Method for identifying and operating and maintaining alarm of abnormal energy consumption of heating system for smart city

Abstract

The present application relates to the technical field of heating and ventilation abnormal energy consumption identification, in particular to a heating and ventilation system abnormal energy consumption identification and operation and maintenance alarm method for smart city. It comprises the following steps: collecting real-time operation data of the target building heating and ventilation system and external environment disturbance data, and obtaining equivalent meteorological temperature through thermal resistance capacity hysteresis correction; inputting it into a lightweight gradient boosting tree prediction theory energy consumption with a thermodynamic physical lower limit penalty; calculating the total energy consumption residual and performing orthogonal decomposition, eliminating meteorological and passenger flow collinear disturbance, and extracting pure equipment operation recession component; finally matching the operation and maintenance expert rule base, generating an alarm work order containing fault space location and processing suggestion and pushing it. The present application eliminates algorithm black box prediction distortion, completely removes false alarms caused by external disturbance, and realizes the automatic closed loop from macroscopic abnormal perception to microscopic accurate fault elimination.

Description

Technical Field

[0001] This invention relates to the field of HVAC abnormal energy consumption identification technology, and more specifically, to a method for identifying and alarming abnormal energy consumption in HVAC systems for smart cities. Background Technology

[0002] In the refined energy management system of smart city building complexes, the Heating, Ventilation, and Air Conditioning (HVAC) system, as the largest flexible energy-consuming carrier, directly determines the energy conservation and emission reduction effectiveness of large commercial complexes and public buildings. Under actual physical conditions, the surge in transient total energy consumption of HVAC systems is essentially a macroscopic manifestation of the deep coupling and intertwining of "external dynamic load disturbances" (such as drastic changes in urban micro-meteorology and sudden surges in passenger flow) and "internal physical equipment degradation" (such as severe scaling of heat exchangers and wear of hydraulic pipelines). Currently, most mainstream building automation systems (BMS) and existing energy consumption monitoring and early warning algorithms rely on static energy efficiency thresholds or simple surface-level data-driven models. However, facing the complex and ever-changing urban building micro-environment, these existing technologies generally encounter the following serious technical bottlenecks in terms of filtering out environmental white noise and accurately locating real equipment-level anomalies:

[0003] In actual operation, when encountering extreme high temperatures outdoors or sudden surges in customer traffic in shopping malls, system energy consumption will naturally increase. However, existing technologies often trigger alarms by simply comparing the residuals between "actual energy consumption" and "predicted energy consumption," failing to address the underlying mathematical logic that exposes the false high energy consumption induced by sudden changes in micro-weather conditions and a surge in pedestrian traffic. This interference from multicollinearity leads to a massive number of false alarms being received by smart city operation and maintenance terminals during peak load periods, significantly reducing the trust of operation and maintenance personnel in the system. Therefore, there is an urgent need to provide a method for identifying abnormal energy consumption and issuing operation and maintenance alarms for HVAC systems in smart cities. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying abnormal energy consumption and issuing operation and maintenance alarms for HVAC systems in smart cities, in order to solve the problem of interference from multicollinearity features mentioned in the background art, which often leads to a large number of false alarms being received by smart city operation and maintenance terminals during peak load periods, resulting in a significant decrease in the trust of operation and maintenance personnel in the system.

[0005] To achieve the above objectives, the present invention aims to provide a method for identifying abnormal energy consumption and issuing maintenance alarms for HVAC systems in smart cities, comprising the following steps:

[0006] S1. Collect real-time operating data of the target building's HVAC system, as well as external environmental disturbance data including micro-meteorological data and real-time building pedestrian density data, and perform time-series correction on the micro-meteorological data to obtain the equivalent outdoor meteorological temperature.

[0007] Specifically, the real-time operating data, the equivalent outdoor meteorological temperature, and the real-time building pedestrian density data are preprocessed to extract equipment status features and external load demand features, respectively. The equipment status features are then cached in a time series, and a historical operating feature set is constructed accordingly.

[0008] S2. Real-time acquisition of the external load demand characteristics at the current moment, and inputting them into a pre-trained lightweight gradient boosting tree model to calculate the theoretical expected energy consumption value for maintaining thermal balance of the HVAC system.

[0009] S3. Collect the actual total energy consumption value of the HVAC system at the current moment, and calculate the total energy consumption residual between the actual total energy consumption value and the theoretical expected energy consumption value. Perform orthogonal decomposition on the total energy consumption residual to directly extract the equipment operation degradation component that characterizes the performance degradation of the equipment itself.

[0010] A real device-level abnormal energy consumption event is determined to be triggered only when the continuous set time window of the device operation degradation component exceeds the preset device health tolerance threshold.

[0011] S4. After triggering the real device-level abnormal power consumption event, extract the device status features within the trigger event time window;

[0012] Using a pre-defined degradation feature classification regression tree model, the feature contribution of each device state feature to the device operation degradation component is calculated, and the core degradation parameters that cause a sudden increase in system energy consumption are identified accordingly.

[0013] S5. Based on the identified core degradation parameters, match them with a preset operation and maintenance expert rule base to determine the corresponding specific fault type;

[0014] An operation and maintenance alarm work order containing the fault location, the specific fault type, and handling suggestions is generated and pushed to the smart city operation and maintenance terminal.

[0015] As a further improvement to this technical solution, the real-time operating data includes at least the real-time power of the chiller unit, the chilled water supply / return temperature, the chilled water supply / return pressure, the cooling water supply / return temperature, the operating frequency of the cooling pump, the operating frequency of the chilled water pump, and the operating frequency of the cooling tower fan.

[0016] Micrometeorological data should include at least outdoor dry-bulb temperature, solar radiation intensity, and real-time outdoor wind speed;

[0017] Real-time pedestrian density data for buildings should include at least the real-time pedestrian density index.

[0018] As a further improvement to this technical solution, time-series correction is performed on the micrometeorological data, involving the following specific steps:

[0019] Obtain the current time from micrometeorological data. outdoor dry bulb temperature and solar radiation intensity And the solar radiation absorption coefficient of the target building envelope is introduced. and outdoor surface heat transfer coefficient Calculate the current outdoor temperature. ;

[0020] Based on the combined outdoor temperatures at the current and historical times, a thermal resistance-capacitance hysteresis time constant for the building envelope is introduced. and data sampling period Dynamic weighted calculation of phase lag and amplitude attenuation is performed to output the equivalent outdoor meteorological temperature. .

[0021] As a further improvement to this technical solution, the construction and training steps of the lightweight gradient boosting tree model are as follows:

[0022] A lightweight gradient boosting tree model is constructed based on the additive model and the forward distribution algorithm. This lightweight gradient boosting tree model consists of... The classification and regression trees are composed of the sum of the tree topology and the weights of the leaf nodes;

[0023] Sample data from the historical operating feature set during the normal operating period of the HVAC system are extracted, and the external load demand features therein are used as input variables, with the actual total energy consumption value of the same period as the true label value.

[0024] Construct a joint loss function for iterative training of the model, wherein the joint loss function consists of a data fitting error term between the true label value and the predicted energy consumption value of the current iteration of the model and a thermodynamic energy conservation penalty term;

[0025] During the training process of the lightweight gradient boosting tree model, the first and second derivatives of the joint loss function with respect to the predicted energy consumption value are calculated and substituted into the split gain calculation of the leaf nodes of the classification and regression trees.

[0026] The pre-trained lightweight gradient boosting tree model is obtained by minimizing the joint loss function.

[0027] During the online prediction phase, the external load demand characteristics at the current moment are acquired in real time and input into the pre-trained lightweight gradient boosting tree model to output the theoretical expected energy consumption value.

[0028] As a further improvement to this technical solution, the specific steps for constructing the thermodynamic energy conservation penalty term are as follows:

[0029] Extract the equivalent outdoor meteorological temperature and the real-time building pedestrian density data from the external load demand characteristics;

[0030] By introducing the comprehensive heat transfer coefficient of the building envelope and the average heat dissipation per person, and combining the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building, the dynamic cooling and heating load of the target building under the current working conditions is calculated.

