AI-based property energy consumption anomaly diagnosis method and system

Abstract

The application provides an AI-based property energy consumption anomaly diagnosis method and system, which comprises the following steps: constructing an operation state feature set of a target property building; inputting the operation state feature set into a preset sequence neural network model to generate a corresponding baseline energy consumption central value and fluctuation width corresponding to a corresponding time slice; comparing the actual energy consumption result of a current time slice with the baseline energy consumption central value and fluctuation width corresponding to the corresponding time slice to generate an energy consumption deviation degree, combining a device control strategy execution state to perform anomaly accumulation score calculation, and identifying a persistent energy consumption anomaly event; and analyzing the control strategy execution state in the time section of the identified energy consumption anomaly event to generate a diagnosis conclusion and a processing suggestion. The application improves the accuracy of anomaly judgment and the operability of the result, and is suitable for complex, multi-strategy parallel property building energy consumption management scenarios.

Description

Technical Field

[0001] This invention belongs to the field of abnormal energy consumption diagnosis in properties, and particularly relates to an AI-based method and system for abnormal energy consumption diagnosis in properties. Background Technology

[0002] In modern urban property management, centralized monitoring and digital data collection of building energy consumption have become standard practice. Major energy-consuming systems such as air conditioning and lighting are typically connected to a unified energy management and building control platform, generating continuous energy consumption data and control status records. However, in actual operation, accurately identifying energy consumption anomalies and determining their causes based on this existing data remains a prominent challenge in property operation and maintenance. Existing energy consumption anomaly detection methods are mostly based on fixed thresholds or simple statistical models, focusing only on changes in the energy consumption values ​​themselves. They fail to consider the building's operating conditions within specific time segments, leading to frequent misjudgments of reasonable energy consumption fluctuations as anomalies under conditions such as environmental changes, strategy switching, or operational mode adjustments. This seriously affects the reliability and usability of the system.

[0003] On the other hand, even if anomalies can be identified, existing systems typically only provide the time range or approximate equipment range of the anomaly, failing to further explain the relationship between the anomaly and the execution of control policies. This makes it difficult for maintenance personnel to quickly locate the root cause of the problem. Especially in residential buildings, energy consumption levels are often directly affected by the execution status of control policies, such as equipment start-up and shutdown plans, operating permit status, and policy switching times. These factors determine the reasonable range of energy consumption over different time periods, but existing methods generally fail to incorporate the execution semantics of control policies into the anomaly judgment and diagnosis process. This results in anomaly analysis remaining at the data level, lacking the ability to interpret data consistent with actual operational logic.

[0004] Therefore, without introducing additional complex data sources, how to combine existing operational status and control strategy information to construct dynamic energy consumption references and achieve interpretable anomaly diagnosis has become a pressing technical problem in the field of property energy management. Summary of the Invention

[0005] The purpose of this invention is to propose an AI-based method and system for diagnosing abnormal energy consumption in properties, thereby solving the aforementioned problems.

[0006] To achieve the above objectives, a first aspect of the present invention provides an AI-based method for diagnosing abnormal energy consumption in properties, the method comprising the following steps: The system acquires environmental status information, equipment control strategy execution status, and actual energy consumption results of the target property building within a continuous time slice, performs time alignment, and constructs a set of operational status features. The operating state features of each time slice in the set of operating state features are input into a preset sequence neural network model, so that the sequence neural network model generates the baseline energy consumption center value and fluctuation width of the corresponding time slice item by item according to the operating state features of each time slice; the sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on actual energy consumption results, indications of whether the execution state of the equipment control strategy has switched, baseline energy consumption center value and fluctuation width; The actual energy consumption result of the current time slice is compared with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation. Based on the energy consumption deviation, an abnormal cumulative score is calculated in combination with the execution status of the equipment control strategy to identify continuous energy consumption abnormal events. For the identified energy consumption anomalies, the execution status of the control strategy within the time period is analyzed, and diagnostic conclusions and handling suggestions are generated to explain the correlation between the anomalies and the control strategy.

[0007] Furthermore, the sequence neural network model consists of a feature projection layer, a GRU temporal modeling layer, and a dual-output head, including: The feature projection layer feeds the running state features of each time slice into a fully connected linear transformation layer and combines it with nonlinear activation to map the environmental state information and the execution state of the device control strategy to a unified latent space. The GRU timing modeling layer uses a gated loop unit to process data of length... The sequence is encoded to obtain the hidden state that reflects the evolution within the window; The dual output head includes one output head that generates the baseline energy consumption center value and another output head that generates the fluctuation width related to the operating state.

