Air conditioner energy efficiency diagnosis optimization system and method

Abstract

This invention relates to the technical field of energy efficiency diagnosis, and particularly to an air conditioning energy efficiency diagnosis and optimization system and method. The directed edges of the causal topology graph in the air conditioning energy efficiency diagnosis and optimization method are constrained by the thermodynamic conservation law of air conditioning and prior knowledge of the physical characteristics of components. Diagnostic conclusions with causal chain interpretations are output through counterfactual reasoning. The loss function of the physical information neural network architecture used in digital twin simulation is embedded in the residuals of the partial differential equations of the heat exchanger and the energy conservation equation of the compressor as physical constraint terms. The air conditioning energy efficiency diagnosis and optimization method, constrained by thermodynamic laws and prior knowledge of component physical characteristics, improves the credibility of diagnostic conclusions, enables real-time response under limited computing power, and improves the efficiency of air conditioning energy efficiency diagnosis and optimization. This solves the problem of low reliability of diagnostic results and inability to achieve real-time response under limited computing power in existing air conditioning energy efficiency diagnosis and optimization methods, leading to low efficiency.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency diagnostics, and in particular to an air conditioning energy efficiency diagnostics and optimization system and method. Background Technology

[0002] Air conditioning systems account for approximately 40% to 60% of a building's total energy consumption, making in-service energy efficiency testing and diagnosis a core aspect of energy-saving retrofits. Currently, artificial intelligence (AI) technology is being introduced for data analysis in the diagnosis and optimization of air conditioning energy efficiency, and high-precision digital twin technology is being used to predict the effects of energy efficiency optimization. However, existing AI energy efficiency diagnostic methods typically employ purely data-driven black-box models (such as neural networks and multiple regression), whose diagnostic conclusions cannot provide a clear causal chain or accurately output the specific reasons affecting energy efficiency (e.g., the system outputs "low energy efficiency" but cannot accurately answer whether it's due to a clogged condenser or refrigerant leakage). They may even produce counter-physical diagnostic conclusions (e.g., misjudging normal fluctuations in condensing temperature as compressor failure), resulting in low credibility of AI diagnostic results. Furthermore, existing high-precision digital twin technologies (such as CFD-based simulations) require large computational loads and rely on high-performance cloud servers, making real-time response impossible in the limited computing power of portable testing kits, thus hindering the efficiency of air conditioning energy efficiency diagnosis and optimization. Summary of the Invention

[0003] One objective of this invention is to provide an air conditioner energy efficiency diagnosis and optimization method that, constrained by the laws of thermodynamics and prior knowledge of the physical characteristics of components, improves the reliability of diagnostic conclusions, enables real-time response under limited computing power, and improves the efficiency of air conditioner energy efficiency diagnosis and optimization. This solves the problem that existing air conditioner energy efficiency diagnosis and optimization methods have low reliability of diagnostic results and cannot achieve real-time response under limited computing power, resulting in low efficiency of air conditioner energy efficiency diagnosis and optimization.

[0004] Another objective of this invention is to provide an air conditioning energy efficiency diagnosis and optimization system that outputs diagnostic conclusions that are thermodynamically interpretable and conform to physical laws under small sample conditions, enabling real-time response under limited computing power and improving the efficiency of air conditioning energy efficiency diagnosis and optimization.

[0005] To achieve the above objectives, this invention proposes an air conditioning energy efficiency diagnosis and optimization method, comprising the following steps: collecting real-time multi-dimensional operating data of the air conditioning system; processing the collected multi-dimensional operating data to obtain abnormal feature data; establishing a causal topology graph with the multi-dimensional operating data as nodes, wherein the directed edges of the causal topology graph are constrained by the thermodynamic conservation law of the air conditioning system and prior knowledge of the physical characteristics of the components; performing counterfactual reasoning based on the abnormal feature data and the causal topology graph to output a diagnostic conclusion with causal chain explanation, and proposing candidate optimization strategies based on the causal attribution of the diagnostic conclusion; and performing forward reasoning and simulation verification based on the diagnostic conclusion and the candidate optimization strategies using a physical information neural network architecture, wherein the loss function of the physical information neural network architecture is embedded in the residuals of the partial differential equation of the heat exchanger and the energy conservation equation of the compressor as physical constraint terms.

[0006] Optionally, the types of nodes in the causal topology graph include intervention variables, environmental variables, intermediate physical states, and target variables; wherein, the intervention variables represent operating parameters in the air conditioning system that are actively adjusted by the controller or by human intervention, the environmental variables represent uncontrollable boundary conditions outside the air conditioning system, the intermediate physical states represent measurable or calculable physical quantities of the thermal state of key components in the air conditioning system, and the target variables represent evaluation indicators of the energy efficiency of the air conditioning system.

[0007] Optionally, the data processing of the collected multidimensional operating data to obtain abnormal feature data includes the following steps: cleaning the collected multidimensional operating data, which includes cooling water supply and return temperatures, cooling water inlet and outlet temperatures, and compressor power; calculating the variance based on the cooling water supply and return temperatures, cooling water inlet and outlet temperatures, and compressor power in each sliding window; determining a steady-state operating condition when all variances within the sliding window are lower than the corresponding preset thresholds, and obtaining the corresponding feature data segment under the steady-state operating condition; extracting features from the corresponding feature data segment under the steady-state operating condition to obtain feature vectors, which include energy efficiency feature data, thermal feature data, and operating feature data; and applying sliding windows to the extracted feature vectors respectively. The system is normalized using a standardization method. A three-level threshold discrimination method is used to classify the operating status of the air conditioning system into three levels: normal, concerning, and abnormal, based on the standardized feature vectors belonging to the intermediate physical state of the thermal feature data and the operating feature data. The feature vectors corresponding to the concerning or abnormal operating status are input as abnormal feature data into a pre-trained random forest classifier, which outputs the corresponding air conditioning system operating status, abnormal equipment location, suspected fault type, abnormality level, and energy efficiency loss risk level.

[0008] Optionally, the abnormal feature data belonging to the intermediate physical state type is counterfactually changed to the baseline value, while keeping the environmental variables and other variables that have not experienced anomalies unchanged. The calculation is performed forward along the directed edge direction of the causal topology graph until the target variable in the counterfactual scenario is obtained. The difference between the target variable in the counterfactual scenario and the target variable in the current actual scenario is used as a quantitative indicator to quantify the contribution of the abnormal features to energy efficiency loss, thereby outputting a diagnostic conclusion with causal chain explanation, and proposing candidate optimization strategies based on the causal attribution of the diagnostic conclusion.

[0009] Optionally, the physical information neural network architecture includes an input layer, a frozen layer, an adaptation layer, and an output layer; wherein, the input layer receives the multi-dimensional operating data; the frozen layer is a physical constraint layer, and the network weights of the frozen layer are determined and frozen in a pre-trained state on the physical information neural network architecture before leaving the factory, the pre-training uses the loss function as the optimization objective, and the loss function embeds the residuals of the heat transfer partial differential equation of the heat exchanger and the residuals of the energy conservation equation of the compressor as physical constraint terms; the frozen layer includes four layers, each containing 128 neurons, with the activation function being tanh; the adaptation layer is an online fine-tuning layer, and the weights of the adaptation layer are the identity matrix initialized when the equipment leaves the factory; the adaptation layer includes two layers, each containing 32 neurons, with the activation function being ReLU; the output layer outputs two-dimensional energy efficiency indicators, including the system comprehensive performance coefficient and the compressor input power consumption.

[0010] Optionally, the physical information neural network architecture receiving the diagnostic conclusion and performing forward inference and simulation verification on the candidate optimization strategies further includes an online calibration step: collecting 48-72 hours of on-site operating data of the air conditioning system and extracting a feature matrix X; inputting the feature matrix X into a pre-trained condition generator G(z|X) to output the current degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the air conditioning system, where γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate; freezing the frozen layer, substituting the generated degradation factor vector γ into the loss function to update the baseline physical parameters in the heat exchanger partial differential equation and the compressor energy conservation equation, and fine-tuning the weights of the adaptation layer online: during online fine-tuning, the optimizer uses Adam and the learning rate is set to 1×10. -4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data, as well as the residual of the physical constraint term after updating the parameters. After online fine-tuning, the physical information neural network architecture performs subsequent forward inference and simulation verification based on the loss function after updating the baseline physical parameters and the fine-tuned adaptation layer.

[0011] Optionally, the expression for the loss function is: L total = L data + λ pde · L pde + λ bc · L bc ; Where: λ pde and λ bc These are the preset penalty weights for balancing the physical constraint gradient and the boundary condition gradient; L data This represents a data fitting term, used to penalize the deviation between the physical state quantities and energy efficiency indicators output by the physical information neural network architecture and the real-time acquired multidimensional operating data; L pde The physical constraint term, used to penalize the residuals of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger and the isentropic compression energy conservation equation of the compressor, is expressed as: ;in, This represents the residual term of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger; L represents the residual term of the isentropic compression energy conservation equation for the compressor; bc The boundary condition constraint term is used to penalize the mean square error between the predicted physical state quantity output by the physical information neural network architecture at the set spatial domain boundary and the measurement value collected in real time at the corresponding real physical boundary of the device.