[0031] Based on the dynamic heating and cooling loads and the theoretical maximum energy efficiency ratio of the HVAC system, the theoretical minimum energy consumption physical lower limit for maintaining thermodynamic balance of the HVAC system is calculated.

[0032] When the predicted energy consumption value of the current iteration of the lightweight gradient boosting tree model is lower than the theoretical minimum energy consumption physical lower limit, the thermodynamic energy conservation penalty term is activated, and a penalty calculation is applied to the predicted energy consumption value.

[0033] As a further improvement to this technical solution, in step S3, the equipment operation degradation component characterizing the performance degradation of the equipment itself is extracted. The specific steps involved are as follows:

[0034] Obtain the actual total energy consumption value and the theoretical expected energy consumption value at the current moment, and calculate the difference to obtain the total energy consumption residual;

[0035] Based on a sliding analysis time window of a set length, the total energy consumption residual obtained by continuous calculation is cached in time series, and the total energy consumption residual within the sliding analysis time window is concatenated in time series to form a total energy consumption residual vector; at the same time, the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building corresponding to the sliding analysis time window are extracted and constructed as micro-meteorological disturbance basis vector and pedestrian surge disturbance basis vector, respectively.

[0036] The micro-meteorological disturbance basis vector and the population surge disturbance basis vector are orthogonalized to construct a multi-dimensional disturbance feature space;

[0037] The total energy consumption residual vector is projected onto the multidimensional disturbance feature space to obtain the micro-meteorological disturbance projection vector mapped to the direction of the micro-meteorological disturbance basis vector, and the population surge disturbance projection vector mapped to the direction of the population surge disturbance after collinearity elimination.

[0038] Using vector subtraction, the micro-meteorological disturbance projection vector and the population surge disturbance projection vector are filtered and separated from the total energy consumption residual vector to obtain the equipment degradation residual vector;

[0039] Extract the last element scalar in the device decay residual vector corresponding to the current moment, and use it as the device operation decay component at the current moment.

[0040] As a further improvement to this technical solution, the specific steps involved in identifying the core degradation parameters that cause a sudden increase in system energy consumption are as follows:

[0041] After triggering the real device-level abnormal power consumption event, the device state characteristics synchronously cached within the trigger event time window are obtained as the input set of independent variables;

[0042] The pre-trained degradation feature classification and regression tree model is invoked. The degradation feature classification and regression tree model uses the equipment state features of the historical operation stage as input variables and the corresponding equipment operation degradation component as fitting labels to obtain a nonlinear mapping network.

[0043] The independent variable input set is input into the degradation feature classification regression tree model, and the SHAP interpretability algorithm based on game theory is introduced to calculate the marginal contribution value of each device state feature in the independent variable input set to the current output device operation degradation component, and aggregate to generate global feature contribution.

[0044] The device status features are sorted in descending order according to the contribution of the global features, and the device status features that meet the preset contribution threshold are used as the core degradation parameters, thereby generating an operation and maintenance alarm work order for the target building HVAC system.

[0045] As a further improvement to this technical solution, in step S5, the specific steps for generating an operation and maintenance alarm work order by matching the identified core degradation parameters with a preset operation and maintenance expert rule base include:

[0046] Construct a pre-defined operation and maintenance expert rule base, which includes a multi-dimensional mapping matrix;

[0047] The mapping key values ​​of the multidimensional mapping matrix include at least: the physical label of the core degradation parameter, the deviation direction of the feature value, and the abnormal state of the combined associated parameters; the mapping target value of the multidimensional mapping matrix is ​​the corresponding specific fault type and processing suggestion.

[0048] Extract the physical tags of the core degradation parameters identified in step S4, calculate the average value of the core degradation parameters within the set time window that triggers the real device-level abnormal energy consumption event, and compare it with the baseline expected value in the historical absolute healthy operation phase to determine the corresponding deviation direction of the feature value.

[0049] The deviation direction of the physical label and the feature value is combined with the real-time deviation status of the combined associated parameters belonging to the same physical subsystem as the physical label, and then input into the operation and maintenance expert rule base for multi-dimensional mapping and matching to determine the specific fault type and the handling suggestion corresponding to the current target building HVAC system.

[0050] Extract the unique device identifier directly bound to the core degradation parameters, perform spatial address addressing in the device topology network, obtain the corresponding physical device node spatial coordinates as the fault location, and use it as the final fault location.

[0051] The fault location, the specific fault type, the handling suggestions, and the global feature contribution of the core degradation parameters are encapsulated into a structured message form of an operation and maintenance alarm work order and pushed to the smart city operation and maintenance terminal.

[0052] As a further improvement to this technical solution, the specific mapping logic of the multidimensional mapping matrix stored in the preset operation and maintenance expert rule base includes at least the following:

[0053] When the physical label of the extracted core degradation parameter includes cooling water approximation degree, and the corresponding feature value deviates in the positive direction and is relatively high, and the feature value of the combined associated cooling water pump inlet and outlet pressure difference parameter increases abnormally, the specific fault type output by the multidimensional mapping matrix is ​​cooling side heat exchange abnormality, and the underlying physical fault source corresponding to the mapping is located as severe scaling of cooling tower heat dissipation packing or blockage of cooling water pipe network filter.

[0054] When the physical label of the extracted core degradation parameter includes the temperature difference between chilled water supply and return water, and the corresponding feature value deviates negatively and is low, while the feature value of the associated chiller unit operating frequency parameter deviates positively and is high, the specific fault type output by the multidimensional mapping matrix is ​​a high flow rate and low temperature difference syndrome. The underlying physical fault source corresponding to this mapping is located as severe dust accumulation on the surface cooler of the terminal air handling unit or a malfunctioning water-side two-way regulating valve.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. In this method for identifying abnormal energy consumption and alarming operation and maintenance of HVAC systems for smart cities, the physical lag effect of heat conduction is accurately restored by introducing the thermal resistance-capacitance hysteresis time constant of the building envelope for dynamic weighting.

[0057] Meanwhile, when constructing the lightweight gradient boosting tree model, the theoretical minimum energy consumption physical lower limit for maintaining thermodynamic equilibrium is directly written into the underlying loss function as a strong constraint penalty term; this invention not only possesses the nonlinear fitting capability of machine learning, but is also strictly limited to the framework that conforms to the real thermodynamic heat transfer law, providing an extremely accurate and absolutely legal energy consumption comparison benchmark for subsequent diagnosis.

[0058] 2. In this method for identifying abnormal energy consumption and alarming operation and maintenance of HVAC systems for smart cities, high-frequency time-series residuals are transformed into vectors and then subjected to inner product projection and stripping in mutually independent orthogonal feature spaces of meteorology and human traffic. Multicollinear external environmental disturbance white noise is filtered out, and the decay residual vectors that represent only the deterioration of equipment physical performance are accurately purified.

[0059] Combined with a health tolerance threshold adaptively generated based on the Gaussian distribution 3σ criterion, it ensures that the warning is only triggered when real mechanical or thermal degradation occurs, which greatly reduces the invalid dispatch rate of smart city operation and maintenance centers. Attached Figure Description

[0060] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0061] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1: Please refer to Figure 1 As shown, this embodiment provides a method for identifying abnormal energy consumption and issuing operation and maintenance alarms for HVAC systems in smart cities, including the following steps:

[0063] S1. Collect real-time operating data of the target building's HVAC system, as well as external environmental disturbance data including micro-meteorological data and real-time building pedestrian density data, and perform time-series correction on the micro-meteorological data to obtain the equivalent outdoor meteorological temperature.

[0064] Specifically, the real-time operating data, the equivalent outdoor meteorological temperature, and the real-time building pedestrian density data are preprocessed to extract equipment status features and external load demand features, respectively. The equipment status features are then cached in a time series, and a historical operating feature set is constructed accordingly.

[0065] Furthermore, the real-time operating data includes at least the real-time power of the chiller unit, the chilled water supply / return temperature, the chilled water supply / return pressure, the cooling water supply / return temperature, the operating frequency of the cooling pump, the operating frequency of the chilled water pump, and the operating frequency of the cooling tower fan.