[0008] Furthermore, the Softplus function is used at the end of the fluctuation width output head to ensure that the output is positive, which is used to construct the baseline interval.

[0009] Furthermore, the indicator of whether the device control strategy execution state has switched is obtained by comparing the device control strategy execution state in the operating state characteristics of adjacent time windows: when the control strategy execution state changes, the indicator is taken as follows: Take when it remains unchanged .

[0010] Furthermore, during training, the loss function of the sequence neural network model includes a policy-switching-aware smoothing constraint term. This smoothing constraint term imposes a stronger smoothness requirement on time slices where the execution state of the device control policy remains unchanged, and relaxes the smoothing constraint near time slices where the execution state of the device control policy changes.

[0011] Furthermore, the energy consumption deviation is calculated by taking the absolute value of the difference between the actual energy consumption and the baseline center value, and then combining it with the fluctuation width.

[0012] Furthermore, in the calculation of the abnormal cumulative score: the value of the abnormal cumulative score is updated in each time slice. When updating, the abnormal cumulative score of the previous time slice is decayed and the energy consumption deviation of the current time slice is added. If the indicator of whether the execution state of the device control strategy has changed in the current time slice changes, the energy consumption deviation is suppressed or cleared to zero.

[0013] Furthermore, the identification of persistent energy consumption anomalies includes: When the cumulative abnormal score exceeds a preset trigger threshold within multiple consecutive time slices, the energy consumption abnormal event is determined to begin; when the cumulative abnormal score falls back below the release threshold, the energy consumption abnormal event is determined to end.

[0014] Furthermore, the generation of diagnostic conclusions includes: The execution status of the device control strategy that occurs most frequently within the time period of the energy consumption anomaly event is extracted as the dominant strategy. Based on the dominant strategy, a consistency score is calculated by combining the time segment of energy consumption anomaly events, the current execution status of the equipment control strategy, and the degree of energy consumption deviation. Based on the consistency score, energy consumption anomalies are categorized into predefined explanatory categories; wherein, the content of the explanatory category involves at least one of control strategy configuration review, equipment operation status check, or strategy execution effectiveness verification.

[0015] In a second aspect of the invention, an AI-based property energy consumption anomaly diagnosis system is provided, the system comprising: The property operation status acquisition unit is used to acquire environmental status information, equipment control strategy execution status and actual energy consumption results of the target property building in a continuous time slice, and perform time alignment to construct a set of operation status features. The state analysis unit is used to input the operating state features of each time slice of the operating state feature set into a preset sequence neural network model, so that the sequence neural network model generates the corresponding baseline energy consumption center value and fluctuation width for each time slice according to the operating state features of each time slice. The sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on actual energy consumption results, indications of whether the execution state of the equipment control strategy has switched, baseline energy consumption center value and fluctuation width. An anomaly determination unit is used to compare the actual energy consumption result of the current time slice with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation; based on the energy consumption deviation, combined with the execution status of the equipment control strategy, an anomaly cumulative score is calculated to identify continuous energy consumption anomaly events. The suggestion generation unit is used to analyze the execution status of the control strategy within the time period of the identified energy consumption anomaly event, and generate diagnostic conclusions and handling suggestions that explain the correlation between the anomaly and the control strategy.

[0016] The beneficial technical effects of the present invention are at least as follows: This invention establishes a dynamic energy consumption baseline that changes with the operating state of a building within each time slice by uniformly modeling the building's operational status. This is achieved by incorporating environmental conditions, control strategy execution status, and actual energy consumption results into a single analytical framework. Instead of simply comparing energy consumption values, this invention uses the dynamic baseline to characterize reasonable energy consumption ranges under different operating conditions. By combining temporal continuity and control strategy switching characteristics, it identifies anomalies in energy consumption behavior that continuously deviates from the reasonable range. Furthermore, this invention performs semantic correlation analysis between abnormal events and the control strategy execution status during their occurrence, transforming anomalies from simple numerical deviations into operational problem explanations corresponding to specific control strategy behaviors, and generating processing suggestions that can be directly used for operation and maintenance decisions. Through this diagnostic mechanism based on operational status and centered on control strategy semantics, this invention achieves a complete technical path from anomaly identification to cause explanation and handling suggestions. This ensures that the energy consumption anomaly diagnosis results are consistent with the actual operational logic of the property, improving the accuracy and operability of anomaly judgments, and is applicable to complex, multi-strategy parallel energy consumption management scenarios in property buildings. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a flowchart of the AI-based property energy consumption anomaly diagnosis method of the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] like Figure 1As shown in the embodiment of the present invention, an AI-based method for diagnosing abnormal energy consumption in properties includes: S1. Obtain the environmental status information, equipment control strategy execution status and actual energy consumption results of the target property building within a continuous time slice, and perform time alignment to construct a set of operating status features.

