Building energy consumption abnormity diagnosis and early warning system based on artificial intelligence algorithm

Abstract

The invention discloses a building energy consumption abnormity diagnosis and early warning system based on an artificial intelligence algorithm, and relates to the technical field of building energy consumption monitoring and abnormity diagnosis, and the system comprises a multi-source energy consumption and environment data collection module, a sensor heterogeneous network connected with an edge gateway through an LoRaWAN protocol for communication, four-stage granularity collection and storage, and equipment redundancy deployment. The data preprocessing module is used for cleaning, denoising and standardizing a sliding window; the multi-dimensional energy consumption feature extraction module is used for constructing a four-dimensional feature system, performing principal component analysis and dimensionality reduction, and reserving key features; the mixed AI algorithm anomaly diagnosis module fuses the LSTM, the gradient boosting tree and the knowledge graph; the abnormal risk grading module is used for carrying out multi-factor coupling evaluation in four stages; the early warning information generating and pushing module is used for pushing differentiated contents in multiple channels; the invention discloses a system self-optimization and iterative increment and federated learning optimization model. According to the method, the building energy consumption data quality and the abnormity diagnosis accuracy are improved, the risk grading and early warning response are optimized, and the green low-carbon operation of the building is supported.

Description

Technical Field

[0001] This invention relates to the field of building energy consumption monitoring and anomaly diagnosis technology, and in particular to a building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithms. Background Technology

[0002] As the construction industry transitions towards green and low-carbon practices, building energy consumption, as a significant component of total social energy consumption, has become a core industry requirement for refined management and efficient control. Modern buildings are increasingly complex, with significant differences in energy consumption characteristics among different building types, such as office, commercial, and residential buildings. Multiple energy-consuming devices, including central air conditioning, lighting, water supply and drainage, and elevators, operate simultaneously, resulting in multi-dimensional, highly dynamic, and strongly coupled energy consumption data. Current building energy consumption monitoring largely relies on single-type sensors, limiting data acquisition dimensions and only providing basic energy consumption values. It lacks simultaneous acquisition of key auxiliary data such as voltage, current, harmonics, and environmental parameters. Furthermore, sensor deployment coverage is incomplete, easily creating monitoring blind spots in critical equipment and peripheral areas. In the renovation of older buildings, traditional wired sensors are costly to install and have long construction cycles, making them difficult to adapt to the renovation needs of existing buildings. Simultaneously, inconsistent communication protocols among different sensors make data transmission susceptible to interference, leading to data delays, loss, or distortion, failing to provide high-quality data support for anomaly diagnosis.

[0003] Limitations in data processing and feature extraction restrict the accuracy of anomaly diagnosis. Existing systems employ simplistic preprocessing methods for collected data, often using a single threshold to filter noise. This fails to adequately consider the temporal characteristics of energy consumption data and the randomness of environmental interference, resulting in low data quality and insufficient accuracy in missing value imputation and outlier removal. At the feature extraction level, there is a focus primarily on temporal features, neglecting frequency domain, spatial domain, and cross-device correlation features, thus failing to comprehensively characterize the operational status and anomaly patterns of building energy consumption. For example, energy consumption anomalies in central air conditioning systems may be related to multiple factors such as cooling tower operating efficiency, fluctuations in supply and return water temperature differences, and changes in pump frequency. Single-dimensional features struggle to capture these complex correlations, leading to weak identification capabilities for hidden anomalies such as equipment inefficiency and pipeline leaks. Furthermore, different building energy consumption patterns are strongly correlated with seasonal variations and usage scenarios; existing feature systems lack specificity and are ill-suited to the diverse energy consumption characteristics of buildings.

[0004] The imperfections in anomaly diagnosis models and early warning mechanisms further impact energy consumption management. Traditional diagnostic methods often employ single algorithms, judging anomalies based on fixed thresholds or simple rules, failing to adapt to dynamic changes in building energy consumption, resulting in high false alarm and false negative rates. For complex anomalies such as continuous pipeline micro-leaks and intermittent valve failures, existing models lack effective identification capabilities, and the diagnostic results lack interpretability, making it difficult to assist maintenance personnel in locating the root cause. Risk classification is often based on a single energy consumption deviation indicator, failing to comprehensively consider factors such as anomaly duration, impact range, and equipment importance, resulting in insufficient scientific rigor in the classification results. Early warning information is pushed through limited channels with vague content, lacking differentiated handling suggestions and emergency procedures, leading to untimely responses and low handling efficiency for maintenance personnel. Furthermore, the system lacks continuous optimization capabilities; model parameters are fixed and cannot be iteratively upgraded through accumulated operational data, making it difficult to adapt to long-term changes in building energy consumption patterns and the emergence of new types of anomalies, thus failing to fully tap the building's energy-saving potential. Summary of the Invention

[0005] The present invention proposes a building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithms to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithms, comprising the following modules: The multi-source energy consumption and environmental data acquisition module deploys smart meters, water meters, gas meters, heating and cooling meters, and multi-dimensional environmental sensors to form a heterogeneous acquisition network; various meters and sensors collect relevant data according to corresponding standards; all devices are connected to the edge computing gateway through the LoRaWAN protocol; the gateway supports communication; data is stored in four granularities: building, building, floor, and room; and redundant acquisition nodes are deployed on key devices. The data preprocessing module adopts a processing architecture that combines time-series sliding window and intelligent cleaning. It removes gross errors using the Laida criterion, fills in missing data using an improved KNN algorithm, and processes noise using a wavelet threshold denoising algorithm; and establishes a data quality assessment system. The multi-dimensional energy consumption feature extraction module constructs a four-dimensional feature system encompassing time domain, frequency domain, spatial domain, and time series, covering various core features and extracting exclusive features for different energy-consuming systems. The anomaly diagnosis module based on hybrid AI algorithms integrates an improved LSTM neural network, gradient boosting tree, and knowledge graph to build a collaborative diagnosis framework; it takes four-dimensional features and environmental parameters as input and outputs anomaly probability, anomaly type, and anomaly confidence to identify multiple types of anomalies. The abnormal risk classification module constructs a multi-factor coupled risk assessment model, divides the risk level into four levels, dynamically adjusts the threshold in combination with the energy consumption period, and calculates the risk value by weighted summation. The early warning information generation and push module generates differentiated content based on the risk level. The early warning information includes core key information, and international projects support switching between Chinese and English. The system's self-optimization and iteration module constructs a data-driven, full lifecycle optimization mechanism. Through incremental learning and federated learning, it optimizes the model and establishes a device health assessment submodule and a digital twin verification mechanism.

[0007] Furthermore, it also includes a submodule for quantifying the degree of energy consumption anomalies, which calculates the cumulative impact of energy consumption anomalies per unit time through integration, using the following formula: Where Q is the cumulative impact value of the anomaly. This is the abnormal start time. This is the abnormal termination time. Let the actual energy consumption be a function of time. This is the predicted value for normal energy consumption. As time weight, For spatial weights.