[0066] Micrometeorological data should include at least outdoor dry-bulb temperature, solar radiation intensity, and real-time outdoor wind speed. In addition, it may also include conventional meteorological parameters such as outdoor relative humidity and atmospheric pressure.

[0067] Real-time pedestrian density data for buildings includes at least the number of people residing in a given area, collected by Wi-Fi probes, visual security cameras, or access control gates deployed within the smart building, and converted into a real-time pedestrian density index (unit: people / ).

[0068] It is worth noting that the specific steps involved in performing time-series correction on the aforementioned micrometeorological data are as follows:

[0069] Obtain the current time from micrometeorological data. outdoor dry bulb temperature and solar radiation intensity And the solar radiation absorption coefficient of the target building envelope is introduced. and outdoor surface heat transfer coefficient Calculate the current outdoor temperature. :

[0070] ;

[0071] In the formula, Indicates the current time The outdoor temperature (usually in °C); Indicates the current time The outdoor dry-bulb temperature (in °C) is the actual air temperature directly measured by a weather station or external temperature sensor. This represents the solar radiation absorption coefficient (dimensionless, ranging from 0 to 1) of the target building envelope (such as a glass curtain wall or concrete exterior wall). The darker the color and the less reflective the material of the target building envelope, the larger this value will be. (Specifically, if the exterior wall is a light-colored glass curtain wall or a high-reflective coating...) The value range is set to 0.15~0.30; if the exterior wall is light-colored concrete or brick structure, The value range is set to 0.40~0.55; if the exterior wall is dark concrete, asphalt, or dark stone, The value range is set to 0.65~0.85). Indicates the current time Solar radiation intensity (unit: This reflects the heat flux density of sunlight hitting the exterior wall; The outdoor surface heat transfer coefficient (or convective heat transfer coefficient) of the target building, in units of This indicates the ability of outdoor air to carry away or transfer heat when it flows over the wall surface, and is greatly affected by outdoor wind speed.

[0072] Specifically, in this embodiment, the current real-time outdoor wind speed is obtained from the collected micrometeorological data. (Unit: m / s) Dynamic calculation, then ;

[0073] In the formula, The base represents the natural convection and radiation heat transfer coefficients under windless conditions (values ​​ranging from 5.0 to 6.0). ); Wind speed correlation coefficient (unit: (The value ranges from 3.5 to 4.0); in this embodiment, the outdoor surface heat transfer coefficient of the target building is... Using formula Perform real-time dynamic solution;

[0074] It is worth noting that large commercial complexes or CBD office buildings in smart cities typically use large areas of glass curtain walls; in summer, even if the outdoor temperature ( The temperature is only 30℃, but if exposed to direct sunlight (radiation intensity) The actual perceived temperature on the exterior wall surface can be as high as 45°C or higher. Using only the outdoor temperature would severely underestimate the actual air conditioning cooling load. Therefore, utilizing the solar radiation absorption coefficient of the exterior surface (…) ) and convective heat transfer coefficient ( This directly aligns the temperature load and radiation load in one dimension, greatly improving the real physical fidelity of the model input;

[0075] Furthermore, concrete walls or floors have extremely high thermal inertia (heat storage capacity), and the sun is strongest outdoors at noon (at the current time). outdoor comprehensive temperature While the heat reaches its peak, the heat transfer to the indoor environment causes a surge in air conditioning load, often not until 2 PM or even 3 PM (i.e., thermal delay and attenuation). If noon meteorological data is used directly to predict noon energy consumption, false alarms will occur, with predicted values ​​higher than actual values. Therefore, based on the current and historical outdoor temperatures, and addressing the lag in heating and cooling loads caused by the heat storage effect of the building envelope, a thermal resistance-capacitance hysteresis time constant of the building envelope is introduced to characterize the heat storage characteristics of the target building. and data sampling period Dynamic weighted calculation of phase lag and amplitude attenuation is performed to output the equivalent outdoor meteorological temperature. :

[0076] ;

[0077] In the formula, Indicates the current time The equivalent outdoor meteorological temperature (in °C); For the previous sampling period (time) The equivalent outdoor meteorological temperature calculated; Indicates the sampling period; This represents the thermal resistivity hysteresis time constant of the building envelope (in time, such as hours, minutes, or seconds).

[0078] In this embodiment, the time constant Specifically:

[0079] ;

[0080] In the formula, The number of structural layers in the building envelope (such as insulation layer, structural layer, and finishing layer). For the first Thickness of structural layer (unit: ); For the first Thermal conductivity of the structural layer material (unit: ); For the first Density of structural layer material (unit: ); For the first Specific heat capacity of structural layer material (unit: ); This represents the overall thermal resistance of the target building envelope; This represents the overall area heat capacity of the target building envelope.

[0081] In this embodiment, the real-time operating data, the equivalent outdoor meteorological temperature, and the real-time building pedestrian density data come from different heterogeneous systems (BMS typically samples at the second / minute level, while micro-meteorological and pedestrian flow data are typically sampled at the 15-minute or hour level). Therefore, by setting a unified reference time window (e.g., 15 minutes), the high-frequency continuous operating data is downsampled using a time window-based moving average; the low-frequency raw micro-meteorological data and pedestrian flow data are upsampled using cubic spline interpolation, ensuring that all data are strictly aligned at the same timestamp; and a density-based Local Anomaly Factor (LOF) algorithm is used to identify hidden sensor noise spikes, and the abnormal missing values ​​are filled using the average of historical operating conditions at the same time.

[0082] After the data is preprocessed, the equivalent outdoor weather temperature, the real-time passenger flow density index, and the derived time representation features (such as the weekday identifier DayOfWeek and the business hours identifier HourOfDay) are packaged together to form the external load demand features used to represent the external passive load pressure.

[0083] Meanwhile, the equipment status characteristics include at least: the temperature difference between chilled water supply and return (characterizing the terminal heat exchange efficiency), the proximity of cooling water (characterizing the heat dissipation performance of the cooling tower), and the pressure difference between the inlet and outlet of the water pump.

[0084] Furthermore, the effective timestamp after arbitrary alignment Next, the extracted features of each dimension are encapsulated into column vectors; and the same timestamps are... External load demand feature vector (dimension: ), device status feature vector (dimension: The scalar labels of the actual total energy consumption of the system during the same period are horizontally spliced ​​together in the feature dimension (column direction) to form a single-step comprehensive feature row vector;

[0085] Subsequently, in chronological order, the historical observation windows were... The single-step integrated feature row vectors corresponding to each discrete timestamp are stacked vertically along the sample dimension (row direction) to construct a dimension of The system generates a high-dimensional temporal feature matrix. To prevent dimensional bias in machine learning models during gradient descent or tree node splitting, the system performs Z-score normalization (or Min-Max normalization) on the feature columns of the concatenated high-dimensional temporal feature matrix to achieve dimensionless processing. Finally, a column slice mask index is set in the high-dimensional temporal feature matrix to generate a historical running feature set.

[0086] S2. Obtain the external load demand characteristics at the current moment in real time, and input them into the pre-trained lightweight gradient boosting tree model to calculate the theoretical expected energy consumption value of the HVAC system to maintain thermal balance under the current physical boundary conditions.

[0087] The construction and training steps of the lightweight gradient boosting tree model are as follows:

[0088] A lightweight gradient boosting tree model is constructed based on the additive model and the forward distribution algorithm. This lightweight gradient boosting tree model consists of... The classification and regression trees are composed of the sum of the tree topology and the weights of the leaf nodes;

[0089] Sample data from the historical operating feature set during the normal operating period of the HVAC system are extracted, and the external load demand features therein are used as input variables, with the actual total energy consumption value of the same period as the true label value.