[0021] Specifically, this step revolves around the fundamental issue of diagnosing property energy consumption anomalies. Before any anomaly judgment or diagnosis, the system must be able to reconstruct the property's true operational status within a specific time slice in an engineering-feasible manner. Here, "operational status" is not an abstract business concept, but rather a result determined by data already existing in the property system and capable of being stably collected. The goal of this step is to organize this information, scattered across different systems, into a strictly time-aligned and semantically consistent description of the operational status through a clear data acquisition path and processing logic. This provides reliable input for subsequent model-based energy consumption baseline construction. In actual deployments, energy consumption data is typically collected uniformly by an energy management system. This system continuously monitors major energy-consuming objects through sub-meters, loop collectors, or smart energy terminals. Taking an office building scenario as an example, air conditioning, public lighting, and elevator systems often correspond to different collection loops, with each loop generating an energy consumption record at preset time intervals. This step directly uses these raw collection results without cross-loop merging or complex statistics, only requiring each record to have a clear time identifier for subsequent alignment with other data.

[0022] Furthermore, the control strategy execution data comes from the operational layer interface of the building control system. In actual engineering, control strategies are often configured in the form of schedules or rules, but these configurations do not directly reflect the actual operational permission status of equipment within a certain time slice. Therefore, this step does not read static configuration files, but instead reads the status records generated by the control system during operation, such as whether the equipment is allowed to start within a certain time period or whether it is within the permitted operational range. This type of information can usually be obtained through the real-time status points or operational logs of the building control system. For each type of major energy-consuming equipment, this step extracts a corresponding control status within each time slice to indicate whether the equipment is in an operable state within that time slice, thereby avoiding the mistaken confusion between energy consumption generated under "strategy not allowing operation" conditions and normal operating energy consumption in subsequent analysis.

[0023] External environmental status data is provided by the meteorological service accessed by the property management system. In engineering implementation, this type of data is typically updated at fixed time intervals and includes key indicators reflecting the changing trends of the building's external environment. This step does not perform spatial interpolation or prediction processing on the environmental data; instead, it directly maps it to the same time slice as energy consumption and control status according to the timestamp. If there is a difference between the update granularity of the environmental data and the energy consumption collection granularity, alignment is achieved using the nearest time slice matching method to ensure that environmental status, control status, and energy consumption results can be referenced simultaneously within each analysis time slice.

[0024] Further, after completing data acquisition and time alignment, this step proceeds to the runtime state construction phase. For each time slice... The system first verifies whether the control strategy execution status within the time slice is complete. If the control status of a certain type of critical equipment is missing, the time slice will not participate in subsequent modeling to avoid introducing uncertain operating conditions. Subsequently, the environmental state, control strategy execution status, and actual energy consumption results within the time slice are combined to form a complete operational status description. This combination is not a simple splicing, but emphasizes the consistency of the three types of information in terms of time and semantics. That is, they jointly describe "at the same moment, the property operates according to a predetermined control strategy under certain environmental conditions and produces corresponding energy consumption results."

[0025] The above operating states are formally represented as a state vector: ; in, Indicates time slice The running status; This indicates that the data was obtained from the meteorological interface and compared with the time slice. Alignment of environmental status information; Indicates time slice The execution status of equipment control strategies is obtained by parsing the building control system's operation records. Indicates time slice The actual energy consumption results are collected by the energy management system. This state vector is used to fully represent the property's energy consumption in a given time slice. The system records the actual operation of the system without assuming any anomaly detection or prediction capabilities. To further illustrate the feasibility of this process, consider the following scenario: In the morning hours of an office building, the control system records that the air conditioning system has entered a permissible operating state, the environmental status shows that external conditions are at a high load level, and the energy management system records that air conditioning energy consumption has increased significantly. The final output of this step is a set of operating status characteristics arranged in chronological order, which consists of multiple... It is composed of elements and stored in time series format.

[0026] S2. Input the operating state features of each time slice of the operating state feature set into a preset sequence neural network model, so that the sequence neural network model generates the baseline energy consumption center value and fluctuation width of the corresponding time slice according to the operating state features of each time slice; the sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on the actual energy consumption results, the indication of whether the execution state of the equipment control strategy has switched, the baseline energy consumption center value and the fluctuation width.