[0008] Furthermore, it also includes a normal energy consumption dynamic prediction submodule, which integrates historical data and real-time environmental parameters to construct a prediction model, the formula of which is... ,in Let t be the predicted normal energy consumption value. This represents the historical average energy consumption for the same period. This is the temperature correction factor. This is the humidity correction factor. This is the illumination correction factor. , , These are the actual temperature, humidity, and light intensity, respectively. =25℃ =50%RH =500 lux is the standard value.

[0009] Furthermore, the multi-source energy consumption and environmental data acquisition module adopts an edge-cloud collaborative architecture. Edge nodes are deployed in the building's low-voltage electrical room and have local data preprocessing capabilities such as preliminary filtering of outliers and data compression. The sensors support self-diagnosis. For the renovation of old buildings, wireless mountable sensors are provided, which use both magnetic and adhesive fixing methods, eliminating the need for wiring during installation. Sensor health records are also established.

[0010] Furthermore, the data preprocessing module introduces a time-series anomaly pre-detection mechanism, using the isolated forest algorithm to perform preliminary screening of the raw data and identify potential abnormal segments; for buildings with strong seasonality, such as shopping malls and hotels, seasonal and trend decomposition algorithms are used to separate the seasonal, trend, and residual items in the energy consumption data, and noise of different components is processed separately; missing data filling introduces spatiotemporal correlation, when a single node's data is missing, the weighted calculation is performed by referring to the contemporaneous data of three adjacent nodes in the same area, and the closer the distance, the higher the weight.

[0011] Furthermore, the multi-dimensional energy consumption feature extraction module adds cross-device correlation features to calculate the energy consumption ratio of central air conditioning and cooling tower, the correlation between water pump flow and pressure, and the linkage rate between lighting system and human body sensing sensor; it introduces equipment operating efficiency features, such as chiller COP (cooling capacity or power consumption), boiler thermal efficiency, and transformer load rate; for new energy buildings, it extracts features such as photovoltaic power generation and load matching degree, energy storage equipment charging and discharging efficiency, and microgrid power balance.

[0012] Furthermore, the anomaly diagnosis module based on the hybrid AI algorithm introduces an interpretability enhancement mechanism, adding a SHAP value analysis layer at the model output to calculate the contribution of each feature to the diagnostic results. Positive values ​​promote anomaly detection, while negative values ​​inhibit it. Features with an absolute contribution value > 0.2 are listed as key influencing factors. The knowledge graph adopts a dynamic update mechanism, with new anomaly cases automatically added to the database after expert confirmation, and related features matched to the existing rule base using cosine similarity. For complex anomalies, such as multi-device linkage failures, a causal inference algorithm is used to trace the root cause.

[0013] Furthermore, the anomaly risk classification module introduces risk diffusion prediction, constructing a diffusion model based on equipment topology relationships and historical data, with the formula as follows: ,in for The scope of the subsequent impact, The current area of ​​influence is represented by α, where α is the diffusion coefficient. Due to the current abnormal energy consumption, This refers to the rated energy consumption of the equipment.

[0014] Furthermore, the early warning information generation and push module integrates intelligent linkage control functions. After high-risk anomalies are confirmed, control commands are automatically sent to the building automation system. It supports customized early warning strategies. Early warning information is integrated into the building BIM model, and the abnormal location is highlighted in the 3D view. The surrounding related equipment and evacuation routes are marked. Clicking on the equipment allows you to view historical energy consumption curves and maintenance records.

[0015] Furthermore, the system's self-optimization and iteration module constructs a dual-loop optimization mechanism. The inner loop fine-tunes the model's hyperparameters hourly using real-time data, while the outer loop retrains the model monthly based on full data. A reinforcement learning optimization and diagnostic strategy is adopted to intelligently adjust feature weights and decision thresholds. Equipment health assessment incorporates remaining life prediction, using LSTM neural networks to analyze energy consumption data, operating temperature, and vibration frequency to predict the equipment's remaining lifespan.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention, through end-to-end technological innovation, constructs an intelligent and refined building energy consumption anomaly diagnosis and early warning system, effectively addressing many limitations of existing technologies. At the data acquisition level, it employs a multi-dimensional heterogeneous sensor network covering various data types, including energy consumption, electrical, and environmental data. It supports four-level granularity storage and redundant acquisition, and combines wireless communication and edge computing architecture to adapt to both new and renovated building scenarios, achieving comprehensive data coverage across all areas and devices. Sensor self-diagnosis and dual-backup communication design ensure the stability and integrity of data transmission, providing a high-quality data foundation for anomaly diagnosis and fundamentally changing the traditional acquisition methods' limitations of single-dimensionality and low data quality.

[0017] Significant improvements in data processing and feature extraction capabilities provide strong support for accurate diagnosis. Through time-series sliding window preprocessing, multi-algorithm collaborative cleaning, and seasonal-trend decomposition techniques, data quality is effectively improved, ensuring the accuracy of missing value imputation and outlier removal. A four-dimensional feature system encompassing time, frequency, spatial, and time-series domains is constructed, integrating cross-device correlation features and equipment operating efficiency features to comprehensively characterize the operational status and anomaly patterns of building energy consumption, accurately capturing hidden and complexly correlated anomalies. Automated feature screening and targeted feature design adapt to the characteristics of different building types and energy-consuming equipment, greatly improving the comprehensiveness and targeting of anomaly identification and solving the problems of traditional feature systems being singular and lacking adaptability.

[0018] The scientific rigor and effectiveness of anomaly diagnosis and early warning have been significantly enhanced. A hybrid AI algorithm and knowledge graph collaborative diagnostic framework, combined with interpretability analysis and causal inference techniques, improves the accuracy and interpretability of anomaly diagnosis, enabling precise identification of various explicit and implicit anomalies and tracing their root causes. A multi-factor coupled risk grading model and risk diffusion prediction make the grading results more aligned with actual operational scenarios. Differentiated early warning content and a multi-channel push mechanism ensure that early warning information reaches relevant personnel in a timely manner. Supporting handling suggestions and emergency procedures improve operational response efficiency and handling quality. The system's self-optimization mechanism, through incremental learning and federated learning, achieves continuous model iteration and upgrades. Combined with equipment health assessment and remaining life prediction, it provides early maintenance reminders, reducing equipment failure rates. Digital twin verification and energy-saving potential analysis provide a scientific basis for building energy-saving renovations, helping to explore energy-saving potential and promote green and low-carbon management of building energy consumption, demonstrating broad application value in various building energy consumption management scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithms proposed in this invention; Figure 2 Bar chart showing the diagnostic accuracy for different abnormality types; Figure 3 A line graph showing system uptime versus false alarm rate; Figure 4 A bar chart showing the early warning response time for different risk levels. Detailed Implementation