[0090] Construct a joint loss function for iterative training of the model, wherein the joint loss function consists of a data fitting error term between the true label value and the predicted energy consumption value of the current iteration of the model and a thermodynamic energy conservation penalty term;

[0091] Specifically, joint loss function Specifically:

[0092] ;

[0093] In the formula, The joint loss function of the representation in the current training iteration (the smaller the value of this function, the more the model fits the historical real operation pattern and the more it conforms to the physical heat transfer baseline). The data fitting error term (i.e., mean square error, MSE) between the true label value and the predicted energy consumption value in the current iteration of the model. This represents the true label feature vector space composed of the actual total energy consumption of the HVAC system from the historical operating feature set; The feature vector space representing the predicted energy consumption value output by the model in the current iteration; This variable represents the traversal index of discrete timestamps or data samples participating in the training. Indicates the first The actual total energy consumption of the system during the same period at each timestamp (the unit is usually 1). or ), which are the objective values ​​collected by physical sensors; This indicates that the lightweight gradient boosting tree model is in the first... The predicted energy consumption value calculated in the current iteration at each timestamp; The thermodynamic energy conservation penalty term is implemented using a one-sided activation mechanism. The preset penalty weight coefficient (dimensionless, preferably ranging from 10 to 50 in this embodiment); Indicates that based on the target building in the first The dynamic heating and cooling loads at each time stamp and the theoretical maximum energy efficiency ratio are rigorously derived to form the physical lower limit of the theoretical minimum energy consumption for maintaining thermodynamic balance in a heating, ventilation, and air conditioning system.

[0094] In this embodiment, the construction steps of the thermodynamic energy conservation penalty term are as follows:

[0095] Extract the equivalent outdoor meteorological temperature and the real-time building pedestrian density data from the external load demand characteristics;

[0096] By introducing the comprehensive heat transfer coefficient of the building envelope and the average heat dissipation per person, and combining the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building, the dynamic cooling and heating load of the target building under the current working conditions is calculated.

[0097] In this embodiment, the dynamic heating and cooling load is specifically as follows: ;

[0098] In the formula, Indicates the target building at the current moment. Dynamic heating and cooling loads (in units) ); This indicates the discrete timestamp step size for real-time acquisition or iterative calculation. Represents the current time in the feature set. The equivalent outdoor meteorological temperature (in °C); The target building's exterior surface area (in units of) ), used to quantify the passive heat transfer area; The indoor air conditioning design control temperature (in °C) is usually set according to the building operation and management specifications (e.g., 24 °C or 26 °C under summer cooling conditions). Total floor area of ​​the building actually in operation (unit: ), used to quantify the area of ​​heat generated by personnel residence; Indicates the overall heat transfer coefficient of the target building envelope (unit: The project values ​​are assigned by looking up tables based on the main exterior material type of the target building. Specifically, if the exterior facade is primarily composed of modern low-emissivity double-glazed curtain walls... The value range is 1.5 to 2.5. If the exterior facade is mainly composed of standard concrete walls with insulation or aerated concrete blocks, The value range is 0.5 to 0.8. If the exterior facade is an old brick-concrete structure without insulation, The value range is 1.2 to 2.0. ); This represents the average heat dissipation per person (the range of values ​​is...). people); Represents the current time in the feature set. Real-time pedestrian density data for buildings (unit: people / ); is a dimensional conversion factor used to convert the heat unit watt within parentheses in an equation into kilowatt;

[0099] Based on the dynamic heating and cooling loads and the theoretical maximum energy efficiency ratio of the HVAC system, the theoretical minimum energy consumption physical lower limit for maintaining thermodynamic balance of the HVAC system is calculated. :

[0100] ;

[0101] In the formula, The theoretical maximum energy efficiency ratio (COP) of the target building's HVAC system (i.e., the system's COP limit value under the most ideal physical conditions, with a range of 5.5 to 6.5).

[0102] When the predicted energy consumption value of the current iteration of the lightweight gradient boosting tree model is lower than the theoretical minimum energy consumption physical lower limit, the thermodynamic energy conservation penalty term is activated, and a penalty calculation is applied to the predicted energy consumption value.

[0103] In this embodiment, the activation mechanism is specifically as follows: when the predicted energy consumption value of the lightweight gradient boosting tree model in the current iteration... Normal and above the physical lower limit When the penalty term is 0, the model fits the data normally; when the predicted energy consumption value is... Physical distortion occurs, meaning it falls below the theoretical minimum physical limit for energy consumption. When (e.g., when the predicted energy consumption is insufficient to offset physical heat transfer). The function outputs a positive value, activating the penalty term and applying a severe penalty calculation to the predicted energy consumption value.

[0104] During the training process of the lightweight gradient boosting tree model, the first and second derivatives of the joint loss function with respect to the predicted energy consumption value are calculated and substituted into the split gain calculation of the leaf nodes of the classification and regression trees.

[0105] By minimizing the joint loss function, the model's fitting and optimization process is strictly constrained by the physical boundary of the building's thermal balance, thereby obtaining the pre-trained lightweight gradient boosting tree model.

[0106] During the online prediction phase, the characteristics of the external load demand at the current moment are obtained in real time and input into the pre-trained lightweight gradient boosting tree model, which outputs the theoretical expected energy consumption value that satisfies the thermodynamic lower limit constraint.

[0107] In this embodiment, when encountering extreme weather conditions or sudden large passenger flows, the predicted values ​​are prone to violating the physical laws of heat transfer (for example, in the height of summer, the predicted theoretical cooling energy consumption may not even be sufficient to offset the heat transferred from outdoors to indoors), causing the subsequent residual-based fault diagnosis logic to completely collapse; moreover, the computing power of the smart city edge side (such as the building BMS gateway) is limited, making it impossible to deploy a large-scale deep learning model; therefore, this embodiment constructs a lightweight gradient boosting tree model and introduces a thermodynamic energy conservation penalty term into its underlying objective function, specifically:

[0108] Sample data from the historical operation feature set during the normal operation period of the HVAC system (i.e., the historical period without fault alarm tags) is extracted; the external load demand features (including equivalent outdoor meteorological temperature, real-time building pedestrian density data, and time representation features, etc.) in the feature set are used as the input variable vector. The actual total energy consumption of the system during the same period is used as the true label value. ;

[0109] Furthermore, this lightweight gradient boosting tree model is composed of If a classification and regression tree is composed of a sum of tree topologies and leaf node weights, then for any input variable vector... Predicted energy consumption The mathematical expression is:

[0110] ;

[0111] In the formula, For the first A classification and regression tree; Let be the function space containing all possible tree topologies and leaf node scores, i.e., the set space containing all possible classification and regression tree topologies (such as tree depth, splitting feature selection, splitting threshold) and their corresponding leaf node scores; constraints. This indicates that the new tree found by the model in each iteration is the optimal solution within the defined function space; Represents the input variable vector After the first After the topology of the classification and regression tree is split and mapped to nodes, the final weight score of the leaf node corresponding to the specific leaf node (i.e. the prediction output contribution value of a single tree) is determined. Indicates for the first The timestamp (or the first) (a discrete sample point), the predicted energy consumption value finally output by the model (in this invention, it is the theoretical expected energy consumption value for the HVAC system to maintain thermal balance under the current physical boundary conditions). Indicates the first In this embodiment, the input variable vector corresponding to each timestamp specifically refers to the external load demand feature constructed and indexed by the preceding steps (which encapsulates the equivalent outdoor weather temperature, real-time building pedestrian density data, and time representation features). This represents the total number of classification and regression trees (CART) contained in the lightweight gradient boosting tree model (i.e., the total number of base learners constructed through the additive model iteration). The traversal index variable represents the base learner, and its value range is... ;

[0112] Furthermore, in the specific training process of the lightweight gradient boosting tree model (such as the XGBoost architecture), the tree node splitting optimization of the model depends on the second-order Taylor expansion of the objective function, and the joint loss function is calculated with respect to the predicted energy consumption value. First derivative (gradient) ) and second derivative (Hessian matrix) );

[0113] When the penalty term is activated, the first and second derivatives will generate huge numerical feedback. The system substitutes the first and second derivatives into the standard split gain calculation formula for the leaf nodes of the classification and regression tree. The huge gradient feedback will force the split gain to become negative, thereby preventing the classification and regression tree from splitting and growing in a direction that violates the thermodynamic boundary.