[0027] Specifically, this step establishes a dynamically changing energy consumption baseline based on the operational state feature set formed in the previous step. The operational state feature set binds the environmental state, control strategy execution state, and energy consumption results within the same time slice to the same semantic coordinate system, enabling the model to learn "what range of reasonable energy consumption typically falls under certain environmental conditions and control strategy execution states." Therefore, the core work of this step is to transform the time-series operational state feature set into a dynamic energy consumption baseline model that can be used for inference, and to ensure that the model output provides both the center level of the baseline energy consumption and the fluctuation width matching the operational state, thus providing a direct reference for anomaly detection in the next step. The input to this step is the operational state feature set output from the previous step, i.e., the operational state features arranged in chronological order. In engineering implementation, The environmental status and control strategy execution status are obtained from continuous data collection records from the property's existing systems, while the energy consumption results are obtained from the sub-item data collection records from the energy management system. All three are aligned by time slice and organized using the same time index. This step uses time slices as the basic sample unit, employing historical time slice sequences during the training phase and real-time or near-real-time rolling time slice sequences during the inference phase.

[0028] Furthermore, the model's input is not a state vector at a single moment, but a fixed-length time window to capture common "delay effects" and "policy switching effects" in property operations. For example, when an air conditioning policy switches from permitted operation to restricted operation, energy consumption does not drop instantly; there is often a transition period. Similarly, the impact of changes in the external environment on energy consumption may also be delayed. Therefore, in each time slice... Model reading from arrive The continuous state vector is used as an input sample, and the reasonable energy consumption level and fluctuation range at that moment are inferred by combining the environmental state change trajectory within the window with the control strategy execution trajectory.

[0029] The model structure employs a deployable sequence network, explicitly injecting the control policy execution state as the dominant signal for "runtime semantics" into the network. The specific implementation can be composed of three concatenated parts. The first part is the feature projection layer: which projects the features of each time slice... A fully connected linear transformation layer, combined with nonlinear activation, maps the environmental state and the control strategy execution state to a unified latent space. The control strategy execution state, in engineering, often originates from discrete state points or planned execution flags of the building control system. This can be converted into a continuous vector through embedding mapping before entering the fully connected layer, and then concatenated with the environmental state vector. The second part is the timing modeling layer: a gated recurrent unit (GRU) layer is used to model the sequence of elements with a length of... The sequence is encoded to obtain the hidden state reflecting the operational evolution within the window. The third part is a dual-output head: one output head generates the baseline energy consumption center value, and the other output head generates the fluctuation width related to the operational state; the fluctuation width output head uses the Softplus function at the end to ensure that the output is positive, so that it can be directly used to construct the baseline interval. The creativity of this structure lies in taking the "strategy execution state" as an important condition for sequence encoding, so that the model learns not a pure numerical trend, but a baseline expression consistent with the actual control logic of the property, which is particularly suitable for the operational characteristics of "strategy planning determines the feasible energy consumption interval" in property scenarios.

[0030] Furthermore, the model in time slices The inference output format is as follows: ; in, This indicates that the sequence neural network model constructed in this step consists of a feature projection layer, a GRU temporal modeling layer, and a dual output head; Indicates length is The sequence of running state windows is taken item by item from the input set of running state features; Indicates time slice The corresponding baseline energy consumption center value; Indicates time slice The corresponding fluctuation width is used to construct the baseline interval. The window length is set during system deployment based on the data collection granularity and typical latency. For example, in common property operations, it can cover several consecutive time slices to reflect the transition period of strategy switching.

[0031] Furthermore, to ensure the baseline both closely reflects operational status and maintains temporal continuity consistent with common property management practices, this step introduces a "smooth constraint for strategy switching awareness" into the training objective. A stable pattern in property management scenarios is that within a continuous timeframe where the control strategy execution state remains unchanged, the reasonable energy consumption baseline typically exhibits a smooth change; however, near the point where the control strategy execution state switches, the baseline is allowed to show more significant jumps or slope changes. Therefore, during training, not only are historical energy consumption results fitted, but the temporal continuity of the baseline is also constrained, and different constraint strengths are applied to the "remain unchanged" and "switching" time slices using the strategy switching indicator. The training objective can be written as: ; in, Indicate the training objective; Indicates time slice The actual energy consumption results come directly from the input sequence. The corresponding energy consumption component in the result; and The model output is derived from the previous formula; With denominator Together, they create a constraint relationship between the model's fitting error and the fluctuation width, preventing the fitting deviation from being masked simply by increasing the fluctuation width; This represents the smoothing constraint strength coefficient, which is set by the training configuration. Indicates the execution status of the control strategy. arrive An indicator of whether a switch has occurred, it is determined by... and The comparison of the control policy execution state components yields the following result: when the control policy execution state changes, the following is taken: Take when it remains unchanged This ensures that the baseline center value remains constant over a period of time. It tends to evolve smoothly; near the time slice of strategy switching, the smoothing constraints are naturally relaxed, allowing the baseline to undergo reasonable structural changes as the strategy changes, which is more in line with the real laws of property operation.