[0020] 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.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithms includes the following modules: The multi-source energy consumption and environmental data acquisition module deploys a heterogeneous acquisition network consisting of smart meters, water meters, gas meters, heating and cooling meters, and multi-dimensional environmental sensors. The smart meters use a three-phase four-wire system to acquire voltage, current, active power, and reactive power data at a sampling frequency of 1Hz, with a metering accuracy of 0.5 class and support for harmonic analysis up to the 31st order. The water and gas meters employ ultrasonic Doppler metering technology with a sampling frequency of 0.1Hz, an error ≤±2%, and anti-magnetic interference capabilities. The heating and cooling meters simultaneously acquire supply and return water temperature, flow rate, and cumulative heat, with an update cycle of 10 seconds. Temperature measurement... Accuracy ±0.1℃; the environmental sensor integrates temperature and humidity, light intensity, CO2 concentration, PM2.5, and noise decibel detection, with a sampling frequency of 0.5Hz, a temperature measurement range of -40℃ to 80℃, and a humidity measurement range of 0-100%RH; all sensors are connected to the edge computing gateway via the LoRaWAN protocol, the gateway supports 5G / Ethernet dual backup communication, data transmission latency ≤1s, and data is stored at four levels of granularity: building, unit, floor, and room. Redundant acquisition nodes are deployed for critical equipment such as central air conditioning units and transformers, achieving 100% data acquisition coverage. The data preprocessing module employs a processing architecture combining a time-series sliding window and intelligent cleaning. The window size is set to 10 minutes. It identifies and removes gross errors in energy consumption data using the Raida criterion (3σ principle), with a removal rate ≤0.5%. For missing data, an improved KNN algorithm combined with historical trends is used for filling, with a filling error ≤3%. A sensor fault warning is triggered when there are more than 5 minutes of consecutive missing data. A wavelet threshold denoising algorithm is used to process high-frequency noise, with a soft threshold function threshold λ=0.6σ, where σ is the data standard deviation, resulting in a signal-to-noise ratio improvement of ≥15dB. Z-score standardization maps the data to the [-1,1] interval to eliminate the influence of dimensions. A data quality assessment system is established, including completeness ≥98%, consistency timestamp synchronization error ≤1s, and accuracy measurement deviation ≤2%. When indicators fail to meet the standards, data re-collection and sensor calibration processes are initiated. The entire preprocessing process is traceable and auditable. A multi-dimensional energy consumption feature extraction module constructs a four-dimensional feature system encompassing time domain, frequency domain, spatial domain, and time series. Time domain features include average daily energy consumption, peak energy consumption, valley energy consumption, load factor, fluctuation coefficient, and peak-valley difference. Frequency domain features obtain fundamental frequency, harmonic amplitude, total harmonic distortion (THD), and spectral entropy through Fast Fourier Transform. Spatial domain features include energy consumption correlation in different regions, equipment energy consumption ratio, and spatial load density. Time series features cover trend slope, periodic intensity, seasonality index, number of abrupt change points, and autocorrelation coefficient with a lag of 1-24 hours. Features are extracted separately for systems such as central air conditioning, lighting, water supply and drainage, and elevators, such as the fluctuation rate of supply and return water temperature difference in central air conditioning, the pump operating frequency curve, and the start-stop frequency and illumination matching degree of lighting systems. The feature dimensions are ≥50, which are reduced to 20-30 dimensions through principal component analysis, retaining key features with a cumulative contribution rate ≥90%. The anomaly diagnosis module based on a hybrid AI algorithm integrates an improved LSTM neural network, a gradient boosting tree, and a knowledge graph to construct a collaborative diagnostic framework. The LSTM network contains four hidden layers (64-128-64-32 neurons), employing Dropout (rate=0.2) and L2 regularization (λ=0.001) to suppress overfitting. The learning rate is dynamically adjusted from an initial 0.001, decaying by 10% every 50 rounds. The gradient boosting tree has 150 decision trees, a maximum depth of 8, a learning rate of 0.05, and a minimum number of sample splits of 20. The knowledge graph... The system stores association rules for spectral storage devices, historical anomaly cases, and expert experience, such as "a chiller unit outlet water temperature > 7℃ and energy consumption increase of 15% indicates a decrease in heat exchange efficiency." It inputs four-dimensional features and environmental parameters, outputting anomaly probabilities of 0-1, anomaly types including equipment failure, pipeline leakage, parameter misalignment, inefficient operation, and environmental interference (10 categories in total), with anomaly confidence ≥ 80%. The diagnostic accuracy is ≥ 96%, with a false negative rate ≤ 2% and a false positive rate ≤ 3%. It supports continuous anomaly identification such as pipeline micro-leakage, sudden anomaly identification such as equipment short circuits, and intermittent anomaly identification such as periodic valve failures, with a diagnostic delay ≤ 2 seconds. The abnormal risk classification module constructs a multi-factor coupled risk assessment model. Assessment factors include: energy consumption deviation magnitude (ratio of actual energy consumption to normal energy consumption), abnormal duration, affected area (area / number of devices), equipment importance (key equipment weight 1.0, ordinary equipment 0.6), and diffusion rate (increase rate per unit time). Risk levels are divided into four levels: no risk (deviation magnitude ≤ ±5%), low risk (5%-10% and duration < 10 min), medium risk (10%-20% or duration 10-30 min), and high risk (> 20% or duration > 30 min or involving safety hazards). Thresholds are dynamically adjusted based on energy consumption periods: during peak periods (e.g., office buildings 9:00-18:00), the threshold is lowered by 20%; during off-peak periods (e.g., 0:00-6:00), the threshold is raised by 10%. Risk values ​​are calculated using a weighted summation, and high risk automatically triggers an emergency response. The early warning information generation and push module generates differentiated content based on risk level. No risk is simply recorded by the system; low-risk messages include the abnormal location and preliminary assessment; medium-risk messages include additional investigation steps and historical similar cases; and high-risk messages include emergency response procedures, contact information for responsible personnel, and recommendations to cut off energy supply. Early warning information includes time accurate to the second, location (building-floor-room-equipment), parameters (energy consumption before and after the anomaly), environmental values, type, risk level, and processing priority. It pushes information through multiple channels including web platform, mobile APP, SMS, and voice calls. For key personnel, if an APP push is not read within 10 seconds, it is automatically converted to SMS; if there is no response within 30 seconds, a voice call is triggered. It supports multi-level confirmation feedback—discovery-handling-closure—with an anomaly status updated every 2 minutes until resolved if the loop is not closed. International projects support bilingual (Chinese and English) switching with a response delay of ≤4 seconds. The system's self-optimization and iteration module constructs a data-driven, full lifecycle optimization mechanism, collecting real-time indicators such as diagnostic accuracy, false alarm rate, missed alarm rate, early warning response time, and handling completion rate, generating a performance report every 24 hours. It employs an incremental learning algorithm, fine-tuning the model parameters with a learning rate of 0.0001 for every 1000 labeled cases accumulated, while retaining historical knowledge. A federated learning framework is introduced, jointly training multiple building nodes; local nodes only upload encrypted model gradients, while the central node aggregates and updates, protecting data privacy. An equipment health assessment submodule is established, calculating a health index of 0-100 based on energy consumption trends, operational parameter drift, and fault frequency; a maintenance prompt is pushed when the index is <60. Digital twin verification is integrated, comparing actual energy consumption with virtual building model simulation results; the diagnostic model is corrected when the deviation exceeds 5%. A monthly energy consumption analysis report is generated, including anomaly statistics, energy-saving potential ≥5%, and optimization suggestions such as equipment replacement and operational strategy adjustments.