[0114] Meanwhile, to ensure the lightweight nature of the gradient boosting tree model, a tree structure regularization term containing the number of leaf nodes and the L2 norm of the leaf node weights is further introduced into the objective function; through continuous iteration of the forward distribution algorithm, the model's fitting and optimization process is strictly constrained by the physical boundary of building thermal balance, and finally the pre-trained lightweight gradient boosting tree model is obtained.

[0115] In the online prediction phase, the external load demand characteristics at the current moment are acquired in real time and input into the pre-trained lightweight gradient boosting tree model. Through traversal and weight accumulation of each classification and regression tree, the theoretical expected energy consumption value that satisfies the thermodynamic lower limit constraint is output, providing a reliable baseline for subsequent residual abnormal decoupling.

[0116] S3. Collect the actual total energy consumption value of the HVAC system at the current moment, and calculate the total energy consumption residual between the actual total energy consumption value and the theoretical expected energy consumption value. Perform orthogonal decomposition on the total energy consumption residual to directly extract the equipment operation degradation component that characterizes the performance degradation of the equipment itself. The specific steps involved are as follows:

[0117] Obtain the actual total energy consumption value and the theoretical expected energy consumption value at the current moment, and calculate the difference to obtain the total energy consumption residual;

[0118] Based on a sliding analysis time window of a set length, the total energy consumption residual obtained by continuous calculation is cached in time series, and the total energy consumption residual within the sliding analysis time window is concatenated in time series to form a total energy consumption residual vector; at the same time, the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building corresponding to the sliding analysis time window are extracted and constructed as micro-meteorological disturbance basis vector and pedestrian surge disturbance basis vector, respectively.

[0119] The micro-meteorological disturbance basis vector and the population surge disturbance basis vector are orthogonalized to eliminate multicollinearity between features and construct an independent multidimensional disturbance feature space.

[0120] The total energy consumption residual vector is projected onto the multidimensional disturbance feature space to obtain the micro-meteorological disturbance projection vector mapped to the direction of the micro-meteorological disturbance basis vector, and the population surge disturbance projection vector mapped to the direction of the population surge disturbance after collinearity elimination.

[0121] Using vector subtraction, the micro-meteorological disturbance projection vector and the population surge disturbance projection vector are filtered and separated from the total energy consumption residual vector to obtain the equipment degradation residual vector;

[0122] Extract the last element scalar in the device degradation residual vector corresponding to the current moment, and use it as the device operation degradation component that purely represents the physical device degradation at the current moment.

[0123] In this embodiment, the total energy consumption residual contains both "pseudo-anomalies" caused by sudden changes in the external environment and "true anomalies" caused by actual equipment degradation. Since outdoor temperature and pedestrian density are often physically coupled (e.g., high temperatures often accompany a surge in customer traffic seeking respite from the heat in shopping malls), they must be separated through strict orthogonal decoupling in a multi-dimensional feature space. Specifically:

[0124] Set the length of the sliding analysis time window to Each sampling step size, at the current moment Extract consecutive cached data The total energy consumption residual scalars are concatenated to form a total energy consumption residual vector. (dimensions are) ), Indicates the current time, corresponding to the [number]th [time]. Each discrete sampling time; The length of the sliding analysis time window represents the number of consecutive sampling steps involved in the analysis. This represents the vector transpose operation; Indicates time The latest scalar residual data point calculated is placed at the end of the vector as the last element; This represents the initial scalar residual data point at the very front (i.e., the oldest) of the current sliding analysis time window in the cache sequence; Indicates the starting time index of the current sliding analysis time window; Indicates continuity The total energy consumption residual vector, constructed in chronological order from the total energy consumption residuals, has the following dimensions: It is used to characterize the temporal distribution of the total energy consumption residual within the sliding time window;

[0125] Similarly, the equivalent outdoor meteorological temperature sequence corresponding to the feature set is extracted to construct the micro-meteorological perturbation basis vector. (Dimensions are in °C); Extract real-time pedestrian density sequences from buildings to construct pedestrian surge disturbance basis vectors. (Dimensions are human) In the formula, Indicates in At this point, the equivalent outdoor meteorological temperature scalar is dynamically weighted and output after considering the building's heat storage hysteresis effect in the preceding steps. This indicates the starting time within the current sliding analysis time window (i.e., the earliest historical time furthest from the current time). ) Calculate the equivalent outdoor weather temperature scalar of the cache; Indicates the current time. (i.e., the latest last moment of the sliding analysis time window) Real-time building pedestrian density scalar (unit: person / This reflects the latest transient passenger flow load status inside the building at the current moment of judgment; Indicates the start time (time) within the current sliding analysis time window. Synchronously collect and cache historical pedestrian density scalars for buildings; This represents the basis vector of the surge in pedestrian flow (physical dimension: human / ) constructed synchronously within the same sliding analysis time window. In terms of data structure, it is also a dimension. The column vector encapsulates the temporal fluctuation characteristics of the passenger flow load inside the target building during that time period;

[0126] To construct mutually independent multidimensional perturbation feature spaces, for and Perform orthogonal normalization:

[0127] Obtain standardized micrometeorological unit direction vector In the formula, This represents the micro-meteorological unit direction vector output after standardization. This represents the micro-meteorological disturbance basis vector constructed within the current sliding analysis time window (composed of continuous data within the sliding analysis time window). (A column vector composed of equivalent outdoor meteorological temperature scalars concatenated in time sequence). Represents the original micro-meteorological disturbance basis vector The Euclidean norm;

[0128] Where, the vector is divided by its own Euclidean norm, making It is transformed into a purely dimensionless unit direction vector to ensure the dimensionality validity of subsequent inner product projection operations.

[0129] Furthermore, by removing components collinear with meteorological data from the pedestrian flow basis vector, a pure pedestrian flow unit direction vector is obtained. :

[0130] , ;

[0131] In the formula, This represents the orthogonal vector of the transitional human flow after removing components collinear with meteorological data; Represents the orthogonal vector of the transitional human flow. The Euclidean norm; This represents the pure passenger flow unit direction vector obtained after scalar division and normalization, which represents the pure passenger flow load fluctuation direction completely independent of external weather interference.

[0132] Thus, from the dimensionless and This constitutes an absolutely orthogonal two-dimensional basis feature space;

[0133] Furthermore, the total energy consumption residual vector is calculated. Projection onto this orthogonal space (an absolutely orthogonal two-dimensional basis feature space):

[0134] Among them, the micro-meteorological disturbance projection vector ;

[0135] Population surge disturbance projection vector ;

[0136] In the formula, Represents the total energy consumption residual vector In an absolutely orthogonal two-dimensional basis space, calculate the inner product projection scalar in the meteorological direction; Represents the total energy consumption residual vector In an absolutely orthogonal two-dimensional basis space, calculate the inner product projection scalar in the direction of the pure human flow; Represents the total energy consumption residual vector The micro-meteorological disturbance projection vector obtained by mapping (physical dimension is This represents the portion of the total energy consumption residual that is falsely high in energy consumption, which is purely caused by sudden changes in local weather (such as heavy rain and sudden cooling, or extreme high temperatures). This represents the projection vector of the surge in pedestrian flow disturbance obtained from the mapping (physical dimension is...). This represents the portion of the total energy consumption residual that is falsely high in energy consumption, which is purely caused by sudden surges in customer traffic (such as during shopping mall promotions or holidays).