[0032] In practical engineering, the construction method of the training samples closely corresponds to the operational rhythm of the property project. The system extracts a length of [length missing] from the historical operational status feature set by sliding over time. The window is used, and the actual energy consumption result of the time slice at the end of the window serves as the supervision signal. To reduce the interference of a single abnormal event on baseline learning, time slices corresponding to confirmed power outages, maintenance, and other atypical operating days can be removed during training data selection according to property operation and maintenance routines, or the weight of samples with extreme deviations can be reduced; these operations only involve the selection strategy of training samples, do not change the input and output definitions, and do not introduce additional data types. During the inference phase, the latest operating status sequence is input in a rolling window manner, and the data is output in real time. and This yields the baseline interval that varies over time. This step outputs two results. The first is the dynamic energy consumption baseline model. Its parameters are obtained through training on a set of historical operational state features and then deployed in a fixed manner; the second term is for any given time slice. The model outputs baseline results based on the input window. This result is used to compare with the actual energy consumption results of the same time slice in the next step to identify anomalous behavior that deviates from the baseline range.

[0033] S3. Compare the actual energy consumption result of the current time slice with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation. Based on the energy consumption deviation, calculate the abnormal cumulative score in combination with the execution status of the equipment control strategy to identify continuous energy consumption abnormal events.

[0034] Specifically, based on the dynamic energy consumption baseline results obtained in the previous step, this step determines the real-time energy consumption behavior and generates abnormal events that can be used for subsequent diagnosis. The baseline results output in the previous step... The "reasonable energy consumption level and its fluctuation range under a given operating condition" has already been quantified. This step uses this as a reference to set the operating condition characteristics. Actual energy consumption results recorded in Mapped as a continuous "deviation trajectory," and combined with the reality that anomalies in property operations are usually persistent and strongly correlated with control strategy switching points, an anomaly determination mechanism with engineering operability is formed. This mechanism outputs the anomaly time segment and retains key context related to the execution status of the control strategy, enabling the next step to directly explain and generate suggestions based on "what the strategy was doing when the anomaly occurred."

[0035] Furthermore, in practical implementation, this step first converts the "difference between actual energy consumption and baseline center value" into comparable deviations under different operating conditions. Because... The reasonable fluctuation range under this operating state has already been expressed. This step adopts the following approach: The deviation intensity is described by a normalized deviation on a given scale, thus ensuring that the meaning of the deviation remains consistent under different environmental conditions such as high load in summer and low load in spring and autumn. The deviation is defined as follows: ; in, Indicates time slice Energy consumption deviation; For time slices The actual energy consumption results come from Energy consumption component; The baseline center value is derived from the baseline model. Time slice The inference output; The fluctuation width, derived from the same inference output, is used to express the normal fluctuation scale under this operating state. (By...) This step unifies the energy consumption differences under different operating conditions into a discernible deviation coordinate, providing a basis for subsequent section determination.

[0036] Furthermore, after obtaining the deviation sequence, this step, based on the actual needs of property operation and maintenance, defines anomalies as "continuous and directional deviations" and introduces a "strategy switching sensitive" segment determination. A common pattern in property scenarios is that control strategies are allowed to exhibit brief transitional deviations near switching points, while when the control strategy remains stable, continuous deviations are more likely to correspond to abnormal equipment operation, unreasonable strategy configuration, or abnormal control execution. Therefore, this step... Construct a policy switching indicator (an indicator of whether the execution status of the device control policy has changed). This is used as a weighting factor in anomaly detection, reducing the probability of momentary misjudgment near policy switching points and enhancing sensitivity to persistent deviations during policy stabilization periods. Specifically, a "cumulative score with policy weights" is applied to the deviation, creating triggerable anomaly segment scores over time. ; in, Indicates the deadline. Abnormal cumulative scores; This is the cumulative rating for the previous time slot; This is the attenuation coefficient, used to preserve the effects of persistent deviations over time and naturally reduce transient fluctuations; Indicates the execution status of the control strategy. arrive Whether it has changed depends on and The comparison yields the result; when a change occurs, the value is taken. Take when it remains unchanged ; This represents the deviation obtained from the previous formula; This is used to filter out normal fluctuations as the deviation from the allowable threshold. Only deviations exceeding the allowable range are accumulated into the score. The meaning of this accumulated score is: when the control strategy remains unchanged, deviations that continuously exceed the allowable range will gradually accumulate over time and form a high score; when the control strategy is switched, the accumulated items are suppressed by weights, making it less likely for short-term deviations during the switching transition period to trigger abnormal segments.