[0024] This invention also includes a submodule for quantifying the degree of energy consumption anomalies, which calculates the cumulative impact of energy consumption anomalies per unit time through integration, using the following formula: Where Q is the cumulative impact value of the anomaly. This is the abnormal start time. This is the abnormal termination time. Let the actual energy consumption be a function of time. This is the predicted value for normal energy consumption. The time weights are: peak 1.5, off-peak 1.0, and trough 0.8. With a spatial weight of 1.2 for key areas and 1.0 for ordinary areas, this quantification method comprehensively assesses the impact of anomalies by considering both spatiotemporal factors, providing quantitative support for resource allocation and priority management.

[0025] This invention also includes a normal energy consumption dynamic prediction submodule, which integrates historical data and real-time environmental parameters to construct a prediction model and introduces multi-factor correction coefficients, the formula of which is: ,in Let t be the predicted normal energy consumption value. This represents the historical average energy consumption for the same period. The temperature correction factor ranges from 0.02 to 0.05. The humidity correction factor ranges from 0.01 to 0.03. The illumination correction factor ranges from 0.005 to 0.01. , , These are the actual temperature, humidity, and light intensity, respectively. =25℃ =50%RH =500 lux is the standard value. The normal energy consumption baseline is refined by correcting multiple environmental factors to improve the differentiation of abnormal diagnosis.

[0026] In this invention, the multi-source energy consumption and environmental data acquisition module adopts an "edge + cloud" collaborative architecture. The edge nodes are deployed in the building's low-voltage electrical room and have local data preprocessing capabilities such as preliminary filtering of outliers and data compression with a compression rate of ≥50%, uploading only valid data. The sensors support self-diagnosis and periodically output operating status codes: 0 normal, 1 low battery, 2 communication failure, and 3 metering anomaly. The battery life is ≥2 years, and it can continue to work when the solar-assisted power supply is provided with a light intensity of ≥500 lux. For the renovation of old buildings, wireless mountable sensors are provided using a magnetic and adhesive dual fixing method, requiring no wiring for installation. The acquisition interval can be remotely adjusted from 1 second to 10 minutes. A sensor health record is established, recording the installation time, calibration records, and fault history. When the health level is <80%, it actively reminds the user to replace the sensor.

[0027] In this invention, the data preprocessing module introduces a temporal anomaly pre-detection mechanism. The isolated forest algorithm is used to initially screen the raw data, identifying potential anomaly segments with an anomaly percentage ≤5%. For these segments, a stricter cleaning strategy is employed, such as lowering the wavelet denoising threshold to 0.4σ. For buildings with strong seasonality, such as shopping malls and hotels, the Seasonal-Trend Decomposition (STL) algorithm is used to separate the seasonal, trend, and residual components in the energy consumption data, processing noise of different components separately. For missing data imputation, spatiotemporal correlation is introduced. When a single node's data is missing, the weighted calculation is performed by referencing the contemporaneous data of three adjacent nodes in the same area; the closer the distance, the higher the weight, improving the imputation accuracy to over 97%. The preprocessing results generate a data quality score range of 0-100 points. Data above 90 points is considered high-quality and can be directly used for modeling; data between 70-90 points requires manual sampling confirmation; and data below 70 points triggers sensor maintenance.

[0028] In this invention, the multi-dimensional energy consumption feature extraction module adds cross-device correlation features, calculates the energy consumption ratio of central air conditioning and cooling tower, the correlation between water pump flow and pressure, and the linkage rate between lighting system and human body sensing sensor. When the correlation is lower than the threshold of 0.6, it is marked as a potential abnormal feature. Equipment operating efficiency features are introduced, such as the COP of chiller unit (cooling capacity / power consumption), boiler thermal efficiency, and transformer load rate. When the efficiency is lower than the industry standard value by 10%, the weight of this feature is strengthened. For new energy buildings, features such as photovoltaic power generation and load matching degree, energy storage equipment charging and discharging efficiency, and microgrid power balance are extracted. Feature engineering adopts automated tools (AutoML) and combines genetic algorithms to select the optimal feature combination, which reduces the number of input features of the model by 30% and improves the diagnostic accuracy by 2%-3%.

[0029] In this invention, the anomaly diagnosis module based on a hybrid AI algorithm introduces an interpretability enhancement mechanism. A SHAP value analysis layer is added to the model output to calculate the contribution of each feature to the diagnostic results. Positive values ​​promote anomaly detection, while negative values ​​inhibit it. Features with an absolute contribution value > 0.2 are listed as key influencing factors and displayed along with the diagnostic results. The knowledge graph adopts a dynamic update mechanism. New anomaly cases are automatically added to the database after being confirmed by more than 3 experts. Related features are matched to the existing rule base using cosine similarity. When the matching degree is ≥ 0.8, similar rules are merged. For complex anomalies, such as multi-device linkage failures, a causal inference algorithm is used to trace the root cause, with a location accuracy of ≥ 90%. The output is a three-level chain of "abnormal phenomenon - direct cause - root cause", such as "surge in air conditioning energy consumption - decrease in chiller output - condenser scaling".

[0030] In this invention, the abnormal risk classification module introduces risk diffusion prediction. Based on equipment topology and historical data, a diffusion model is constructed to predict the risk impact range in the next 10 minutes, 30 minutes, and 60 minutes. The formula is as follows: ,in for The scope of the subsequent impact, The current range of influence is represented by α, which is the diffusion coefficient, taken as 0.05 when the equipment coupling degree is high and 0.01 when it is low. Due to the current abnormal energy consumption, By predicting the risk spread trend, resources can be deployed in advance to reduce the scope of accidents and ensure the rated energy consumption of equipment.