[0137] Based on the principle of superposition in linear spaces, the aforementioned pseudo-external disturbances are filtered out using vector subtraction:

[0138] ;

[0139] In the formula, To obtain the total energy consumption residual After thoroughly filtering out the two pseudo-external disturbances mentioned above, the remaining residual vector of device degradation (due to...) and Dimensionless, this vector inherits the residual. physical dimensions Extract the vector corresponding to the current time step; Last element scalar This is the pure device operation degradation component that is output to subsequent warning logic;

[0140] Based on the above orthogonal decomposition steps, the equipment operation degradation component sequence during the healthy phase is calculated, and the mean of the equipment operation degradation component sequence is also calculated. with standard deviation Based on Gaussian distribution The criteria (i.e., the 99.73% confidence interval) are set as follows:

[0141] ;

[0142] In the formula, This represents the preset device health tolerance threshold (dimension: ). This serves as a rigid mathematical boundary for whether the system triggers a real abnormal alarm. Represents the probability density function of the Gaussian distribution. The criterion represents the confidence boundary of the maximum allowable health fluctuations of the system;

[0143] During the real-time monitoring phase, the device's operational degradation component is observed if and only if the currently extracted scalar value is... Within a continuously defined time window (e.g., 5 consecutive sampling periods), it is strictly greater than At that time, the system confirmed that the false high-energy-consumption white noise had been completely eliminated, triggered a real device-level abnormal energy consumption event, and entered the degradation parameter attribution process;

[0144] If and only if the equipment operation degradation component continuously exceeds the preset equipment health tolerance threshold within a set time window (the preset equipment health tolerance threshold is based on: extracting the time series of the equipment operation degradation component of the target building HVAC system during the historical absolute healthy operation phase, calculating its statistical mean and standard deviation, and setting the upper bound of the distribution under the preset confidence level as the preset equipment health tolerance threshold based on the confidence interval mechanism of normal distribution), it is determined that a real equipment-level abnormal energy consumption event is triggered.

[0145] Furthermore, it should be noted that although the lightweight gradient boosting tree model has incorporated equivalent outdoor weather temperature and real-time pedestrian density features into the input, under sudden conditions such as extreme weather changes or unprecedented pedestrian surges, the theoretical expected energy consumption value output by the model will still produce nonlinear fitting residuals due to limitations in the model's generalization ability and the transient acquisition lag of physical sensor data. Therefore, performing a second orthogonal projection of the total energy consumption residual onto the multidimensional perturbation feature space in step S3 is a second line of defense established at the underlying mathematical logic level. This step utilizes the strict geometric isolation characteristics of orthogonal decomposition to thoroughly filter out external white noise that the model fails to fully fit, ensuring that the final extracted residual vector of equipment degradation purely and uniquely reflects the deterioration of the physical entity.

[0146] S4. After triggering the real device-level abnormal power consumption event, extract the device status features within the trigger event time window;

[0147] A pre-defined degradation feature classification regression tree model is used to calculate the feature contribution of each equipment state feature to the equipment operation degradation component, and based on this, the core degradation parameters causing a sudden increase in system energy consumption are identified. The specific steps involved are as follows:

[0148] After the actual device-level abnormal power consumption event is triggered, the device state characteristics synchronously cached within the trigger event time window are obtained as the input set of independent variables to be attributed.

[0149] The pre-trained degradation feature classification and regression tree model is invoked. The degradation feature classification and regression tree model uses the equipment state features of the historical operation stage as input variables and the corresponding equipment operation degradation component as fitting labels to obtain a nonlinear mapping network.

[0150] The independent variable input set is input into the degradation feature classification regression tree model, and the SHAP interpretability algorithm based on game theory is introduced to calculate the marginal contribution value of each device state feature in the independent variable input set to the current output device operation degradation component, and aggregate to generate global feature contribution.

[0151] The device status features are sorted in descending order according to the contribution of the global features, and the device status features that meet the preset contribution threshold are used as the core degradation parameters, thereby generating an operation and maintenance alarm work order for the target building HVAC system.

[0152] Specifically, for any one of the data instances to be attributed within the time window that triggers a real device-level anomaly. Let its feature set be For the set of the first Individual device status characteristics;

[0153] The SHAP interpretability algorithm calculates the Shapley value by iterating through the marginal contributions of a feature across all possible combinations of feature subsets. :

[0154] ;

[0155] In the formula, This indicates the specific moment at which the exception was triggered (i.e., the specific data instance to be attributed). (Next, the first) The local Shapley marginal contribution of a specific equipment condition characteristic (such as a specific chilled water temperature difference or a specific pump pressure difference) to the final predicted equipment operation degradation component; This represents a specific data instance to be attributed (i.e., a feature vector containing all device state features at this moment) that triggers a real device-level abnormal power consumption event. The complete set of equipment state features (i.e., including all introduced dimensions such as temperature difference, pressure difference, approximation degree, frequency, etc.) is used to model the degradation feature classification regression tree model. Not including the first Any subset of features; This represents the total number of feature dimensions (i.e., the total number of features) contained in the complete set of the above-mentioned device status features. Represents the current feature subset The number of feature dimensions (i.e., the number of features) contained therein; This indicates that the degradation feature classification regression tree model only observes a subset. When considering the characteristics in the data, the conditional expected prediction value (i.e., the baseline degradation prediction value) for the equipment operation degradation component is obtained. Indicates from the complete set of features Remove the first The set of remaining features after identifying the target features; This represents the weighting coefficient for the combination and permutation, indicating the subset of features in all possible feature addition orders. Just in the feature The probability weights previously added to the model are used to ensure absolute fairness in the allocation of feature contributions; This indicates that the degraded feature classification regression tree model is in the input feature subset and jointly introduced the first When considering information about a target feature, for the current instance The joint expected forecast value made by the component of equipment operating decline;

[0156] Furthermore, after acquiring a partial view of all data instances to be attributed within the time window set by the triggering event... Then, the system aggregates and generates the first value by taking the absolute value of the Shapley values ​​of all sample points within the window and calculating the mean (or summation). Contribution of individual device status characteristics to global features in this abnormal event .

[0157] The SHAP interpretability algorithm guarantees that the sum of the Shapley values ​​of each feature is strictly equal to the difference between the current predicted value and the baseline expected value; subsequently, the system processes all sample points within the time window that triggers the anomaly. Calculate the mean to obtain the first value. Global feature contribution of individual device status characteristics .

[0158] After obtaining the global feature contribution of all device status features, sort them according to the global feature contribution. Sort the values ​​in descending order;

[0159] For the preset contribution threshold, this embodiment preferably uses the Pareto principle (80 / 20 rule) or a cumulative contribution rate cutoff mechanism for setting; specifically, the system calculates the top... Cumulative contribution rate of each feature In the formula, This indicates the top-ranked items after sorting by global feature contribution in descending order. The cumulative contribution rate (dimensionless percentage) of each device state characteristic to the current device operational degradation component reflects this... The weighting of each feature in explaining the cause of the abnormal decline). This indicates the number of head device status features selected by truncation (i.e., the number of features initially identified as suspected core degradation parameters), and its value changes dynamically until... Until the preset contribution threshold (e.g., 80%) is met; This represents the ranking index variable of the extracted features after the global feature contribution ranking is completed in descending order (the value range is...). ); This represents the total number of dimensions of the entire set of device state features participating in this attribution diagnosis; This indicates the rank in the descending sorted sequence. The global feature contribution of the device state characteristics of a bit; This represents a global index variable that iterates through all device state features involved in attribution (its value range is...). ); Represents the set of elements. The global feature contribution of each device's status characteristics;

[0160] The system gradually increases The value of (from) (Starting to increase), dynamically calculate the corresponding cumulative contribution rate. ;when When the contribution percentage is first greater than or equal to the preset contribution threshold (preferably set to 80% or 85%, conforming to the Pareto 80 / 20 rule), the system stops traversing and sets the corresponding top 20% contribution percentage at that time as the highest. The device status characteristics are formally locked and output as the core degradation parameters;

[0161] Specifically, when the cumulative contribution rate When the value first exceeds or equals a preset threshold (e.g., 80%), this will be... These features are identified as core degradation parameters. For example, if the attribution results show that the cumulative contribution of "cooling water proximity" and "cooling tower fan frequency" reaches 85%, the system will directly generate a targeted alarm work order, prompting maintenance personnel that "there is a real equipment anomaly, which is highly likely to be caused by severe scaling on the cooling tower fins or insufficient cooling water flow. Please prioritize checking the cooling-side equipment."