[0012] Optionally, the air conditioning energy efficiency diagnosis and optimization method further includes a missing data processing mechanism. The missing data processing mechanism consists of the following steps: when sensor data is lost during the collection of real-time multidimensional operating data of the air conditioning system, the missing data is reconstructed based on Markov equivalence classes of causal topology. The conditional independence relationship of the missing nodes is determined using Markov equivalence classes. Then, the energy conservation equation of the heat exchanger and the data of the parent and child nodes that have not been lost are used as constraints. The Gibbs sampling algorithm is used to perform maximum a posteriori estimation of the missing physical state quantities, so that the residual of the reconstructed complete dataset that satisfies the energy conservation equation is less than a preset threshold. The number of Gibbs sampling iterations is 1000, of which the first 200 iterations are discarded as an aging period, and the a posteriori estimation is the mean of the remaining samples.

[0013] This invention also proposes an air conditioning energy efficiency diagnosis and optimization system to implement any of the above-described air conditioning energy efficiency diagnosis and optimization methods. The air conditioning energy efficiency diagnosis and optimization system includes: a portable data acquisition terminal for acquiring real-time multi-dimensional operating data of the air conditioning system; an energy efficiency diagnosis engine for processing the acquired multi-dimensional operating data to obtain abnormal feature data; a causal reasoning module, which embeds an air conditioning thermodynamic structure causal model, a counterfactual reasoning engine, and an optimization strategy generation module; the air conditioning thermodynamic structure causal model is used to establish a causal topology graph using the multi-dimensional operating data as nodes; the counterfactual reasoning engine performs counterfactual reasoning based on the abnormal feature data and the causal topology graph, outputting a diagnostic conclusion with causal chain explanation; the optimization strategy generation module proposes candidate optimization strategies based on the causal attribution of the diagnostic conclusion; and a digital twin simulation module, which embeds a physical information neural network architecture for performing forward reasoning and simulation verification based on the diagnostic conclusion and the candidate optimization strategies.

[0014] Optionally, the digital twin simulation module also embeds an online calibration unit, which performs the following steps: collecting 48-72 hours of on-site operating data of the air conditioning system and extracting a feature matrix X; inputting the feature matrix X into a pre-trained condition generator G(z|X) and outputting the degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the current air conditioning system, where γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate; freezing the frozen layer, substituting the generated degradation factor vector γ into the loss function to update the baseline physical parameters in the heat exchanger partial differential equation and the compressor energy conservation equation, and fine-tuning the weights of the adaptation layer online: during online fine-tuning, the optimizer uses Adam and the learning rate is set to 1×10. -4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data, as well as the residual of the physical constraint term after updating the parameters. After online fine-tuning, the physical information neural network architecture performs subsequent forward inference and simulation verification based on the loss function after updating the baseline physical parameters and the fine-tuned adaptation layer.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The air conditioning energy efficiency diagnosis and optimization method of the present invention forms a closed-loop workflow of data acquisition, initial screening diagnosis, detailed diagnosis and attribution, and digital twin simulation. The attribution process is based on the prior knowledge constraints of the thermodynamic conservation law of air conditioning and the physical characteristics of components, ensuring that the output diagnostic conclusions are thermodynamically interpretable. Furthermore, it quantifies the attribution of energy efficiency losses through counterfactual reasoning, thereby accurately calculating the contribution of abnormal characteristic data to energy efficiency losses. The twin simulation process employs a lightweight digital twin based on the PINN architecture. Simultaneously, the residuals of the partial differential equations of the heat exchanger and the energy conservation equation of the compressor are embedded as physical constraint terms into the loss function. This eliminates the need for the PINN architecture to rely on large amounts of data to learn physical laws, significantly reducing the hypothesis space. This allows the air conditioning energy efficiency diagnosis and optimization method of the present invention to conform to physical laws under small sample conditions, facilitating real-time response under limited computing power and improving the efficiency of air conditioning energy efficiency diagnosis and optimization. This solves the problem of low reliability of diagnostic results and inability to achieve real-time response under limited computing power in existing air conditioning energy efficiency diagnosis and optimization methods, leading to low efficiency in air conditioning energy efficiency diagnosis and optimization.

[0016] 2. The energy efficiency diagnosis engine, causal reasoning module, and digital twin simulation module of the air conditioner energy efficiency diagnosis and optimization system of the present invention form a progressive collaborative workflow. The causal reasoning module embeds prior knowledge constraints of thermodynamic constraints and component physical characteristic constraints. The physical information neural network architecture adopts the partial differential equation of the air conditioner heat exchanger and the energy conservation equation of the compressor as constraints, thereby realizing an interpretable, anti-physics reasoning immune, edge computing power adaptable, and small sample robust energy efficiency diagnosis and optimization closed loop. Attached Figure Description

[0017] Figure 1 This is a flowchart of an embodiment of an air conditioner energy efficiency diagnosis and optimization method; Figure 2 A schematic diagram of the causal topology of an embodiment of the air conditioner energy efficiency diagnosis and optimization method of the present invention (each node in the causal topology diagram only shows some feature data, the arrows represent causal chain relationships, and all arrow directions are constrained by the prior knowledge of the thermodynamic conservation law of air conditioner and the physical properties of components). Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0020] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a fixed connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal connection of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the word "and / or" throughout the text means including three parallel solutions; taking "A and / or B" as an example, it includes solution A, solution B, or a solution that simultaneously satisfies A and B. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0022] To address the aforementioned technical problems, this invention proposes a method for diagnosing and optimizing air conditioning energy efficiency, such as... Figure 1 As shown, it includes the following steps: Collect real-time, multi-dimensional operational data of the air conditioning system; The collected multidimensional operational data is processed to obtain abnormal feature data; A causal topology graph is established using the multidimensional operational data as nodes. The directed edges of the causal topology graph are constrained by the thermodynamic conservation law of air conditioning and prior knowledge of the physical properties of the components. Based on the abnormal feature data and the causal topology graph, counterfactual reasoning is performed to output a diagnostic conclusion with causal chain explanation, and candidate optimization strategies are proposed based on the causal attribution of the diagnostic conclusion. The physical information neural network architecture performs forward inference and simulation verification based on the diagnostic conclusions and the candidate optimization strategies. The loss function of the physical information neural network architecture is embedded in the residuals of the partial differential equations of the heat exchanger and the energy conservation equation of the compressor as physical constraint terms.

[0023] The air conditioning energy efficiency diagnosis and optimization method of this invention forms a closed-loop workflow of data acquisition, initial screening diagnosis, detailed diagnosis and attribution, and digital twin simulation. The attribution process is based on prior knowledge of the thermodynamic conservation law of air conditioning and the physical characteristics of components, and uses counterfactual reasoning to ensure the thermodynamic interpretability of the output diagnostic conclusions. The twin simulation process employs a lightweight digital twin based on a Physical Information Neural Network (PINN) architecture. Simultaneously, the residuals of the partial differential equations of the air conditioning heat exchanger and the energy conservation equation of the compressor are embedded as physical constraint terms into the loss function. This eliminates the need for the PINN architecture to rely on large amounts of data to learn physical laws, reducing the hypothesis space. This allows the air conditioning energy efficiency diagnosis and optimization method of this invention to conform to physical laws under small sample conditions (limited on-site data collection conditions) and helps achieve real-time response in environments with limited computing power, thereby improving the efficiency of air conditioning energy efficiency diagnosis and optimization.

[0024] The diagnostic conclusions output by the air conditioner energy efficiency diagnosis and optimization method of this invention follow a causal chain of "because A→B→C, energy efficiency is reduced," with each step having a clear physical meaning. The construction of the directed edges of the causal topology graph is constrained by prior knowledge of thermodynamic laws and component physical characteristics, fundamentally eliminating absurd diagnostic conclusions (such as anti-physics diagnostic conclusions) that may be generated by purely data-driven models, and significantly improving the credibility of the diagnostic conclusions. At the same time, a lightweight PINN architecture and a loss function with embedded physical constraints are constructed for simulation verification, which can achieve real-time response under limited computing power, improving the efficiency of air conditioner energy efficiency diagnosis and optimization. This solves the problem of low reliability of diagnostic results and inability to achieve real-time response under limited computing power in existing air conditioner energy efficiency diagnosis and optimization methods, resulting in low efficiency of air conditioner energy efficiency diagnosis and optimization.