[0031] In this invention, the early warning information generation and push module integrates intelligent linkage control functions. After high-risk anomalies such as electrical fire hazards and gas leaks are confirmed, control commands such as shutting off valves, starting ventilation, and shutting down equipment are automatically sent to the building automation system. The command transmission delay is ≤1s, and the execution feedback time is ≤3s. It supports customized early warning strategies. Managers can set "do-not-disturb periods" (e.g., no push notifications during meetings unless it is a high-risk event), "responsible person grouping" (divided by area / equipment type), and "handling process templates" (e.g., checking the filter first and then the compressor for air conditioning anomalies). The early warning information is integrated into the building BIM model, highlighting the anomaly location in the 3D view, marking surrounding related equipment and evacuation routes, and clicking on the equipment allows viewing historical energy consumption curves and maintenance records, improving handling efficiency by more than 30%.

[0032] In this invention, the system's self-optimization and iteration modules construct a dual-loop optimization mechanism. The inner loop fine-tunes model hyperparameters such as learning rate and window size hourly using real-time data, while the outer loop retrains the model monthly based on full data. A reinforcement learning-based optimization and diagnostic strategy is employed, using "minimum false alarm rate + fastest response time" as the reward function to intelligently adjust feature weights and decision thresholds. After 100 days of iteration, the false alarm rate is reduced by 40%. Equipment health assessment incorporates remaining life prediction, using LSTM neural networks to analyze energy consumption data, operating temperature, and vibration frequency to predict the equipment's remaining lifespan (RUL) with an error ≤8%, and pushes replacement plans 30 days in advance. Energy-saving recommendations are generated using cost-benefit analysis, calculating the payback period for each measure, such as the payback period for replacing LED lights = lamp cost / annual electricity savings. Schemes with a payback period <3 years are prioritized, providing quantitative decision support for building energy-saving renovations.

[0033] The following two examples further illustrate the specific implementation of this system: Example 1: Implementation of Energy Consumption Anomaly Diagnosis and Early Warning in Large Commercial Buildings This embodiment targets a large commercial building with a floor area of ​​80,000 square meters, encompassing shopping areas, dining areas, cinemas, underground parking garages, and other functional areas. The main energy-consuming equipment includes 4 central air conditioning units, 12 water pumps, over 8,000 lighting fixtures, 30 elevators, and various gas-fired equipment for catering. The building experiences an average daily customer flow of over 20,000 people, with peak energy consumption concentrated on weekends and holidays from 10:00 AM to 9:00 PM. It needs to adapt to a highly dynamic, multi-device, and strongly coupled energy consumption scenario. The specific implementation is as follows: 1. System module deployment and parameter configuration Multi-source energy consumption and environmental data acquisition module deployment: Acquisition devices are deployed at four levels: building, floor, and room. Eight multi-dimensional environmental sensors are deployed on each floor of the shopping area to simultaneously collect data on temperature, humidity, light intensity, CO2 concentration, PM2.5, and noise levels. In the catering area, each shop is equipped with smart meters for electricity, gas, and water. The smart meters use a three-phase four-wire system to collect voltage, current, active power, and reactive power at a sampling frequency of 1Hz, with a metering accuracy of 0.5 class and support for harmonic analysis up to the 31st order. The gas and water meters use ultrasonic Doppler metering technology with a sampling frequency of 0.1Hz, an error of ≤±2%, and anti-magnetic interference capabilities. The central air conditioning room is equipped with heating and cooling meters to simultaneously collect supply and return water temperature, flow rate, and cumulative heat, with an update cycle of 10 seconds and a temperature measurement accuracy of ±0.1℃. Wireless surface-mount sensors are deployed in the underground parking garage and edge areas, using a magnetic and adhesive dual-fixation method, requiring no wiring, and the data collection interval is set to 5 seconds. All sensors are connected to 12 edge computing gateways via the LoRaWAN protocol. The gateways support dual backup communication of 5G / Ethernet with a data transmission latency of ≤1s. Redundant acquisition nodes are deployed for key equipment such as central air conditioning units and transformers, achieving 100% data acquisition coverage. The sensors support self-diagnosis, periodically output operating status codes, and have a battery life of ≥2 years. The sensors in the high-humidity environment of the catering area have an IP54 protection rating.

[0034] Data preprocessing module configuration: A processing architecture combining a time-series sliding window and intelligent cleaning is adopted, with a window size set to 10 minutes. The Laida criterion (3σ principle) is used to identify and remove gross errors in energy consumption data, with the removal rate controlled within 0.3%. For data gaps caused by equipment start-ups and shutdowns in the catering area, an improved KNN algorithm combined with historical trends is used for data filling, with a filling error ≤3%. A sensor fault warning is triggered when consecutive gaps exceed 5 minutes. A wavelet thresholding algorithm is used to process high-frequency noise, with a soft threshold function threshold λ = 0.6σ, where σ is the data standard deviation. After denoising the central air conditioning current data, the signal-to-noise ratio is improved by 18 dB. Z-score standardization maps all data to the [-1,1] interval to eliminate the influence of dimensions. Establish a data quality assessment system, setting the following criteria: completeness ≥ 98%, consistency timestamp synchronization error ≤ 1s, and accuracy measurement deviation ≤ 2%. When the indicators fail to meet the standards, data supplementation and sensor calibration processes are initiated, and the entire preprocessing process is traceable and auditable. In view of the strong seasonality of commercial buildings, the Seasonal-Trend Decomposition (STL) algorithm is used to separate the seasonal, trend, and residual items in the energy consumption data, and noise of different components is processed separately.

[0035] Multi-dimensional energy consumption feature extraction module configuration: Construct a four-dimensional feature system encompassing time domain, frequency domain, spatial domain, and time series. Time domain features include average daily energy consumption, peak energy consumption, valley energy consumption, load factor, fluctuation coefficient, and peak-valley difference. For gas equipment in the catering area, additionally, average daily usage time and peak-hour energy consumption ratio are extracted. Frequency domain features obtain fundamental frequency, harmonic amplitude, total harmonic distortion (THD), and spectral entropy through Fast Fourier Transform. For central air conditioning motor equipment, the amplitudes of the 3rd and 5th harmonics are analyzed. Spatial domain features include the correlation of energy consumption between different floors, the energy consumption ratio of each area, and spatial load density, calculating the energy consumption coupling coefficient between the shopping area and the catering area. Time series features cover trend slope, periodic intensity, seasonality index, number of abrupt change points, and autocorrelation coefficient with a lag of 1-24 hours. For the central air conditioning system, the fluctuation rate of supply and return water temperature difference, the water pump operating frequency curve, and the COP value of the chiller unit are extracted separately. For the lighting system, the start-stop frequency and the matching degree of illumination are extracted. The feature dimension reached 55 dimensions, which was reduced to 25 dimensions through principal component analysis. Key features with a cumulative contribution rate of ≥90% were retained. Cross-equipment correlation features were added to calculate the energy consumption ratio between the central air conditioning and cooling tower, and the correlation between water pump flow and pressure. When the correlation degree is less than 0.6, it is marked as a potential abnormal feature.