[0037] The method for generating abnormal sections is consistent with the handling practices of property operation and maintenance. This step involves real-time calculation. Set trigger conditions during the process: when When the price rises and exceeds a preset trigger threshold within a series of consecutive time slices, the corresponding time slice sequence is marked as a "candidate start" of an abnormal segment; when in subsequent time slices... When the temperature falls below the threshold after the decay effect, the end time of the abnormal segment is determined. To facilitate subsequent diagnosis, this step also retains the dominant characteristics of the deviation direction within the abnormal segment, i.e., statistical analysis of the segment... The symbol percentage is used to distinguish between "consistently above the baseline" and "consistently below the baseline." In property scenarios, consistently above the baseline often aligns with phenomena such as excessive equipment operation, premature policy activation, and control failure to shut down as planned; consistently below the baseline is often associated with equipment not starting according to policy, control policy execution failure, or decreased equipment capacity. This directional information naturally connects with the subsequent interpretation of the control policy semantics, enabling the next step to generate more actionable recommendations based on typical patterns such as "policy allows but energy consumption is abnormally high" or "policy allows but energy consumption is abnormally low."

[0038] Furthermore, in terms of project deployment, this step operates in a streaming manner: when each new time slice arrives, the system synchronously retrieves the corresponding... and And read the output from the previous step. and ,calculate and This step updates the current status of the abnormal section. If the triggering condition is met, an abnormal event record is immediately generated and written to the event queue; if the abnormality persists, the end time of the event and the statistical information within the section are continuously updated. This implementation method can meet the real-time requirements of property monitoring, and at the same time generate section-level results consistent with operation and maintenance handling habits, rather than isolated single-point alarms. This step outputs a set of abnormal events, where each abnormal event includes at least the abnormal start time, abnormal end time, and deviation direction identifier.

[0039] S4. For the identified energy consumption anomaly events, analyze the execution status of the control strategy within the time period, and generate diagnostic conclusions and handling suggestions to explain the correlation between the anomaly and the control strategy.

[0040] Specifically, this step revolves around "how to transform identified anomalies into actionable operational conclusions." In the previous step, anomalies were clearly identified in the form of time segments, but this result remained at the numerical and temporal level, insufficient to directly guide property operations personnel in taking action. This step clarifies the relationship between anomalies and control strategies by restoring, abstracting, and semantically analyzing the execution status of control strategies within the anomaly segments. Based on this, it generates handling suggestions for specific strategy behaviors, enabling the anomaly diagnosis results to directly correspond to actual property management actions.

[0041] Furthermore, in the specific implementation process, this step first uses abnormal events as indexes to reconstruct the execution state of the control strategy through time slices. For each abnormal event... The system is based on its time segment ,from Extracting the corresponding time slices This forms a complete sequence of control strategy execution states. In engineering, this sequence can be directly obtained through the building control system's operation log or status point query interface, reflecting what the control system "actually allowed or restricted" during the occurrence of an anomaly.

[0042] Furthermore, after obtaining the control policy execution state sequence, this step does not directly judge a single time slice, but divides the sequence into several consecutive policy behavior segments. The rule for dividing the policy behavior segments is based on whether the control policy execution state changes: when adjacent time slices... When maintaining consistency, they are considered as the same strategic behavior segment; when When a change occurs, it is considered the start of a new segment. In this way, the control behavior within the abnormal segment is expressed as a set of policy behavior segments with clear start and end times, such as "the policy always allows operation throughout the abnormal segment" or "the policy switches from restricted operation to allowed operation shortly after the abnormality begins".