[0025] like Figure 2 As shown, to further illustrate, the types of nodes in the causal topology graph include intervention variables, environmental variables, intermediate physical states, and target variables; wherein, the intervention variables represent operating parameters in the air conditioning system that are actively adjusted by the controller or by human intervention, the environmental variables represent uncontrollable boundary conditions outside the air conditioning system, the intermediate physical states represent measurable or calculable physical quantities of the thermal state of key components in the air conditioning system, and the target variables represent evaluation indicators of the energy efficiency of the air conditioning system.

[0026] The following are examples illustrating the intervention variables, which include the opening degree of the electronic expansion valve, the operating frequency of the chilled water pump, and the speed of the cooling tower fan; the environmental variables include the outdoor dry-bulb / wet-bulb temperature and the indoor air conditioning load; the intermediate physical states include the condenser terminal temperature difference, the evaporator superheat, the compressor pressure ratio, the evaporation temperature, the condensation temperature, the suction pressure and the discharge pressure, the chilled water supply and return water temperature difference, the cooling water inlet and outlet water temperature difference, and the equipment load rate; the target variables include the system comprehensive performance coefficient (COP) and the compressor input power.

[0027] The directed edges of the causal topological graph are constrained by the thermodynamic conservation law of air conditioning and prior knowledge of the physical properties of its components, specifically including: Mass conservation constraints, for example, the refrigerant mass flow rate under steady state satisfies the following: evaporator flow rate = compressor suction flow rate = condenser flow rate = expansion valve discharge rate; Energy conservation constraints, for example, condenser heat dissipation = evaporator heat absorption + compressor input power; Component physical characteristics are constrained. For example, an increase in condenser terminal temperature difference will inevitably lead to an increase in condensing temperature, which in turn will lead to an increase in compressor pressure ratio and power consumption. The causal topology diagram only allows a unidirectional causal chain of "condenser terminal temperature difference → condensing temperature → compressor pressure ratio → compressor power consumption".

[0028] Anti-physical causal edges are prohibited, and it is explicitly restricted that anti-physical edges pointing from the target variable or intermediate physical state to the environmental variable are not allowed (such as "compressor power consumption increases → outdoor temperature rises").

[0029] To further explain, the data processing of the collected multidimensional operational data to obtain abnormal feature data includes the following steps: The collected multidimensional operating data is cleaned, and the multidimensional operating data includes cooling water supply and return water temperature, cooling water inlet and outlet water temperature, and compressor power. The variance is calculated based on the cooling water supply and return water temperature, cooling water inlet and outlet water temperature, and compressor power in each sliding window. When all the variances in the sliding window are lower than the corresponding preset threshold, it is determined to be a steady-state operating condition, and the corresponding feature data segment under the steady-state operating condition is obtained. Feature extraction is performed on the corresponding feature data segments under the steady-state operating conditions to obtain feature vectors. The feature vectors include energy efficiency feature data, thermal feature data, and operational feature data. Sliding windows are applied to the extracted feature vectors respectively. Standardization methods are used for normalization. The thermal feature data and operational feature data belonging to the intermediate physical state in the standardized feature vector are classified into three levels of normal, concern, and abnormal using a three-level threshold discrimination method. The feature vector corresponding to the operational state of concern or abnormal is used as abnormal feature data and input into a pre-trained random forest classifier to output the corresponding operational state of the air conditioning system, abnormal equipment location, suspected fault type, abnormal level, and energy efficiency loss risk level.

[0030] To further explain, the data cleaning process for the collected multidimensional operational data includes the following steps: amplitude limiting filtering of the collected multidimensional operational data, using sliding window mid-range filtering to suppress impulse noise, and marking missing values. The purpose of amplitude limiting filtering is to ensure that the collected multidimensional operational data is within a reasonable range, such as chilled water temperature of 0~20℃, cooling water temperature of 15~45℃, power of 0~rated value × 1.2, etc. The window width of the sliding window mid-range filtering is 5 sampling points. Missing values ​​include values ​​that were missing during the original acquisition as well as obviously distorted values ​​that were removed after amplitude limiting filtering. To prevent system crashes during subsequent calculations and to prevent spurious fitting from affecting the consistency of physical constraints, the system can trigger a missing data processing mechanism to reconstruct the missing data for the marked missing values, so that the reconstructed complete dataset can participate in subsequent energy efficiency assessment calculations.

[0031] To further explain, the sliding window in the steady-state condition identification step has a window width of 60 sampling points, corresponding to a 5-minute time window. When all variances within the sliding window are below the corresponding preset threshold, it is determined to be a steady-state condition, and the corresponding feature data segment under the steady-state condition can be obtained. When one of the variances in the sliding window is above the corresponding preset threshold, it is determined to be a transitional condition, and the transitional data segment corresponding to the transitional condition is marked and not included in the subsequent energy efficiency assessment calculation.

[0032] Furthermore, the feature data corresponding to the steady-state operating condition corresponds to the intervention variable, the environmental variable, the intermediate physical state, or the target variable in the causal topology graph. Further, the feature vector obtained by feature extraction from the feature data segment corresponding to the steady-state operating condition is a complete 12-dimensional feature vector, specifically divided into energy efficiency feature data, thermal feature data, and operational feature data. The energy efficiency feature data corresponds to the target variable in the causal topology graph, specifically including the system comprehensive performance coefficient (COP) and compressor input power. The thermal feature data and the operational feature data correspond to the intermediate physical state in the causal topology graph. The thermal feature data specifically includes condenser terminal temperature difference, evaporator superheat, compressor pressure ratio, evaporation temperature, condensation temperature, suction pressure, and discharge pressure. The operational feature data specifically includes chilled water supply and return temperature difference, cooling water inlet and outlet temperature difference, and equipment load rate. Furthermore, the extracted abnormal feature data is also constrained by the thermodynamic conservation law of air conditioning and the prior knowledge of the physical properties of components. In the counterfactual reasoning stage, the system counterfactually changes the extracted abnormal feature data belonging to the intermediate physical state to the corresponding physical baseline value, while keeping the environmental variables and other variables that have not experienced anomalies unchanged. Forward propagation calculation is performed along the directed edges of the causal topology graph to obtain the target variable in the counterfactual scenario.

[0033] The physical principle behind anomaly threshold discrimination lies in identifying the specific physical causes of changes in the target variable (energy efficiency) by monitoring deviations in parameters belonging to the intermediate physical state. Including the target variable itself in threshold discrimination would constitute a circular reasoning in causal logic. Therefore, the data used for threshold discrimination should be limited to thermodynamic and operational characteristic data belonging to the intermediate physical state.

[0034] Eigenvalues ​​after standardization x norm The calculation formula is: x norm =( x - μ window ) / σ window , x This represents the original characteristic values ​​(i.e., the energy efficiency characteristic data, the thermal characteristic data, and the operational characteristic data). μ window This represents the mean of the feature within the sliding window. σ window This represents the characteristic standard deviation within the sliding window.

[0035] To further explain, in the step of classifying the operating status of the air conditioning system into three levels—normal, concerning, and abnormal—using a three-level threshold discrimination method on the standardized feature vector, the design value (the design value set according to the actual situation or the nameplate parameter as the design value), the historical mean ± 3σ (standard deviation), and the industry standard limit are used as the three-level threshold discrimination criteria. In this embodiment, it generally refers to comparing the feature vector belonging to the intermediate physical state with the three-level threshold discrimination criteria to classify the operating status of the air conditioning system into three levels: normal, concerning, and abnormal. (Specifically, when the feature vector is close to the design value and within the historical mean ± 3σ range, the operating status is determined to be normal; when the feature vector exceeds the historical mean ± 3σ range but is lower than the industry standard limit, the operating status is determined to be concerning; when the feature vector exceeds the industry standard limit, the operating status is determined to be abnormal.) (For cases deemed abnormal); the training data of the pre-trained random forest classifier includes normal operating conditions, typical fault conditions, and accelerated aging data simulated in the laboratory, covering fault types such as refrigerant leakage, condenser blockage, and compressor efficiency decline. The 12-dimensional standardized feature vector (specifically including the system comprehensive performance coefficient COP, compressor input power, condenser terminal temperature difference, evaporator superheat, compressor pressure ratio, evaporation temperature, condensation temperature, suction pressure and discharge pressure, chilled water supply and return temperature difference, cooling water inlet and outlet temperature difference, and equipment load rate) serves as the input to the random forest classifier. The classifier outputs fault type labels and corresponding confidence scores, thereby enabling precise discrimination of the abnormal feature data corresponding to the operating state being of concern or abnormal, and outputting the corresponding system operating state, abnormal equipment location, suspected fault type, abnormality level, and energy efficiency loss risk level.

[0036] To further explain, the multi-dimensional operating data of the air conditioning system collected in real time includes the system comprehensive performance coefficient (COP), compressor input power, condenser terminal temperature difference, evaporator superheat, compressor pressure ratio, evaporation temperature, condensation temperature, suction pressure and discharge pressure, chilled water supply and return water temperature difference, cooling water inlet and outlet water temperature difference, and equipment load rate, etc.