[0036] The anomaly diagnosis module based on a hybrid AI algorithm is configured as follows: A collaborative diagnostic framework is constructed by integrating an improved LSTM neural network, gradient boosting tree, and knowledge graph. The LSTM network contains four hidden layers (64-128-64-32 neurons), employing Dropout with a rate of 0.2 and L2 regularization with λ of 0.001 to suppress overfitting. The initial learning rate is set to 0.001, decaying by 10% every 50 rounds. The gradient boosting tree has 150 decision trees, a maximum depth of 8, a learning rate of 0.05, and a minimum number of sample splits of 20. The knowledge graph stores association rules, historical anomaly cases, and expert experience, such as "a chiller unit outlet water temperature > 7℃ and energy consumption increase of 15% indicates a decrease in heat exchange efficiency" and "instantaneous flow rate of gas meter > 5m³ / h". 3 "Significant leakage occurs when there is no cooking period." Input four-dimensional features and environmental parameters, output anomaly probability (0-1), anomaly type (equipment failure, pipeline leakage, parameter misalignment, etc., 10 categories), and anomaly confidence ≥80%. An interpretability enhancement mechanism is introduced, adding a SHAP value analysis layer to calculate the contribution of each feature to the diagnostic results. Features with an absolute contribution value >0.2 are listed as key influencing factors. For sudden anomalies such as gas leaks in catering areas, the diagnostic delay is ≤2s, the diagnostic accuracy is ≥96%, the false negative rate is ≤2%, and the false positive rate is ≤3%.

[0037] Configuration of the abnormal risk classification and early warning push module: A multi-factor coupled risk assessment model is constructed. Assessment factors include energy consumption deviation amplitude (ratio of actual energy consumption to normal energy consumption), abnormal duration, affected area (area / number of devices), equipment importance (key equipment weight 1.0, ordinary equipment 0.6), and diffusion rate (increase rate per unit time). Risk levels are divided into four levels: no risk (deviation amplitude ≤ ±5%), low risk (5%-10% and duration < 10 min), medium risk (10%-20% or duration 10-30 min), and high risk (> 20% or duration > 30 min or involving safety hazards). Thresholds are dynamically adjusted based on the energy consumption periods of commercial buildings: a 20% reduction during peak hours (10:00-21:00) and a 10% increase during off-peak hours (0:00-6:00). The cumulative impact value of the abnormality is calculated through integration, using the following formula: If the abnormal start time of gas leak in the catering area =18:00, End Time =18:10, actual energy consumption =8m 3 / h, normal energy consumption prediction =1.5m 3 / h, time weight =1.5 Peak period, spatial weight =1.2 Key region, integral calculation yields Q=∫(18:00-18:10)|8-1.5|×1.5×1.2dt=6.5×1.8×600=7020 (unit: m) 3 •s), judged as high risk. Normal energy consumption dynamic prediction uses the formula In the summer 2 PM central air conditioning energy consumption forecast, the historical average energy consumption for the same period is included. =120kW, actual temperature =35℃, standard temperature =25℃, temperature correction factor =0.04, actual humidity =65%RH, standard humidity =50%RH, humidity correction factor =0.02, actual light intensity =800 lux, standard illuminance =500 lux, illumination correction factor =0.008, calculated as follows Risk diffusion prediction uses the formula If the water pump malfunctions, the current affected area =200㎡, diffusion coefficient α=0.05, high equipment coupling, Δt=300s, current abnormal energy consumption. =80kW, rated energy consumption =50kW, calculated as S(300s)=200×(1+0.05×300×80 / 50)=200×(1+0.05×480)=200×25=5000㎡. Warning information is pushed through multiple channels including Web platform, mobile APP, SMS, and voice call. For high-risk gas leaks in the catering area, if the APP push is not read within 10 seconds, it is automatically converted to an SMS; if there is no response within 30 seconds, a voice call is triggered. The warning information is integrated into the building BIM model, highlighting the leak location and surrounding valve locations in the 3D view, and simultaneously pushing emergency response procedures and contact information for responsible personnel.

[0038] System self-optimization and iteration module configuration: A data-driven, full lifecycle optimization mechanism is constructed, collecting real-time indicators such as diagnostic accuracy, false alarm rate, missed alarm rate, early warning response time, and handling completion rate, generating a performance report every 24 hours; an incremental learning algorithm is adopted, fine-tuning model parameters every 1000 labeled cases, with a learning rate set to 0.0001, retaining historical knowledge; a federated learning framework is introduced, training on three similar commercial building nodes, with local nodes only uploading encrypted model gradients, and the central node aggregating and updating; an equipment health assessment submodule is established, calculating a health index of 0-100 based on energy consumption trends, operating parameter drift, and fault frequency, pushing maintenance prompts when the central air conditioning unit's health index is <60; integrated digital twin verification compares actual energy consumption with virtual building model simulation results, correcting the diagnostic model when the deviation exceeds 5%; a monthly energy consumption analysis report is generated, including anomaly statistics, energy-saving potential analysis, and optimization suggestions, recommending replacement with LED lights for the inefficient lighting system, with a calculated investment payback period of 2.5 years.

[0039] 2. Effect Verification and Table Analysis

[0040] Table 1 Key Operational Indicators of Large Commercial Building Systems in Example 1 Exception types Diagnostic accuracy (%) Warning response time (s) Processing completion efficiency (min) False alarm rate (%) Missed report rate (%) Gas leak (sudden) 99.2 3.5 8 1.1 0.8 Decreased heat exchange efficiency of air conditioner (implicit) 96.5 4.2 25 2.3 1.7 Lighting circuit fault (intermittent) 97.8 3.8 12 1.8 1.2 Abnormal water pump flow rate (continuous type) 98.3 4.0 18 2.0 1.5 Table 1 shows data based on 30 days of continuous system monitoring, comprehensively reflecting the operational effectiveness of this invention in complex scenarios within large commercial buildings. The accuracy rate for diagnosing various anomalies exceeds 96%, with a 99.2% accuracy rate for sudden anomalies such as gas leaks. This is attributed to the comprehensive coverage of multi-source heterogeneous sensors and the precise characterization of the four-dimensional feature system, enabling rapid capture of core anomaly characteristics. The early warning response time is controlled within 4.2 seconds. Multi-channel push mechanisms and BIM model visualization ensure that maintenance personnel receive critical information promptly. The handling efficiency for gas leak anomalies is only 8 minutes, effectively reducing the risk of accident escalation. The false alarm rate and false negative rate are controlled below 2.3% and 1.7% respectively, demonstrating the advantages of hybrid AI algorithms and knowledge graph collaborative diagnosis. Combined with interpretability analysis and causal inference technology, the reliability of diagnostic results is significantly improved. Through continuous self-optimization, the system adapts to the highly dynamic energy consumption patterns of commercial buildings, providing strong support for refined energy consumption management and energy-saving renovations.