[0162] It is worth noting that in this embodiment, since the physical degradation data of the equipment (such as water pump pressure difference and cold water temperature difference) inevitably contains high-frequency white noise from the sensors, using a single decision tree is prone to overfitting. Therefore, the degradation feature classification regression tree model in this embodiment adopts a random forest regression tree architecture based on the bootstrap aggregation method. The specific construction and training steps are as follows:

[0163] Extract time-series data of the target building's HVAC system during its historical operation phase, construct the equipment state features at each historical sampling time as the input vector of independent variables, and use the scalar of the equipment operation degradation component corresponding to the synchronous operation after orthogonal decomposition as the true label of the dependent variable, and splice them together to form a global training set containing multiple sample pairs;

[0164] Set the total number of classification and regression trees in the random forest (set the random forest to consist of...). (Composition of a Classification and Regression Tree (CART))

[0165] Using a random resampling method with replacement, multiple subsets of training data with the same sample size as the total number of trees are extracted from the global training set. These subsets are then used to train the corresponding classification and regression trees (specifically, the process is executed...). Each draw is a random selection with replacement. Generate samples Individual Training Set ));

[0166] In the During the growth of a CART tree, assume that the current node contains a sample feature space of... The algorithm randomly selects a subset of device status feature dimensions to form a candidate feature subset, and iterates through each feature dimension in the candidate subset. and the potential segmentation threshold under this feature. , will space Divide into left child node space With right child node space ,in, This indicates a specific sample instance. In the The specific values ​​that can be taken in each feature dimension; Indicates the current feature dimension The threshold for a potential cutoff point being tested (i.e., a specific continuous value of the feature, such as "2.5℃" or "50kPa"). Representing the feature space Any input variable sample instance (i.e., a complete device status feature vector at a specific timestamp).

[0167] The optimal segmentation feature variable and its optimal segmentation point threshold are found by using the minimum square error criterion, so that the sum of the square errors of the device operation degradation component in the left child node space and the right child node space after the division and the mean of their respective spaces is minimized.

[0168] The node splitting and growth are recursively executed until the preset tree stopping condition is met (when the number of samples in the node is less than the preset threshold or the tree reaches the maximum depth), thus completing the construction of a single classification and regression tree; all the constructed classification and regression trees are integrated to form the pre-trained degraded feature classification and regression tree model, whose prediction mapping value for any input device state feature is the arithmetic mean of the independent prediction results of each classification and regression tree.

[0169] S5. Based on the identified core degradation parameters, match them with a preset operation and maintenance expert rule base to determine the corresponding specific fault type;

[0170] Generate an operation and maintenance alarm work order containing the fault location, the specific fault type, and handling suggestions, and push it to the smart city operation and maintenance terminal. The specific steps involved include:

[0171] Construct a pre-defined operation and maintenance expert rule base, which includes a multi-dimensional mapping matrix stored in the form of a trie or knowledge graph;

[0172] The mapping key values ​​of the multidimensional mapping matrix include at least: the physical label of the core degradation parameter, the deviation direction of the feature value, and the abnormal state of the combined associated parameters; the mapping target value of the multidimensional mapping matrix is ​​the corresponding specific fault type and processing suggestion.

[0173] Extract the physical tags (e.g., coolant proximity) of the core degradation parameters identified in step S4, and extract the continuous historical sequence of these parameters within a set time window that triggers the anomaly. Calculate the characteristic average value of the core degradation parameters within the set time window that triggers the actual device-level abnormal energy consumption event. And compare it with its baseline expected value during the historical period of absolute healthy operation. Determine the direction of deviation of the corresponding feature value (by calculating the difference). ,when ( When the tolerance threshold is used, a positive value indicates a higher tolerance, and vice versa.

[0174] The deviation direction of the physical label and the feature value is combined with the real-time deviation status of the combined associated parameters belonging to the same physical subsystem as the physical label, and then input into the operation and maintenance expert rule base for multi-dimensional mapping and matching to determine the specific fault type and the handling suggestion corresponding to the current target building HVAC system.

[0175] Extract the unique device identifier (such as the data point number in the BMS system) directly bound to the core degradation parameters. (or the MAC address of the device), spatial address addressing is performed in the device topology network of the building information model (BIM) or building automation system of the target building to obtain the spatial coordinates of the corresponding physical device node as the fault location (e.g., mapping the abstract point number back to the three-dimensional physical device node spatial coordinates in the BIM model). And a plain text description of the location (such as "Underground Level 2, Cold Station B, Cooling Tower"), and use it as the final location of the fault;

[0176] The fault location, the specific fault type, the handling suggestions, and the global feature contribution of the core degradation parameters are encapsulated into a structured message form of an operation and maintenance alarm work order according to a preset data communication protocol (such as MQTT or HTTP Restful API) and pushed to the smart city operation and maintenance terminal.

[0177] In this embodiment, the specific mapping logic of the multi-dimensional mapping matrix stored in the preset operation and maintenance expert rule base includes at least the following:

[0178] When the physical label of the extracted core degradation parameter includes cooling water approximation (difference between outlet water temperature and outdoor wet-bulb temperature), and its corresponding feature value deviates in a positive direction and is relatively high, and the feature value of the combined associated cooling water pump inlet and outlet pressure difference parameter increases abnormally, the specific fault type output by the multidimensional mapping matrix is ​​cooling-side heat exchange anomaly. The underlying physical fault source corresponding to this mapping is located as severe scaling of the cooling tower heat dissipation packing or clogging of the cooling water pipe network filter (specifically, after matching the matrix, the output specific fault type is "cooling-side heat exchange anomaly"; the physical fault source is precisely located as "severe scaling of the cooling tower heat dissipation packing or clogging of the cooling water pipe network filter"; the output treatment suggestion is "arrange chemical cleaning of the cooling tower packing and check the cooling water pump Y-type filter").

[0179] When the physical label of the extracted core degradation parameter includes the chilled water supply and return water temperature difference, and the corresponding feature value deviates negatively and is low, while the feature value of the associated chiller unit operating frequency parameter deviates positively and is high, the specific fault type output by the multidimensional mapping matrix is ​​"high flow rate, low temperature difference syndrome". The underlying physical fault source corresponding to this mapping is located as severe dust accumulation on the surface cooler of the terminal air handling unit or a malfunctioning two-way regulating valve on the water side (specifically, after matching the matrix, the specific fault type output is "high flow rate, low temperature difference syndrome"; the physical fault source is precisely located as "dust accumulation on the surface cooler of the terminal AHU / FCU or a malfunctioning two-way regulating valve"; the output processing suggestion is: "Investigate the filter of the terminal air handling unit in the key defense area for dirt and blockage, and whether the mechanical actuator of the two-way valve is stuck").

[0180] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying abnormal energy consumption and generating operation and maintenance alarms for HVAC systems in smart cities, characterized in that, Includes the following steps: S1. Collect real-time operating data of the target building's HVAC system, as well as external environmental disturbance data including micro-meteorological data and real-time building pedestrian density data, and perform time-series correction on the micro-meteorological data to obtain the equivalent outdoor meteorological temperature. Specifically, the real-time operating data, the equivalent outdoor meteorological temperature, and the real-time building pedestrian density data are preprocessed to extract equipment status features and external load demand features, respectively. The equipment status features are then cached in a time series, and a historical operating feature set is constructed accordingly. S2. Real-time acquisition of the external load demand characteristics at the current moment, and inputting them into a pre-trained lightweight gradient boosting tree model to calculate the theoretical expected energy consumption value for maintaining thermal balance of the HVAC system. S3. Collect the actual total energy consumption value of the HVAC system at the current moment, and calculate the total energy consumption residual between the actual total energy consumption value and the theoretical expected energy consumption value. Perform orthogonal decomposition on the total energy consumption residual to directly extract the equipment operation degradation component that characterizes the performance degradation of the equipment itself. A real device-level abnormal energy consumption event is determined to be triggered only when the continuous set time window of the device operation degradation component exceeds the preset device health tolerance threshold. S4. After triggering the real device-level abnormal power consumption event, extract the device status features within the trigger event time window; Using a pre-defined degradation feature classification regression tree model, the feature contribution of each device state feature to the device operation degradation component is calculated, and the core degradation parameters that cause a sudden increase in system energy consumption are identified accordingly. S5. Based on the identified core degradation parameters, match them with a preset operation and maintenance expert rule base to determine the corresponding specific fault type; An operation and maintenance alarm work order containing the fault location, the specific fault type, and handling suggestions is generated and pushed to the smart city operation and maintenance terminal.

2. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, The real-time operating data includes at least the real-time power of the chiller unit, chilled water supply / return temperature, chilled water supply / return pressure, cooling water supply / return temperature, cooling pump operating frequency, chilled water pump operating frequency, and cooling tower fan operating frequency. Micrometeorological data should include at least outdoor dry-bulb temperature, solar radiation intensity, and real-time outdoor wind speed; Real-time pedestrian density data for buildings should include at least the real-time pedestrian density index.

3. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, The specific steps involved in performing time-series correction on the micrometeorological data are as follows: Obtain the current time from micrometeorological data. outdoor dry bulb temperature and solar radiation intensity And the solar radiation absorption coefficient of the target building envelope is introduced. and outdoor surface heat transfer coefficient Calculate the current outdoor temperature. ; Based on the combined outdoor temperatures at the current and historical times, a thermal resistance-capacitance hysteresis time constant for the building envelope is introduced. and data sampling period Dynamic weighted calculation of phase lag and amplitude attenuation is performed to output the equivalent outdoor meteorological temperature. .

4. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, The construction and training steps of the lightweight gradient boosting tree model are as follows: A lightweight gradient boosting tree model is constructed based on the additive model and the forward distribution algorithm. This lightweight gradient boosting tree model is composed of... The classification and regression trees are composed of the sum of the tree topology and the weights of the leaf nodes; Sample data from the historical operating feature set during the normal operating period of the HVAC system are extracted, and the external load demand features therein are used as input variables, with the actual total energy consumption value of the same period as the true label value. Construct a joint loss function for iterative training of the model, wherein the joint loss function consists of a data fitting error term between the true label value and the predicted energy consumption value of the current iteration of the model and a thermodynamic energy conservation penalty term; During the training process of the lightweight gradient boosting tree model, the first and second derivatives of the joint loss function with respect to the predicted energy consumption value are calculated and substituted into the split gain calculation of the leaf nodes of the classification and regression trees. The pre-trained lightweight gradient boosting tree model is obtained by minimizing the joint loss function. During the online prediction phase, the external load demand characteristics at the current moment are acquired in real time and input into the pre-trained lightweight gradient boosting tree model to output the theoretical expected energy consumption value.

5. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 4, characterized in that, The specific steps for constructing the thermodynamic energy conservation penalty term are as follows: Extract the equivalent outdoor meteorological temperature and the real-time building pedestrian density data from the external load demand characteristics; By introducing the comprehensive heat transfer coefficient of the building envelope and the average heat dissipation per person, and combining the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building, the dynamic cooling and heating load of the target building under the current working conditions is calculated. Based on the dynamic heating and cooling loads and the theoretical maximum energy efficiency ratio of the HVAC system, the theoretical minimum energy consumption physical lower limit for maintaining thermodynamic balance of the HVAC system is calculated. When the predicted energy consumption value of the current iteration of the lightweight gradient boosting tree model is lower than the theoretical minimum energy consumption physical lower limit, the thermodynamic energy conservation penalty term is activated, and a penalty calculation is applied to the predicted energy consumption value.

6. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, In step S3, the equipment operation degradation component, which characterizes the performance degradation of the equipment itself, is extracted. The specific steps involved are as follows: Obtain the actual total energy consumption value and the theoretical expected energy consumption value at the current moment, and calculate the difference to obtain the total energy consumption residual; Based on a sliding analysis time window of a set length, the total energy consumption residual obtained by continuous calculation is cached in time series, and the total energy consumption residual within the sliding analysis time window is concatenated in time series to form a total energy consumption residual vector; at the same time, the equivalent outdoor meteorological temperature and the real-time pedestrian density data of the building corresponding to the sliding analysis time window are extracted and constructed as micro-meteorological disturbance basis vector and pedestrian surge disturbance basis vector, respectively. The micro-meteorological disturbance basis vector and the population surge disturbance basis vector are orthogonalized to construct a multi-dimensional disturbance feature space; The total energy consumption residual vector is projected onto the multidimensional disturbance feature space to obtain the micro-meteorological disturbance projection vector mapped to the direction of the micro-meteorological disturbance basis vector, and the population surge disturbance projection vector mapped to the direction of the population surge disturbance after collinearity elimination. Using vector subtraction, the micro-meteorological disturbance projection vector and the population surge disturbance projection vector are filtered and separated from the total energy consumption residual vector to obtain the equipment degradation residual vector; Extract the last element scalar in the device decay residual vector corresponding to the current moment, and use it as the device operation decay component at the current moment.

7. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, The specific steps involved in identifying the core degradation parameters that cause a sudden increase in system energy consumption are as follows: After triggering the real device-level abnormal power consumption event, the device state characteristics synchronously cached within the trigger event time window are obtained as the input set of independent variables; The pre-trained degradation feature classification and regression tree model is invoked. The degradation feature classification and regression tree model uses the equipment state features of the historical operation stage as input variables and the corresponding equipment operation degradation component as fitting labels to obtain a nonlinear mapping network. The independent variable input set is input into the degradation feature classification regression tree model, and the SHAP interpretability algorithm based on game theory is introduced to calculate the marginal contribution value of each device state feature in the independent variable input set to the current output device operation degradation component, and aggregate to generate global feature contribution. The device status features are sorted in descending order according to the contribution of the global features, and the device status features that meet the preset contribution threshold are used as the core degradation parameters, thereby generating an operation and maintenance alarm work order for the target building HVAC system.

8. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 1, characterized in that, In S5, the specific steps for generating an operation and maintenance alarm work order include: Construct a pre-defined operation and maintenance expert rule base, which includes a multi-dimensional mapping matrix; The mapping key values ​​of the multidimensional mapping matrix include at least: the physical label of the core degradation parameter, the deviation direction of the feature value, and the abnormal state of the combined associated parameters; the mapping target value of the multidimensional mapping matrix is ​​the corresponding specific fault type and processing suggestion. Extract the physical tags of the core degradation parameters identified in step S4, calculate the average value of the core degradation parameters within the set time window that triggers the real device-level abnormal energy consumption event, and compare it with the baseline expected value in the historical absolute healthy operation phase to determine the corresponding deviation direction of the feature value. The deviation direction of the physical label and the feature value is combined with the real-time deviation status of the combined associated parameters belonging to the same physical subsystem as the physical label, and then input into the operation and maintenance expert rule base for multi-dimensional mapping and matching to determine the specific fault type and the handling suggestion corresponding to the current target building HVAC system. Extract the unique device identifier code that is directly bound to the core degradation parameter, perform spatial address addressing in the device topology network, and obtain the corresponding physical device node spatial coordinates as the fault location; The fault location, the specific fault type, the handling suggestions, and the global feature contribution of the core degradation parameters are encapsulated into a structured message form of an operation and maintenance alarm work order and pushed to the smart city operation and maintenance terminal.

9. The method for identifying abnormal energy consumption and providing operation and maintenance alarms for HVAC systems in smart cities according to claim 8, characterized in that, The specific mapping logic of the multidimensional mapping matrix stored in the preset operation and maintenance expert rule base includes at least the following: When the physical label of the extracted core degradation parameter includes cooling water approximation degree, and the corresponding feature value deviates in the positive direction and is relatively high, and the feature value of the combined associated cooling water pump inlet and outlet pressure difference parameter increases abnormally, the specific fault type output by the multidimensional mapping matrix is ​​cooling side heat exchange abnormality, and the underlying physical fault source corresponding to the mapping is located as severe scaling of cooling tower heat dissipation packing or blockage of cooling water pipe network filter. When the physical label of the extracted core degradation parameter includes the temperature difference between chilled water supply and return water, and the corresponding feature value deviates negatively and is low, while the feature value of the associated chiller unit operating frequency parameter deviates positively and is high, the specific fault type output by the multidimensional mapping matrix is ​​a high flow rate and low temperature difference syndrome. The underlying physical fault source corresponding to this mapping is located as severe dust accumulation on the surface cooler of the terminal air handling unit or a malfunctioning water-side two-way regulating valve.

Images