[0037] To further explain, the abnormal feature data belonging to the intermediate physical state type is counterfactually changed to the baseline value, while keeping the environmental variables and other variables that have not experienced anomalies unchanged. The calculation is performed forward along the directed edge direction of the causal topology graph until the target variable in the counterfactual scenario is obtained. The difference between the target variable in the counterfactual scenario and the target variable in the current actual scenario is used as a quantitative indicator to quantify the contribution of the abnormal features to energy efficiency loss, thereby outputting a diagnostic conclusion with causal chain explanation, and proposing candidate optimization strategies based on the causal attribution of the diagnostic conclusion.

[0038] To further explain, the abnormal feature data belonging to the intermediate physical state is counterfactually changed to a baseline value. The baseline value is generally a threshold (such as the corresponding design value, historical statistical average, or nameplate parameter, etc.) or optimized value corresponding to the normal operating state of the air conditioning system. At the same time, the environmental variables and other variables that have not experienced abnormalities remain unchanged. The calculation is performed by forward propagation along the directed edge direction of the causal topology graph to update the values ​​of the downstream intermediate physical state and the target variable until the target variable (the comprehensive performance coefficient of the air conditioning system COP or the compressor input power) in the counterfactual scenario is calculated. The difference between the target variable value in the counterfactual scenario and the target variable in the current actual scenario is used as a quantitative indicator to quantify the contribution of the abnormal feature data to energy efficiency loss. Thus, a diagnostic conclusion with causal chain explanation can be output based on the abnormal feature data.

[0039] To further explain, the physical information neural network architecture includes an input layer, a freeze layer, an adaptation layer, and an output layer; The input layer receives the multidimensional operational data collected in real time. The frozen layer is a physical constraint layer. The network weights of the frozen layer are determined and kept frozen before leaving the factory by performing pre-training on the physical information neural network architecture. The pre-training uses the loss function as the optimization objective. The loss function is embedded with the residuals of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger and the residuals of the isentropic compression energy conservation equation of the compressor as physical constraint terms. The frozen layer consists of four layers, each containing 128 neurons, with the activation function being tanh; The adaptation layer is an online fine-tuning layer, and the weights of the adaptation layer are the identity matrix initialized when the device leaves the factory; the adaptation layer consists of two layers, each containing 32 neurons, and the activation function is ReLU; The output layer outputs two-dimensional energy efficiency indicators, including the system comprehensive performance coefficient and the compressor input power consumption.

[0040] Specifically, the input layer receives 12 dimensions of operational data, including energy efficiency characteristic data, thermal characteristic data, and operational characteristic data. The energy efficiency characteristic data corresponds to the target variable in the causal topology graph, specifically including the system comprehensive performance coefficient (COP) and compressor input power. The thermal characteristic data and operational characteristic data correspond to the intermediate physical states in the causal topology graph. The thermal characteristic data specifically includes condenser terminal temperature difference, evaporator superheat, compressor pressure ratio, evaporation temperature, condensation temperature, suction pressure, and discharge pressure. The operational characteristic data specifically includes chilled water supply and return temperature difference, cooling water inlet and outlet temperature difference, and equipment load rate.

[0041] The frozen layer is a physical constraint layer (i.e., a physical constraint layer containing the residuals of the nonlinear one-dimensional transient heat transfer partial differential equation embedded in the air conditioner heat exchanger and the residuals of the compressor isentropic compression energy conservation equation). The network weights of the frozen layer are determined through pre-training before leaving the factory and kept frozen. The pre-training dataset contains simulation data of at least 100,000 different operating conditions (covering different levels of fouling, refrigerant charge, and ambient temperature combinations). The training objective is to minimize the error between the predicted COP and the simulation true value. At the same time, the residuals of the heat exchanger heat transfer partial differential equation and the compressor energy conservation equation are added as regularization terms to the loss function (the loss function is the overall optimization objective of the pre-training process, calculated after forward propagation to the output layer, and then updated by backpropagation through gradient calculation). The feature representation learned by the frozen layer is constrained by the thermodynamic physical equations (heat exchanger heat transfer equation and energy conservation equation), so that its output features satisfy basic physical consistency, rather than purely data-driven free fitting. The physical constraints of the frozen layer and the "directed edges of the causal topology graph constrained by the prior knowledge of the thermodynamic conservation law and component physical properties of the air conditioner" (mass conservation, energy conservation, component property constraints, etc.) both belong to thermodynamic prior knowledge, but their levels of action are different. The thermodynamic conservation law constraint is a constraint at the structural level of the causal topology graph, determining whether the causal relationship between variables holds; while the physical constraints of the frozen layer are constraints at the feature level of the neural network, determining whether the numerical values ​​output by the physical information neural network architecture model conform to the physical equations. The two complement each other, jointly ensuring the physical consistency of the entire system from causal logic to numerical prediction. The adaptation layer is an online fine-tuning layer. The weights of the adaptation layer are initialized to an identity matrix and can be updated online after collecting multidimensional operating data for a limited period of time on-site.

[0042] Furthermore, the physical information neural network architecture, which receives the diagnostic conclusion and performs forward inference and simulation verification on the candidate optimization strategies, also includes an online calibration step: Collect 48-72 hours of on-site operating data of the air conditioning system and extract the feature matrix X; Input the feature matrix X into the trained condition generator G(z|X) and output the degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the current air conditioning system, where: γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate. On an edge device (such as a portable energy efficiency testing kit), the frozen layer is frozen, and the generated degradation factor vector γ is substituted into the loss function to update the baseline physical parameters in the heat exchanger partial differential equation and the compressor energy conservation equation. The weights of the adaptation layer are then fine-tuned online: the optimizer used during online fine-tuning is Adam, and the learning rate is set to 1×10⁻⁶.-4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data, as well as the residual of the physical constraint term after updating the parameters. After online fine-tuning, the physical information neural network architecture performs subsequent forward inference and simulation verification based on the loss function after updating the baseline physical parameters and the fine-tuned adaptation layer.

[0043] By freezing the frozen layer (physical constraint layer) and only fine-tuning the last two adaptation layers, the physical information neural network architecture model can adapt to the aging parameters of the field equipment within seconds, thereby achieving rapid adaptation to equipment aging drift and avoiding the huge computing power overhead of retraining in the cloud.

[0044] To further explain, the expression for the loss function is: L total = L data + λ pde · L pde + λ bc · L bc ; Where: λ pde and λ bc These are preset penalty weight coefficients for balancing the physical constraint gradient and the boundary condition gradient, respectively, used to determine the magnitude of the data fitting gradient and the physical equation residual gradient in the balanced loss function, and are pre-defined hyperparameters.

[0045] L data This represents a data fitting term used to penalize the deviation between the physical state quantities and energy efficiency indicators output by the physical information neural network architecture and the multidimensional operating data collected in real time. Specifically, the penalty for the data fitting term covers both the "energy efficiency index" output by the physical information neural network architecture and the "physical state quantity" predicted to satisfy the calculation of the physical equation.

[0046] L pde The physical constraint term, used to penalize the residuals of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger and the isentropic compression energy conservation equation of the compressor, is expressed as: in, This represents the residual term of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger; This represents the residual term in the isentropic compression energy conservation equation of the compressor. The expression for the isentropic compression energy conservation equation of the compressor is: In the formula, W is the compressor input power. This is the refrigerant mass flow rate. The specific enthalpy of inhalation. The actual exhaust specific enthalpy, and it satisfies the isentropic compression relationship. ,in The isentropic exhaust enthalpy, The isentropic efficiency of the compressor. During forward inference, the physical information neural network architecture uses the square of the difference between the two sides of the energy conservation equation as the residual term. Perform calculations; L bc The boundary condition constraint term is used to penalize the mean square error between the predicted physical state quantity (such as refrigerant temperature or pressure) output by the physical information neural network architecture at the set spatial domain boundary (e.g., the start and end points of the heat exchanger tube length) and the measured value collected in real time by the sensor installed at the corresponding real physical boundary of the equipment (i.e., the inlet / outlet, air inlet / exhaust port of the actual heat exchanger).

[0047] To further explain, the partial differential equation of an air conditioner heat exchanger is: T / t + v· T / x - α· ²T / x² = - (UA / (ρcV))·(T - T_coolant); The parameters in this equation are defined and obtained as follows: T: Refrigerant temperature (K or °C), representing the physical field to be solved.

[0048] t: time (s), is the independent variable.

[0049] x: Spatial coordinate (m) along the length of the heat exchanger tube, which is the independent variable.

[0050] v: Refrigerant flow rate (m / s), calculated by dividing the volumetric flow rate Q measured by the on-site flow sensor by the cross-sectional area A_c of the flow channel, i.e., v = Q / A_c.