[0041] Example 2: Implementation of Energy Consumption Anomaly Diagnosis and Early Warning in Smart Office Buildings This embodiment targets a 30,000㎡ smart office building with 20 floors, including open-plan office areas, private offices, meeting rooms, server rooms, and underground parking. It integrates a photovoltaic power supply system and energy storage equipment. The main energy-consuming equipment includes two central air conditioning units, eight water pumps, over 5,000 smart lighting fixtures, and ten elevators. Office hours are Monday to Friday, 9:00 AM to 6:00 PM. Energy consumption patterns are regular and strongly correlated with sunlight and personnel density, requiring adaptation to renewable energy access and regular energy consumption scenarios. The specific implementation is as follows: 1. System module deployment and parameter configuration Deployment of Multi-Source Energy Consumption and Environmental Data Acquisition Modules: Acquisition devices are deployed at four levels of granularity. In open office areas, one multi-dimensional environmental sensor is deployed every 50 square meters to collect data on temperature, humidity, light intensity, CO2 concentration, PM2.5, and noise levels at a sampling frequency of 0.5 Hz. In private offices and meeting rooms, smart meters and human body sensors are deployed. The smart meters have a sampling frequency of 1 Hz, a metering accuracy of 0.5, and support harmonic analysis up to the 31st order. In the central air conditioning room and water pump room, heat meters and vibration sensors are deployed. The heat meters have a 10-second update cycle and a temperature measurement accuracy of ±0.1℃. Wireless surface-mounted sensors are deployed in the underground parking lot, with a data collection interval of 10 seconds. Dedicated metering sensors are deployed in the photovoltaic panel array and energy storage equipment to collect data on power generation, charging and discharging power, and remaining power. All sensors are connected to six edge computing gateways via the LoRaWAN protocol. The gateways support dual backup communication via 5G / Ethernet with a data transmission latency of ≤1s. Redundant acquisition nodes are deployed for key equipment such as transformers and photovoltaic inverters, achieving 100% data acquisition coverage. The sensors support self-diagnosis and periodically output operating status codes. The sensors in the rooftop photovoltaic area support solar-assisted power supply and can operate continuously when the light intensity is ≥500 lux.

[0042] Data preprocessing module configuration: A processing architecture combining a time-series sliding window and intelligent cleaning is adopted, with a window size set to 10 minutes. The Laida criterion (3σ principle) is used to identify and remove gross errors in energy consumption data, with the removal rate controlled within 0.2%. For data gaps caused by sparse personnel on weekends, an improved KNN algorithm combined with historical trends is used for filling, introducing spatiotemporal correlation. Data from three adjacent offices on the same floor is weighted for calculation, with closer offices receiving higher weights, improving the filling accuracy to over 97%. A sensor fault warning is triggered when more than 5 minutes of data are missing. A wavelet threshold denoising algorithm is used to process high-frequency noise, with a soft threshold function threshold λ=0.6σ, where σ is the data standard deviation. After denoising, the signal-to-noise ratio of photovoltaic power generation data is improved by 16dB. Z-score standardization maps all data to the [-1,1] interval to eliminate the influence of dimensions. A data quality assessment system was established, setting the following criteria: integrity ≥ 98%, consistency timestamp synchronization error ≤ 1s, and accuracy measurement deviation ≤ 2%. The preprocessed results generated a data quality score range of 0-100 points. Data with scores above 90 points were considered high-quality and directly used for modeling, while data with scores between 70 and 90 points required manual sampling for confirmation, and data with scores < 70 points triggered sensor maintenance. A time-series anomaly pre-detection mechanism was introduced, using the isolated forest algorithm to perform preliminary screening of the raw data and identify potential abnormal segments. The proportion of abnormal segments was ≤ 5%, and a more stringent cleaning strategy was adopted for such segments, with the wavelet denoising threshold lowered to 0.4σ.

[0043] Multi-dimensional energy consumption feature extraction module configuration: Construct a four-dimensional feature system encompassing time domain, frequency domain, spatial domain, and time series. Time domain features include average daily energy consumption, peak energy consumption, valley energy consumption, load factor, fluctuation coefficient, and peak-valley difference. For photovoltaic systems, average daily power generation and peak-valley power generation ratio are additionally extracted. Frequency domain features obtain fundamental frequency, harmonic amplitude, total harmonic distortion (THD), and spectral entropy through Fast Fourier Transform. For elevator equipment, current harmonic distribution is analyzed in detail. Spatial domain features include the correlation of energy consumption on different floors, the energy consumption ratio of each area, and spatial load density. The energy consumption coupling coefficient between open office areas and meeting rooms is calculated. Time series features cover trend slope, periodic intensity, seasonality index, number of abrupt change points, and autocorrelation coefficient with a lag of 1-24 hours. For central air conditioning systems, the fluctuation rate of supply and return water temperature difference and the water pump operating frequency curve are extracted; for lighting systems, the start-stop frequency and light matching degree are extracted; for photovoltaic systems, the power generation and load matching degree and the charging and discharging efficiency of energy storage devices are extracted. The feature dimensions reached 52 dimensions, which were reduced to 22 dimensions through principal component analysis, retaining key features with a cumulative contribution rate of ≥90%; cross-device correlation features were added to calculate the power balance between central air conditioning and energy storage devices, and the linkage rate between lighting systems and human body induction sensors. Features with a correlation degree lower than 0.6 were marked as potential abnormal features.

