponding to the ideal evaporation temperature is obtained as the ideal evaporation pressure. By combining the ideal condensation pressure, the ideal evaporation pressure, and the zero entropy loss reference power, an ideal state vector is generated. Step two includes: A fault characteristic sensitivity matrix is established based on the numerical solution of the differential equations of the refrigeration system mechanism. The partial derivatives of physical parameters with respect to fault variables are determined using the fault characteristic sensitivity matrix, and a sensitivity Jacobian vector is constructed. Among them, the sensitivity Jacobian vector represents the absolute change in physical parameters caused by a unit fault intensity; An additive perturbation model is used to multiply the sensitivity Jacobian vector by a preset fault evolution intensity and then superimpose it onto the ideal state vector to generate a virtual fault state vector. Step four includes: If the fault type is condenser fouling, execute the flow compensation strategy, calculate the additional cooling water flow to compensate for the decrease in heat transfer coefficient, increase the cooling water flow and keep the chiller setpoint unchanged; If the fault type is sensor deviation, execute the parameter locking strategy, lock the current control parameters and disable the automatic adjustment logic, and issue a calibration alarm signal.
2. The energy consumption optimization method for a refrigeration system based on machine learning according to claim 1, characterized in that: The fault evolution intensity is a numerical sequence generated by discretizing within a preset interval according to a step size. The numerical sequence covers the entire range of faults, from minor to severe. The virtual fault state vector corresponds to the evolution state under different fault mechanisms, where the evolution state includes the condenser fouling state or the abnormal refrigerant charge state.
3. The energy consumption optimization method for a refrigeration system based on machine learning according to claim 1, characterized in that: Step three, which generates the real-time feature residual vector and the virtual feature residual vector, includes the following steps: Calculate the difference between the actual running data and the ideal state vector, divide the difference by the corresponding benchmark value in the ideal state vector, and generate a real-time feature residual vector; Calculate the difference between the virtual fault state vector and the ideal state vector, divide the difference by the corresponding reference value in the ideal state vector, and generate a virtual feature residual vector. Among them, the real-time feature residual vector and the virtual feature residual vector convert the power deviation and pressure deviation into relative percentage change rates.
4. The energy consumption optimization method for a refrigeration system based on machine learning according to claim 3, characterized in that: Step three, which involves calculating projection similarity, includes: The projection similarity of the real-time feature residual vector onto the virtual feature residual vector is calculated using the cosine similarity formula. Among them, projection similarity represents the directional consistency between real-world degradation features and theoretical fault fingerprints.
5. The energy consumption optimization method for a refrigeration system based on machine learning according to claim 4, characterized in that: Step three, identifying the fault type, includes: The maximum value among the projection similarities is selected as the maximum projection similarity. The maximum projection similarity is compared with a preset confidence threshold. If the maximum projection similarity is greater than the confidence threshold, the refrigeration system is determined to have experienced the fault type corresponding to the maximum projection similarity. If the maximum projection similarity is less than or equal to the confidence threshold, the refrigeration system is determined to be in an unstructured environmental noise state and no fault identification is triggered.