[0051] α: Thermal diffusivity (m² / s), taken from the refrigerant's physical property parameters at the average temperature of the heat exchanger, and pre-stored in the database of the testing host.

[0052] UA: The product of heat transfer coefficient and heat transfer area (W / K); it should be noted that during the online calibration phase, UA is obtained by multiplying the output degradation factor γ by the reference UA0, where UA0 is taken from the equipment nameplate parameters or design values. Furthermore, if no online calibration step is performed, then UA = UA0.

[0053] ρ: Refrigerant density (kg / m³), taken from the physical property parameter table.

[0054] c: Specific heat capacity of refrigerant (J / (kg·K)), taken from the physical property parameter table.

[0055] V: Volume of the heat exchanger control body (m³), determined by the heat exchanger's geometric parameters.

[0056] T_coolant: Temperature (K or °C) of the cooling medium (water or air), measured by a field temperature sensor.

[0057] Furthermore, the air conditioning energy efficiency diagnosis and optimization method also includes a missing data processing mechanism. The steps of the missing data processing mechanism are as follows: when sensor data loss occurs during the collection of real-time multi-dimensional operating data of the air conditioning system, the missing data is reconstructed based on Markov equivalence classes of causal topology. The conditional independence relationship of missing nodes is determined by using Markov equivalence classes. Then, the energy conservation equation of the heat exchanger and the data of the parent and child nodes that have not been lost are used as constraints. The Gibbs sampling algorithm is used to perform maximum a posteriori estimation of the missing physical state variables, so that the residual of the reconstructed complete dataset that satisfies the energy conservation equation is less than a preset threshold. The number of Gibbs sampling iterations is 1000, of which the first 200 are discarded as an aging period, and the a posteriori estimation is the mean of the remaining samples.

[0058] Currently, in field testing of air conditioning systems, sporadic sensor malfunctions leading to data loss are quite common. Existing air conditioning energy efficiency diagnosis and optimization methods lack effective missing data reconstruction mechanisms, often resulting in interruptions in the diagnostic process or incorrect conclusions. To further explain, the system uses the marking of missing values ​​during the data cleaning step as a trigger condition to initiate Markov equivalence class reconstruction based on the causal topology graph structure. After reconstructing the missing values ​​into a complete dataset, this complete dataset is used as input in subsequent energy efficiency assessment calculations.

[0059] The present invention establishes a missing data reconstruction mechanism based on Markov equivalence classes of the causal topology graph structure. The Markov inference mechanism based on physical constraints uses the energy conservation equation as a priori constraint for probabilistic inference, so that a high diagnostic accuracy can still be maintained when a single sensor fails, thereby achieving robust diagnosis under sensor failure.

[0060] Furthermore, the air conditioner energy efficiency diagnosis and optimization method also includes the following steps: based on the diagnostic conclusion and simulation verification results, an optimization strategy scheme is generated. The optimization strategy scheme can be visualized in the form of a cause-effect graph to present the attribution chain of energy efficiency problems. The optimization strategy scheme may include control parameter adjustment suggestions and / or hardware modification suggestions.

[0061] This invention also proposes an air conditioner energy efficiency diagnosis and optimization system for implementing any of the air conditioner energy efficiency diagnosis and optimization methods described above, wherein the air conditioner energy efficiency diagnosis and optimization system includes: A portable data acquisition terminal, used to collect real-time multi-dimensional operating data of an air conditioning system; An energy efficiency diagnostic engine is used to process the collected multidimensional operating data and obtain abnormal feature data. The causal reasoning module embeds an air conditioning thermodynamic structure causal model (SCM), a counterfactual reasoning engine, and an optimization strategy generation module. The air conditioning thermodynamic structure causal model is used to establish a causal topology graph using the multidimensional operational data as nodes. The counterfactual reasoning engine performs counterfactual reasoning based on the abnormal feature data and the causal topology graph, and outputs a diagnostic conclusion with causal chain explanation. The optimization strategy generation module is used to propose candidate optimization strategies based on the causal attribution of the diagnostic conclusion. A digital twin simulation module, which embeds a physical information neural network architecture, is used for forward inference and simulation verification based on the diagnostic conclusion and the candidate optimization strategies.

[0062] The air conditioning energy efficiency diagnosis and optimization system of the present invention forms a closed-loop workflow of data acquisition, initial screening diagnosis, detailed diagnosis and attribution, and digital twin construction and pre-verification. After the energy efficiency diagnosis engine completes the initial screening diagnosis, it transmits the abnormal feature data to the causal reasoning module for deep attribution. The causal reasoning module outputs the diagnostic conclusion (root cause) and candidate optimization strategy suggestions. The digital twin simulation module performs simulation verification on the candidate optimization strategy suggestions based on the diagnostic conclusion and outputs the simulation verification results. The diagnostic conclusion output by the air conditioning energy efficiency diagnosis and optimization system of the present invention has thermodynamic interpretability and can conform to physical laws under small sample conditions, which helps to achieve real-time response in a limited computing power environment and improves the efficiency of air conditioning energy efficiency diagnosis and optimization.

[0063] Furthermore, the causal reasoning module also includes a missing data processing module. This module is used to reconstruct missing data based on the Markov equivalence classes of the causal topology graph when sensor data is lost during the collection of real-time multidimensional operating data of the air conditioning system. A specific example is provided below for illustration: Assuming the condenser outlet temperature sensor experiences data loss, the parent nodes of this node in the cause-effect graph are the condenser inlet temperature, cooling water flow rate, and refrigerant flow rate; the child nodes are the compressor power consumption and system COP. After determining the conditional independence relationship between this node and other nodes using Markov equivalence classes, the Gibbs sampling algorithm is used for maximum a posteriori estimation. Initialization: Based on the current values ​​of condenser inlet temperature, cooling water flow rate, and refrigerant flow rate, an estimated value T_co_out^(0) is randomly initialized according to the historical data of the condenser outlet temperature distribution under this operating condition.

[0064] Iterative sampling: For the k-th iteration, samples are sequentially taken from the conditional probability distribution P(T_co_out | current values ​​of all other nodes), where the measured values ​​are taken for non-lost nodes and the current estimated values ​​are taken for lost nodes. The conditional probability distribution is obtained through Bayesian network parameter learning, and the prior distribution is set to a Gaussian distribution.

[0065] Physical constraints: After each sampling, check whether the energy conservation equation holds (i.e., the heat dissipation of the condenser ≈ the heat released by the refrigerant). If the residual is greater than the threshold ε=0.01, reject the current sampling and resample.

[0066] Convergence judgment: After 1000 iterations, take the mean of the posterior samples as the final estimate, and discard the first 200 iterations as the aging period.

[0067] Using the above method, the reconstructed complete dataset satisfies a thermodynamic consistency error of less than 5%.

[0068] Furthermore, the digital twin simulation module also includes an online calibration unit, which is used to implement the following steps: Collect 48-72 hours of on-site operating data of the air conditioning system and extract the feature matrix X; Input the feature matrix X into the trained condition generator G(z|X) and output the degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the current air conditioning system, where: γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate. On an edge device (such as a portable energy efficiency testing kit), the frozen layer is frozen, and the weights of the adaptation layer are fine-tuned online. During online fine-tuning, the optimizer is Adam, and the learning rate is set to 1×10. -4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data. After online fine-tuning is completed, the generated degradation factor γ is input as an additional condition variable into the physical information neural network architecture for subsequent forward inference and simulation verification.

[0069] To further explain, the online calibration unit includes a condition generator and a discriminator. During the factory training phase, the discriminator D(γ,X) employs a three-layer fully connected network, taking the degradation factor γ (dimension 3) and the feature matrix X (dimension 12) as inputs, and outputting a matching score between 0 and 1, used to determine whether the generated γ matches the current operating state. The condition generator G(z|X) takes random noise z and conditional features X as inputs and outputs a degradation factor vector γ. The condition generator and the discriminator are alternately updated through adversarial training. The training data comes from the historical normal operation and maintenance data of this model of air conditioning equipment (containing more than 500 sets of sensor sequences under different operating conditions and their corresponding real degradation factors, which are obtained through laboratory aging tests). After the edge device is deployed, only the generator G(z|X) is retained to output the degradation factor, and the discriminator is no longer used to reduce computational overhead.

[0070] Furthermore, the optimization strategy generation module can also be used to output a final optimization strategy based on the diagnostic conclusion (i.e., the result of counterfactual reasoning) and the simulation verification result (i.e., the forward inference result implemented by the physical information neural network architecture and the loss function in the digital twin simulation module). The optimization strategy may include control parameter adjustment schemes or hardware modification suggestions.

[0071] The air conditioning energy efficiency diagnosis and optimization system also includes a human-computer interaction and visualization module, which is used to visualize the attribution chain of energy efficiency problems and expected energy-saving benefits in the form of a cause-and-effect diagram.