[0043] This step then performs a joint analysis of the strategy behavior segments and the abnormal deviation direction. The abnormal deviation direction originates from the deviation degree analyzed in the previous step. The statistical results are used to describe whether the energy consumption within the abnormal section is generally higher or lower than the baseline. Based on property operation experience, this step defines a set of interpretable strategy-abnormality correspondences for different combinations of scenarios. For example, when the dominant strategy behavior segment shows that the equipment is continuously in an allowable operating state within the abnormal section, and the abnormal deviation direction is consistently higher than the baseline, it indicates that under these operating conditions, the strategy configuration or execution method may lead to excessive equipment operation. When the dominant strategy shows allowable operation but the abnormal deviation direction is consistently lower than the baseline, it may indicate that the equipment is not operating effectively according to the strategy, or that there is a deviation in control execution. To avoid subjective judgment, this step quantifies this correspondence and introduces a scoring index to measure the consistency between the abnormality and the strategy.

[0044] The consistency score is calculated as follows: ; in, Indicates an abnormal event Consistency score with the dominant control strategy behavior over its duration; The time range of the abnormal event; For time slices The execution status of the control strategy; The dominant strategy is the control policy execution state that occurs most frequently within the abnormal segment, determined by... Statistical data was obtained; This represents the energy consumption deviation obtained in the previous step. Used to extract deviations from the direction; This is an indicator function used to statistically analyze deviations under the dominant policy conditions. The score reflects whether anomalous deviations primarily occur within the context of the same policy behavior, thus quantifying the strength of the association between anomalies and the control policy.

[0045] Finally, after obtaining the consistency score, this step maps the abnormal events to several predefined explanation categories based on the sign and magnitude of the score. The explanation categories are designed to directly correspond to common problems in property operation and maintenance, such as "the policy's allowed runtime does not match actual demand," "the policy allows operation but equipment response is insufficient," and "frequent policy switching leads to energy consumption fluctuations." Each explanation category corresponds to a set of explicit generation rules and textual description templates. These templates can be stored in a configuration manner during engineering implementation, and the system uses them at runtime... The system automatically selects the appropriate interpretation and fills in the abnormal segment information based on the selected value. After the interpretation is generated, this step further generates processing suggestions based on the interpretation category. The processing suggestions directly target the control policy level or the equipment verification level. For example, it may suggest verifying the start-up and shutdown plans of a certain type of equipment, checking the effectiveness of the control policy execution within the abnormal segment, or manually reviewing the policy configuration. The generation of suggestions does not involve automatic control or policy issuance; it is only output in readable text or structured record form for operation and maintenance personnel to refer to and make decisions.

[0046] This invention also provides an AI-based property energy consumption anomaly diagnosis system, the system comprising: The property operation status acquisition unit is used to acquire environmental status information, equipment control strategy execution status and actual energy consumption results of the target property building in a continuous time slice, and perform time alignment to construct a set of operation status features. The state analysis unit is used to input the operating state features of each time slice of the operating state feature set into a preset sequence neural network model, so that the sequence neural network model generates the baseline energy consumption center value and fluctuation width of the corresponding time slice according to the operating state features of each time slice. The sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on actual energy consumption results, indications of whether the execution state of the equipment control strategy has switched, baseline energy consumption center value and fluctuation width. An anomaly determination unit is used to compare the actual energy consumption result of the current time slice with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation; based on the energy consumption deviation, combined with the execution status of the equipment control strategy, an anomaly cumulative score is calculated to identify continuous energy consumption anomaly events. The suggestion generation unit is used to analyze the execution status of the control strategy within the time period of the identified energy consumption anomaly event, and generate diagnostic conclusions and handling suggestions that explain the correlation between the anomaly and the control strategy.

[0047] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0048] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0049] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An AI-based property energy consumption anomaly diagnosis method, characterized in that, The method includes: The system acquires environmental status information, equipment control strategy execution status, and actual energy consumption results of the target property building within a continuous time slice, performs time alignment, and constructs a set of operational status features. The operating state features of each time slice in the set of operating state features are input into a preset sequence neural network model, so that the sequence neural network model generates the baseline energy consumption center value and fluctuation width of the corresponding time slice item by item according to the operating state features of each time slice; the sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on actual energy consumption results, indications of whether the execution state of the equipment control strategy has switched, baseline energy consumption center value and fluctuation width; The actual energy consumption result of the current time slice is compared with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation. Based on the energy consumption deviation, an abnormal cumulative score is calculated in combination with the execution status of the equipment control strategy to identify continuous energy consumption abnormal events. For the identified energy consumption anomalies, the execution status of the control strategy within the time period is analyzed, and diagnostic conclusions and handling suggestions are generated to explain the correlation between the anomalies and the control strategy. 2.The AI-based property energy consumption anomaly diagnosis method of claim 1, wherein, The sequence neural network model consists of a feature projection layer, a GRU temporal modeling layer, and a dual-output head, including: The feature projection layer feeds the running state features of each time slice into a fully connected linear transformation layer and combines it with nonlinear activation to map the environmental state information and the execution state of the device control strategy to a unified latent space. The GRU timing modeling layer uses a gated loop unit to process data of length... The sequence is encoded to obtain the hidden state that reflects the evolution within the window; The dual output head includes one output head that generates the baseline energy consumption center value and another output head that generates the fluctuation width related to the operating state. 3.The AI-based property energy consumption anomaly diagnosis method of claim 2, characterized in that, The output header of the fluctuation width uses the Softplus function to ensure that the output is positive, which is used to construct the baseline interval.