[0072] To further explain, after implementing the optimization strategy generated by the optimization strategy generation module, the optimized actual operating data can be fed back to the online calibration unit, thereby achieving continuous adaptive optimization.

[0073] The energy efficiency diagnostic engine, the causal reasoning module, and the digital twin simulation module of this invention form a progressive collaborative workflow. The causal reasoning module embeds prior knowledge constraints based on thermodynamic constraints and component physical characteristic constraints. The physical information neural network architecture employs partial differential equations for air conditioner heat exchangers and energy conservation equations for compressors, thereby achieving an interpretable, anti-physics reasoning-immune, edge computing-adaptable, and small-sample-robust energy efficiency diagnostic and optimization closed loop. More specifically, the energy efficiency diagnostic engine receives the collected multi-dimensional operational data and determines "which structure or component in the air conditioning system may have an anomaly" using a three-level threshold discrimination method and a random forest classification method. The causal reasoning module then uses the energy efficiency diagnostic... The abnormal feature data output by the engine is used as input. Based on prior knowledge constraints such as thermodynamic constraints and component physical characteristic constraints, the engine outputs diagnostic conclusions that follow a causal chain through counterfactual reasoning, and quantifies their contribution to energy efficiency loss, thus solving the problem of "why the abnormality occurs and how much the abnormality contributes to energy efficiency loss" in the air conditioning system. The digital twin simulation module, by constructing the lightweight PINN architecture and the online calibration unit, performs forward reasoning and simulation verification on the diagnostic conclusions and candidate optimization strategies output by the causal reasoning module, and then outputs the expected energy efficiency improvement effect and energy saving benefit prediction. The operating data after implementing the corresponding optimization strategy can be fed back to the online calibration unit to achieve continuous adaptive optimization.

[0074] The present invention will be further illustrated below with reference to two specific embodiments, but the scope of protection of the present invention is not limited thereto: In one embodiment, taking a commercial complex air conditioning system as the engineering object, the specific working methods of the energy efficiency diagnostic engine, causal reasoning module, and digital twin simulation module are described: The commercial complex has a building area of ​​65,000 m², and its centralized air conditioning system is configured as follows: 3 centrifugal chiller units (#1, #2, and #3, each with a rated cooling capacity of 950 RT and a rated COP of 6.2, all of which have been in operation for 7 years), 4 chilled water pumps (variable frequency control), 4 cooling water pumps, and 2 sets of square crossflow cooling towers. The system is a primary pump variable flow system, with chilled water supply and return design temperatures of 7℃ / 12℃ and cooling water inlet and outlet design temperatures of 32℃ / 37℃. Normal operating hours are from 8:00 to 22:00, and the peak daily cooling load in summer is approximately 5200 kW.

[0075] At 10:00 AM on a typical summer workday, with an outdoor temperature of 33℃ and relative humidity of 65%, the building's cooling load is approximately 3800kW (approximately 73% load rate). The portable data acquisition terminal described above is deployed on-site, with a sampling period of 5 seconds, to acquire the following data in real time via a wireless sensor network: chilled water supply and return temperatures (Unit #1: 6.8℃ / 11.5℃; Unit #2: 7.2℃ / 12.8℃; Unit #3: 7.5℃ / 12.2℃), cooling water inlet and outlet temperatures (Unit #1: 33.5℃ / 39.2℃; Unit #2: 34.2℃ / 41.8℃; Unit #3: 33.8℃ / 40.5℃), and chilled water flow rate (Unit #1: 420 m³ / h). 3 / h; #2: 180 m 3 / h; #3: 280 m 3 / h), compressor power (#1: 510 kW; #2: 320 kW; #3: 440 kW), operating frequency of each device, and environmental parameters.

[0076] After receiving the aforementioned operating data, the energy efficiency diagnostic engine first performs data cleaning: In the amplitude limiting filtering stage, the chilled water temperature (6.8~7.5℃), cooling water temperature (33.5~41.8℃), and compressor power (320~510 kW) are all within their respective reasonable ranges, with no outliers. A sliding window midpoint filter (window width of 5 sampling points) is used to process measurement noise in the temperature signal, while simultaneously marking missing data points (in this embodiment, communication is normal, and there are no missing data points). Subsequently, a sliding window variance discrimination method (window width of 60 sampling points, corresponding to a 5-minute time window) is used to analyze the data from Unit #2: the variances of the chilled water supply and return water temperature windows are 0.03 and 0.05, respectively; the variances of the cooling water inlet and outlet water temperature windows are 0.07 and 0.09, respectively; and the variance of the compressor power window is 0.12, all below the preset steady-state threshold. This time period is determined to be a steady-state operating condition, and the corresponding steady-state data segment is extracted for subsequent energy efficiency assessment calculations.

[0077] Entering the feature extraction and standardization stage, the energy efficiency diagnostic engine calculates the characteristic parameters of Unit #2 according to the formula: Condenser terminal temperature difference = condenser temperature (corresponding to a saturation temperature of 40.2℃ based on an exhaust pressure of 40.5℃) - cooling water outlet temperature of 41.8℃ = -1.6℃. The calculated value is negative, meaning the condenser temperature is lower than the cooling water outlet temperature—which is physically impossible and is caused by a time synchronization issue between the cooling water outlet temperature sensor and the exhaust pressure sensor during data acquisition. The energy efficiency diagnostic engine triggers a fluctuation check of the values ​​at each measuring point within the steady-state window, confirming that the standard deviation of each measuring point during this period is less than the measurement noise threshold, indicating that the sensor accuracy is normal. Subsequently, the cooling water flow data is checked, revealing that the cooling water inlet temperature of Unit #2 is 34.2℃, the outlet temperature is 41.8℃, and the flow rate is 240 m³ / h, with a calculated heat dissipation of approximately 2120kW, while the unit's cooling capacity is approximately 1050kW, indicating a heat imbalance and abnormal heat exchange at the condenser end. After ruling out sensor time synchronization issues, the preliminary assessment is that the condenser's heat exchange capacity has significantly decreased; the calculated evaporator superheat is 7.2℃, exceeding the normal operating range by 4-6℃; the calculated compressor pressure ratio (discharge pressure 1.45MPa / suction pressure 0.42MPa) is 3.45, significantly deviating from the design value by 2.8. Regarding operational characteristics, the chilled water supply and return temperature difference is 5.6℃, within the normal range of 5-6℃; the equipment load rate is approximately 54%. After Z-score standardization, all characteristic parameters show that the condenser terminal temperature difference deviates from the historical average by 3.8 standard deviations, the compressor pressure ratio by 2.9 standard deviations, the evaporator superheat by 1.8 standard deviations, and the chilled water supply and return temperature difference is within ±1σ.

[0078] The two-level identification architecture is activated. In the first-level three-threshold discrimination, the condenser terminal temperature difference far exceeds the attention threshold (1.5℃) and the alarm threshold (2.0℃), and is directly marked as "abnormal"; the compressor pressure ratio (3.45) exceeds the attention threshold (3.0), and is marked as "attention"; the superheat (7.2℃) exceeds the attention threshold (6.0℃) but does not reach the alarm threshold (8.0℃), and is marked as "attention". Based on the comprehensive judgment, Unit #2 is determined to be in an "abnormal" state. The second-level random forest classifier receives a 12-dimensional standardized feature vector and inputs it into a pre-trained model (the model's training data covers normal operating conditions and typical fault simulation data of this type of unit under laboratory conditions, including 6 types of fault modes such as condenser blockage, refrigerant leakage, compressor efficiency reduction, and abnormal expansion valve opening, with each type of fault labeled according to different degrees). The classifier outputs: For Unit #2, the confidence scores for "condenser blockage" are 0.89, "refrigerant leakage" is 0.12, and "compressor efficiency reduction" is 0.07. The contributing features of condenser blockage are mainly abnormal condenser end difference (ranked 1st in feature importance, contributing 68%) and increased compressor pressure ratio (ranked 2nd, contributing 22%). The final output diagnostic conclusion is: the operating status is "abnormal", the abnormal equipment location is the condenser of Unit #2 chiller, the suspected fault type is "condenser tube wall scaling / blockage", the abnormality level is "severe" (end difference deviates from the baseline value by more than 200%, corresponding to an estimated energy efficiency loss of 18~22%), and the energy efficiency loss risk level is "high".

[0079] The energy efficiency diagnostic engine packages and transmits the above-mentioned abnormal characteristic data—#2 unit condenser abnormality (terminal temperature difference -1.6℃ vs design value 2.5℃, deviation 4.1℃; pressure ratio 3.45 vs design value 2.8; power 320 kW vs normal value approximately 280 kW)—to the causal reasoning module. After receiving the data, the causal reasoning module counterfactually changes the abnormal characteristic data (condenser terminal temperature difference -1.6℃) belonging to the intermediate physical state to the design baseline value of 2.5℃. Keeping the environmental variables (outdoor temperature 33℃, cooling load 3800kW) unchanged, it calculates along the causal graph: the condensing temperature drops from 40.2℃ to 35.5℃, the condensing pressure drops from 1.45MPa to 1.12MPa, the pressure ratio drops from 3.45 to 2.7, the compressor power consumption drops from 320kW to 268kW, and the COP recovers from the current 3.3 (=1050 / 320) to 4.7 (=1050 / 268). The quantification shows that condenser blockage causes a 1.4% decrease in COP, accounting for approximately 85% of the total energy efficiency loss. Based on this, the system generates an optimization strategy: online cleaning of the #2 unit condenser, with an expected COP increase of approximately 42% and annual energy savings of approximately 420,000 kWh.

[0080] The digital twin simulation module receives the "condenser blockage" output by the causal reasoning module as the root cause, loads the pre-trained PINN architecture model (which has been pre-trained for this type of chiller unit), and performs forward reasoning on the optimization strategy: it assumes that after the condenser terminal temperature difference recovers to 2.5℃, the compressor power consumption is predicted to decrease from 320kW to 260kW, and the COP to increase to 4.9, which is basically consistent with the counterfactual reasoning results. The final system outputs the optimization scheme and expected energy-saving benefits. Field implementation verified that the deviation from the prediction is ≤3%.

[0081] In another embodiment, the air conditioning energy efficiency diagnosis and optimization system and method of the present invention are applied to the energy efficiency diagnosis and optimization of the central air conditioning system of a large shopping mall: The large shopping mall has a building area of ​​80,000 m², and its central air conditioning system includes 3 centrifugal chillers (each with a cooling capacity of 1200 RT), 4 chilled water pumps, 4 cooling water pumps, 3 cooling towers, and corresponding control systems. The portable data acquisition terminal is deployed at key nodes such as the control cabinet, chilled water pipeline, cooling water pipeline, and cooling tower of the chiller unit. The acquisition terminal collects the following multi-dimensional operating data continuously for 72 hours via a wireless sensor network: compressor input power and evaporator / condenser inlet and outlet water temperatures of each unit, chilled water main flow rate and supply and return water temperature difference, cooling water inlet and outlet water temperatures and cooling tower fan operating status, indoor and outdoor temperature and humidity and air conditioning load data of each area of ​​the building. A total of 32 sensor nodes are deployed, and all data is uploaded to the portable evaluation host in real time at a 5-second interval (in this example, the host has a computing power of 8 TFLOPS and 6 GB of memory).

[0082] The pre-installed air conditioning thermodynamic structure causal model (SCM) in the causal reasoning module is used to establish a causal topology diagram. Based on the multi-dimensional operational data collected on-site, the following abnormal characteristic data are obtained through the energy efficiency diagnostic engine: Environmental variables: outdoor temperature is 32℃, indoor load is 3800 kW; Intervention variable: chilled water pump frequency was 45 Hz (design value / historical normal value was 45 Hz, which did not change); Intermediate physical conditions: The condenser terminal temperature difference of Unit 2 is 5.2℃ (design value / historical normal value is 2.5℃), and the compressor pressure ratio of Unit 2 is 4.8 (design value / historical normal value is 3.2). Target variables: Unit 2 compressor power consumption is 850 kW (design value / historical normal value is 620 kW), system COP is 3.8 (design value / historical normal value is 5.2); The counterfactual reasoning engine is invoked to perform counterfactual reasoning on the abnormal characteristic data of Unit 2 condenser terminal temperature difference deviating from the baseline value of 2.7℃ (5.2-2.5) belonging to the intermediate physical state: The condenser temperature differential was counterfactually changed from 5.2℃ to 2.6℃ (i.e., reduced by 50% through cleaning), while keeping all environmental parameters (outdoor temperature 32℃, indoor load 3800 kW) unchanged, and keeping other intervention variables (chilled water pump frequency, etc.) unchanged. Forward propagation computation along the directed edges of the causal topological graph: Reduced condenser terminal temperature → condensing temperature drops from 42℃ to 36℃; reduced condensing temperature → compressor pressure ratio drops from 4.8 to 3.4; reduced compressor pressure ratio → compressor power consumption drops from 850 kW to 670 kW; reduced compressor power consumption + unchanged cooling capacity → system COP increases from 3.8 to 4.9. The counterfactual reasoning engine outputs the diagnosis of "increased thermal resistance of fouling on the water side of the condenser" as the root cause, and quantifies it: This abnormal characteristic data (the temperature difference at the condenser terminals of Unit 2 deviates from the baseline value by 2.7℃) causes an additional 180 kW of power consumption in the compressor and a 1.1% reduction in COP (accounting for 78% of the total COP reduction).

[0083] The digital twin simulation module adopts a PINN architecture. The network input is the current operating status (condenser terminal temperature difference of 5.2℃, chilled water flow rate, outdoor temperature, etc.), and the output is the predicted values ​​of system COP and compressor power consumption. In the loss function, λ... pde Take 100, λ bc Take 50, L pde The residuals of the partial differential heat transfer equations for embedded heat exchangers.

[0084] Since Unit #2 has been in operation for 5 years, there is a decrease in the heat transfer coefficient due to heat exchanger aging. The system performs GAN-inspired online calibration: using 72 hours of multi-dimensional operating data as input, the condition generator outputs a degradation factor γ_UA = 0.82 (indicating that the heat transfer coefficient has decreased to 82% of the design value); the frozen layer of the PINN architecture is frozen, and only the adaptation layer is fine-tuned. The optimizer Adam has a learning rate of 1e-4, iterates for 45 rounds, and takes about 8 seconds. After calibration, the model's COP prediction error for the current operating condition is reduced from ±0.35 to ±0.08.

[0085] Based on the counterfactual reasoning engine's diagnostic conclusion of "increased thermal resistance due to fouling on the condenser water side," the optimization strategy generation module generates the following candidate optimization strategies: Optimization Strategy A: Online cleaning of the condenser of Unit 2 is expected to reduce the terminal temperature difference from 5.2℃ to 2.5℃.

[0086] Optimization Strategy B (Control Optimization): Optimize the chilled water pump frequency control curve, adjusting the frequency from 45 Hz to a range of 38~48 Hz that is adjusted in real time according to the load.

[0087] The PINN architecture of the digital twin simulation module receives the diagnostic conclusion of "increased thermal resistance of condenser water-side fouling" output by the counterfactual reasoning engine and performs forward reasoning on strategy A (single reasoning time 32 ms), outputting "After implementing strategy A": condenser terminal temperature difference 2.5℃, compressor power consumption 660 kW, system COP 5.0, which is 31.6% higher than the current COP 3.8.

[0088] The human-computer interaction module visualizes the attribution chain of energy efficiency issues and the expected energy-saving benefits in the form of a cause-and-effect graph. The energy efficiency problem attribution chain is as follows: "Condenser terminal temperature difference of 5.2℃ (abnormal) → condenser temperature rise → compressor pressure ratio increase → compressor power consumption increase of 180 kW → COP drops from 5.2 to 3.8". Energy efficiency loss contribution: Condenser fouling contributes 78%; Simulation verification was also performed on optimization strategy B, and the results showed that its energy efficiency loss contribution was 22%. Therefore, strategy A (online cleaning) should be implemented first, which is expected to save 750,000 kWh of electricity per year.

[0089] Actual verification: Online chemical cleaning was performed on the condenser of Unit #2. Post-cleaning, measured data showed that the condenser terminal temperature difference decreased to 2.4℃, compressor power consumption decreased to 655 kW, and the system COP increased to 5.1. The actual results deviated from the PINN simulation prediction (COP 5.0) by only 2%, demonstrating the high-precision predictive capability of the digital twin model of this invention.

[0090] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can readily conceive of other specific embodiments of the invention without inventive effort, and these embodiments will all fall within the scope of protection of the present invention.

Claims

1. A method for diagnosing and optimizing air conditioning energy efficiency, characterized in that, Includes the following steps: Collect real-time, multi-dimensional operational data of the air conditioning system; The collected multidimensional operational data is processed to obtain abnormal feature data; A causal topology graph is established using the multidimensional operational data as nodes. The directed edges of the causal topology graph are constrained by the thermodynamic conservation law of air conditioning and prior knowledge of the physical properties of the components. Based on the abnormal feature data and the causal topology graph, counterfactual reasoning is performed to output a diagnostic conclusion with causal chain explanation, and candidate optimization strategies are proposed based on the causal attribution of the diagnostic conclusion. The physical information neural network architecture performs forward inference and simulation verification based on the diagnostic conclusions and the candidate optimization strategies. The loss function of the physical information neural network architecture is embedded in the residuals of the partial differential equations of the heat exchanger and the energy conservation equation of the compressor as physical constraint terms.