4. The AI-based method for diagnosing abnormal energy consumption in properties according to claim 1, characterized in that, The indication quantity of whether the device control strategy execution state switches is obtained by comparing the device control strategy execution states in the running state features of adjacent time windows: when the control strategy execution state changes, taking , and when remaining unchanged, taking . 5.The AI-based property energy consumption anomaly diagnosis method of claim 2, wherein, During training, the loss function of the sequence neural network model includes a policy-switching-aware smoothing constraint term. This smoothing constraint term imposes a stronger smoothness requirement on time slices where the execution state of the device control policy remains unchanged, and relaxes the smoothing constraint near time slices where the execution state of the device control policy changes. 6.The AI-based property energy consumption anomaly diagnosis method of claim 1, wherein, The energy consumption deviation is calculated by taking the absolute value of the difference between the actual energy consumption and the baseline center value, and then combining it with the fluctuation width. 7.The AI-based property energy consumption anomaly diagnosis method of claim 1, wherein, In the calculation of the abnormal cumulative score: the value of the abnormal cumulative score is updated in each time slice. When updating, the abnormal cumulative score of the previous time slice is decayed and the energy consumption deviation of the current time slice is added. If the indicator of whether the execution state of the device control strategy has changed in the current time slice changes, the energy consumption deviation is suppressed or cleared to zero. 8.The AI-based property energy consumption anomaly diagnosis method of claim 7, wherein, The identified persistent energy consumption anomalies include: When the cumulative abnormal score exceeds a preset trigger threshold within multiple consecutive time slices, the energy consumption abnormal event is determined to begin; when the cumulative abnormal score falls back below the release threshold, the energy consumption abnormal event is determined to end. 9.The AI-based property energy consumption anomaly diagnosis method of claim 1, wherein, The generation of diagnostic conclusions includes: The execution status of the device control strategy that occurs most frequently within the time period of the energy consumption anomaly event is extracted as the dominant strategy. Based on the dominant strategy, a consistency score is calculated by combining the time segment of energy consumption anomaly events, the current execution status of the equipment control strategy, and the degree of energy consumption deviation. Based on the consistency score, energy consumption anomalies are categorized into predefined explanatory categories; wherein, the content of the explanatory category involves at least one of control strategy configuration review, equipment operation status check, or strategy execution effectiveness verification.

10. An AI-based property energy consumption anomaly diagnosis system, characterized in that, The system includes: The property operation status acquisition unit is used to acquire environmental status information, equipment control strategy execution status and actual energy consumption results of the target property building in a continuous time slice, and perform time alignment to construct a set of operation status features; The state analysis unit is used to input the operating state features of each time slice of the operating state feature set into a preset sequence neural network model, so that the sequence neural network model generates the corresponding baseline energy consumption center value and fluctuation width for each time slice according to the operating state features of each time slice. The sequence neural network model consists of a feature projection layer, a GRU timing modeling layer and dual output heads, and is trained based on actual energy consumption results, indications of whether the execution state of the equipment control strategy has switched, baseline energy consumption center value and fluctuation width. An anomaly determination unit is used to compare the actual energy consumption result of the current time slice with the baseline energy consumption center value and fluctuation width of the corresponding time slice to generate an energy consumption deviation; based on the energy consumption deviation, combined with the execution status of the equipment control strategy, an anomaly cumulative score is calculated to identify continuous energy consumption anomaly events. The suggestion generation unit is used to analyze the execution status of the control strategy within the time period of the identified energy consumption anomaly event, and generate diagnostic conclusions and handling suggestions that explain the correlation between the anomaly and the control strategy.

Images