2. The air conditioning energy efficiency diagnosis and optimization method according to claim 1, characterized in that, The types of nodes in the causal topology graph include intervention variables, environmental variables, intermediate physical states, and target variables; wherein, the intervention variables represent operating parameters in the air conditioning system that are actively adjusted by the controller or by human intervention, the environmental variables represent uncontrollable boundary conditions outside the air conditioning system, the intermediate physical states represent measurable or calculable physical quantities of the thermal state of key components in the air conditioning system, and the target variables represent evaluation indicators of the energy efficiency of the air conditioning system.

3. The air conditioning energy efficiency diagnosis and optimization method according to claim 2, characterized in that, The process of processing the collected multidimensional operational data to obtain abnormal feature data includes the following steps: The collected multidimensional operating data is cleaned, and the multidimensional operating data includes cooling water supply and return water temperature, cooling water inlet and outlet water temperature, and compressor power. The variance is calculated based on the cooling water supply and return water temperature, cooling water inlet and outlet water temperature, and compressor power in each sliding window. When all the variances in the sliding window are lower than the corresponding preset threshold, it is determined to be a steady-state operating condition, and the corresponding feature data segment under the steady-state operating condition is obtained. Feature extraction is performed on the corresponding feature data segments under the steady-state operating conditions to obtain feature vectors. The feature vectors include energy efficiency feature data, thermal feature data, and operational feature data. Sliding windows are applied to the extracted feature vectors respectively. Standardization methods are used for normalization. The standardized feature vectors belonging to the intermediate physical state of the thermal feature data and the operational feature data are used to classify the operating status of the air conditioning system into three levels: normal, attention, and abnormal, using a three-level threshold discrimination method. The feature vectors corresponding to the operating status of attention or abnormal are used as abnormal feature data and input into a pre-trained random forest classifier to output the corresponding air conditioning system operating status, abnormal equipment location, suspected fault type, abnormal level, and energy efficiency loss risk level.

4. The air conditioning energy efficiency diagnosis and optimization method according to claim 3, characterized in that, The abnormal feature data belonging to the intermediate physical state type is counterfactually changed to the baseline value, while keeping the environmental variables and other variables that have not experienced anomalies unchanged. The calculation is performed forward along the directed edge direction of the causal topology graph until the target variable in the counterfactual scenario is obtained. The difference between the target variable in the counterfactual scenario and the target variable in the current actual scenario is used as a quantitative indicator to quantify the contribution of the abnormal features to energy efficiency loss, thereby outputting a diagnostic conclusion with causal chain explanation, and proposing candidate optimization strategies based on the causal attribution of the diagnostic conclusion.

5. The air conditioning energy efficiency diagnosis and optimization method according to claim 1, characterized in that, The physical information neural network architecture includes an input layer, a freeze layer, an adaptation layer, and an output layer; The input layer receives the multidimensional operational data. The frozen layer is a physical constraint layer. The network weights of the frozen layer are determined and kept frozen before leaving the factory by performing pre-training on the physical information neural network architecture. The pre-training uses the loss function as the optimization objective. The loss function is embedded with the residuals of the heat transfer partial differential equation of the heat exchanger and the residuals of the energy conservation equation of the compressor as physical constraint terms. The frozen layer consists of four layers, each containing 128 neurons, with the activation function being tanh; The adaptation layer is an online fine-tuning layer, and the weights of the adaptation layer are the identity matrix initialized when the device leaves the factory; the adaptation layer consists of two layers, each containing 32 neurons, and the activation function is ReLU; The output layer outputs two-dimensional energy efficiency indicators, including the system comprehensive performance coefficient and the compressor input power consumption.

6. The air conditioning energy efficiency diagnosis and optimization method according to claim 5, characterized in that, The physical information neural network architecture receives the diagnostic conclusion and performs forward inference and simulation verification on the candidate optimization strategies, and also includes an online calibration step: Collect 48-72 hours of on-site operating data of the air conditioning system and extract the feature matrix X; Input the feature matrix X into the trained condition generator G(z|X) and output the degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the current air conditioning system, where: γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate. The frozen layer is frozen, and the generated degradation factor vector γ is substituted into the loss function to update the baseline physical parameters in the heat exchanger partial differential equation and the compressor energy conservation equation. The weights of the adaptation layer are then fine-tuned online: the optimizer used during online fine-tuning is Adam, and the learning rate is set to 1×10⁻⁶. -4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data, as well as the residual of the physical constraint term after updating the parameters. After online fine-tuning, the physical information neural network architecture performs subsequent forward inference and simulation verification based on the loss function after updating the baseline physical parameters and the fine-tuned adaptation layer.

7. The air conditioning energy efficiency diagnosis and optimization method according to claim 6, characterized in that, The expression for the loss function is: L total = L data + λ pde · L pde + λ bc · L bc ; Where: λ pde and λ bc These are the preset penalty weight coefficients for balancing the physical constraint gradient and the boundary condition gradient; L data This represents a data fitting term used to penalize the deviation between the physical state quantities and energy efficiency indicators output by the physical information neural network architecture and the multidimensional operating data collected in real time. L pde The physical constraint term, used to penalize the residuals of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger and the isentropic compression energy conservation equation of the compressor, is expressed as: ;in, This represents the residual term of the nonlinear one-dimensional transient heat transfer partial differential equation of the air conditioner heat exchanger; This represents the residual term in the isentropic compression energy conservation equation of the compressor. L bc The boundary condition constraint term is used to penalize the mean square error between the predicted physical state quantity output by the physical information neural network architecture at the set spatial domain boundary and the measurement value collected in real time at the corresponding real physical boundary of the device.

8. The air conditioning energy efficiency diagnosis and optimization method according to claim 1, characterized in that, The air conditioning energy efficiency diagnosis and optimization method also includes a missing data processing mechanism. The steps of the missing data processing mechanism are as follows: when sensor data loss occurs during the collection of real-time multi-dimensional operating data of the air conditioning system, the missing data is reconstructed based on Markov equivalence classes of causal topology. The conditional independence relationship of missing nodes is determined by using Markov equivalence classes. Then, the energy conservation equation of the heat exchanger and the data of the parent and child nodes that have not been lost are used as constraints. The Gibbs sampling algorithm is used to perform maximum a posteriori estimation of the missing physical state variables, so that the residual of the reconstructed complete dataset that satisfies the energy conservation equation is less than a preset threshold. The number of Gibbs sampling iterations is 1000, of which the first 200 are discarded as an aging period, and the a posteriori estimation is the mean of the remaining samples.

9. An air conditioning energy efficiency diagnosis and optimization system, characterized in that, The air conditioning energy efficiency diagnosis and optimization system, used to implement any one of the air conditioning energy efficiency diagnosis and optimization methods according to claims 1 to 8, comprises: A portable data acquisition terminal, used to collect real-time multi-dimensional operating data of an air conditioning system; An energy efficiency diagnostic engine is used to process the collected multidimensional operating data and obtain abnormal feature data. The causal reasoning module embeds a causal model of the air conditioning thermodynamic structure, a counterfactual reasoning engine, and an optimization strategy generation module. The causal model of the air conditioning thermodynamic structure is used to establish a causal topology graph using the multidimensional operational data as nodes. The counterfactual reasoning engine performs counterfactual reasoning based on the abnormal feature data and the causal topology graph, and outputs a diagnostic conclusion with causal chain explanation. The optimization strategy generation module is used to propose candidate optimization strategies based on the causal attribution of the diagnostic conclusion. A digital twin simulation module, which embeds a physical information neural network architecture, is used for forward inference and simulation verification based on the diagnostic conclusion and the candidate optimization strategies.

10. The air conditioning energy efficiency diagnosis and optimization system according to claim 9, characterized in that, The digital twin simulation module also includes an online calibration unit, which is used to perform the following steps: Collect 48-72 hours of on-site operating data of the air conditioning system and extract the feature matrix X; Input the feature matrix X into the trained condition generator G(z|X) and output the degradation factor vector γ = [γ_UA, γ_comp, γ_fan] of the current air conditioning system, where: γ_UA represents the heat transfer coefficient decay rate, γ_comp represents the compressor energy efficiency decay rate, and γ_fan represents the fan efficiency decay rate. The frozen layer is frozen, and the generated degradation factor vector γ is substituted into the loss function to update the baseline physical parameters in the heat exchanger partial differential equation and the compressor energy conservation equation. The weights of the adaptation layer are then fine-tuned online: the optimizer used during online fine-tuning is Adam, and the learning rate is set to 1×10⁻⁶. -4 The number of iterations does not exceed 50, and the objective function is to minimize the mean square error between the simulation output and the measured data, as well as the residual of the physical constraint term after updating the parameters. After online fine-tuning, the physical information neural network architecture performs subsequent forward inference and simulation verification based on the loss function after updating the baseline physical parameters and the fine-tuned adaptation layer.

Images