[0044] The anomaly diagnosis module based on a hybrid AI algorithm is configured as follows: A collaborative diagnosis framework is constructed by integrating an improved LSTM neural network, a gradient boosting tree, and a knowledge graph. The LSTM network contains four hidden layers, i.e., 64-128-64-32 neurons. Dropout with a rate of 0.2 and L2 regularization with λ of 0.001 are used to suppress overfitting. The learning rate is initially set to 0.001 and decays by 10% every 50 rounds. The gradient boosting tree has 150 decision trees, a maximum depth of 8, a learning rate of 0.05, and a minimum number of sample splits of 20. The knowledge graph stores device association rules, historical anomaly cases, and expert experience, such as "a photovoltaic inverter with output power less than 80% of rated power and sufficient sunlight is an inverter fault." Inputting four-dimensional features and environmental parameters, the system outputs anomaly probabilities (0-1), anomaly types (10 categories including equipment failure, parameter misalignment, and operational inefficiency), and anomaly confidence levels ≥80%. An interpretability enhancement mechanism is introduced, adding a SHAP value analysis layer to calculate the contribution of each feature to the diagnostic results. For anomalies linked to photovoltaic and energy storage systems, a causal inference algorithm is used to trace the root cause, achieving a location accuracy ≥90%. The cumulative impact value of the anomaly is calculated through integration, using the following formula: If the photovoltaic inverter malfunction starts at an abnormal time =11:00, End Time =11:20, actual energy consumption =30kW, normal energy consumption prediction value =120kW, time weighting =1.5 Peak period, spatial weight 1.0 In the ordinary area, the integral calculation yields Q = ∫(11:00-11:20)|30-120|×1.5×1.0dt = 90×1.5×1200 = 162000 (unit: kW・s), which is classified as medium risk. Normal energy consumption dynamic prediction uses the formula... In the autumn 2023 2:00 PM office lighting energy consumption forecast, the historical average energy consumption for the same period is included. =80kW, actual temperature =28℃, standard temperature =25℃, temperature correction factor =0.03, actual humidity =55%RH, standard humidity =50%RH, humidity correction factor =0.015, actual light intensity =600 lux, standard illuminance =500 lux, illumination correction factor =0.006, calculated as follows Risk diffusion prediction uses the formula If the lighting circuit fault currently affects the area =100㎡, diffusion coefficient α=0.01, low equipment coupling, Δt=600s, current abnormal energy consumption. =40kW, rated energy consumption =25kW, calculated as S(600s)=100×(1+0.01×600×40 / 25)=100×(1+0.01×960)=100×10.6=1060㎡. Warning information is pushed according to risk level. Medium-risk anomalies are pushed with text prompts, investigation steps, and historical similar cases. Multi-level acceptance feedback—discovery-handling-closure—is supported. If the loop is not closed, the anomaly status is updated every 2 minutes.

[0045] System self-optimization and iterative module configuration: A dual-loop optimization mechanism is constructed. The inner loop fine-tunes model hyperparameters such as learning rate and window size hourly using real-time data, while the outer loop retrains the model monthly based on full data. A reinforcement learning optimization and diagnostic strategy is adopted, using "minimum false alarm rate + fastest response time" as the reward function to intelligently adjust feature weights and decision thresholds. An equipment health assessment submodule is established, introducing remaining lifespan prediction. Based on LSTM neural network analysis of energy consumption data, operating temperature, and vibration frequency, it predicts the remaining lifespan of photovoltaic inverters with an error ≤8%, and pushes replacement plans 30 days in advance. A federated learning framework is used for joint training across multiple building nodes to protect data privacy. A monthly energy consumption analysis report is generated, including anomaly statistics, energy-saving potential analysis, and optimization suggestions. For the problem of inefficient central air conditioning operation, optimized operation strategies are recommended, with an estimated annual energy saving rate of 8%.

[0046] 2. Effect Verification and Table Analysis Table 2 Key Operational Indicators of the Smart Office Building System in Example 2 Exception types Diagnostic accuracy (%) Warning response time (s) Processing completion efficiency (min) False alarm rate (%) Missed report rate (%) Photovoltaic inverter malfunction (equipment-related) 98.5 3.9 22 1.5 1.0 Air conditioning parameter mismatch (latent) 97.2 4.5 30 2.1 1.5 Lighting circuit short circuit (sudden type) 99.0 3.6 10 1.2 0.9 Energy storage charging and discharging anomaly (continuous type) 96.8 4.3 18 2.4 1.6 Table 2 shows data based on 30 consecutive days of system operation monitoring, demonstrating the adaptability and operational effectiveness of this invention in smart office building scenarios. The accuracy rate for diagnosing various anomalies exceeds 96%, with anomalies such as photovoltaic inverter faults and lighting circuit short circuits achieving an accuracy rate of over 98.5%, proving the comprehensiveness of multi-source data collection and the effectiveness of the four-dimensional feature system, capable of accurately capturing anomaly patterns in both new energy and traditional energy-consuming equipment. The early warning response time is controlled within 4.5 seconds. Differentiated early warning content and a multi-level acceptance feedback mechanism ensure efficient handling by maintenance personnel; the handling efficiency for lighting circuit short circuit anomalies is only 10 minutes. The false alarm rate and false negative rate are controlled below 2.4% and 1.6% respectively, thanks to the collaborative diagnosis of hybrid AI algorithms and knowledge graphs, combined with interpretability analysis, improving the reliability and credibility of diagnostic results. The system's self-optimization mechanism and remaining life prediction function can adapt to the energy consumption patterns of office buildings, continuously optimizing the diagnostic model, providing scientific support for energy consumption management and energy-saving potential exploration in new energy buildings, and contributing to the green and low-carbon operation of buildings.

[0047] Reference Figure 2 This figure visually demonstrates the system's diagnostic capabilities for various energy consumption anomalies. Data shows that the diagnostic accuracy for all anomaly types exceeds 96%, with sudden anomalies such as gas leaks achieving 99.2% accuracy and equipment-related anomalies such as photovoltaic inverter failures reaching 98.5%. This result confirms the effectiveness of the hybrid AI algorithm and four-dimensional feature system—by integrating LSTM neural networks, gradient boosting trees, and knowledge graphs, the system can accurately capture the core features of different types of anomalies, maintaining a high accuracy of 96.5% even for latent anomalies (such as decreased air conditioning heat exchange efficiency). This solves the problem of insufficient ability of traditional single algorithms to identify complex anomalies, providing a reliable basis for subsequent risk classification and early warning.

[0048] Reference Figure 3 This figure reflects the actual effect of the system's self-optimization mechanism. As the running time increased from 10 days to 50 days, the false alarm rate continuously decreased from 3.5% to 1.8%, showing a significant optimization trend. This is attributed to the dual-loop optimization mechanism: the inner loop fine-tunes hyperparameters hourly, and the outer loop retrains the model monthly, combining reinforcement learning to adjust feature weights with the goal of "minimizing the false alarm rate." Data demonstrates that the system can continuously iterate through accumulated operational data, gradually adapting to the dynamic changes in building energy consumption patterns, solving the problem of fixed models and high false alarm rates in traditional systems, and significantly improving long-term operational stability.

[0049] Reference Figure 4 This graph illustrates the differentiated response capabilities of the early warning push module. As the risk level increases, the response time decreases from 4.0 seconds for no risk to 0.8 seconds for extremely high risk, and only 1.2 seconds for high-risk anomalies. This is because the system dynamically adjusts its push strategy based on the risk level: high-risk anomalies are pushed through multiple channels including "APP + SMS + voice" and trigger a rapid response mechanism, while no-risk anomalies are only recorded locally. The data shows that the system can prioritize the efficiency of handling high-risk anomalies, solving the problem of critical information delays caused by the traditional "one-size-fits-all" approach to early warnings, and buying time for the rapid handling of emergencies (such as gas leaks).

[0050] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence algorithm-based building energy consumption anomaly diagnosis and early warning system, characterized in that, The system comprises the following modules: Multi-source energy consumption and environmental data acquisition module, deploying smart meters, water meters, gas meters, cold and heat meters and multi-dimensional environmental sensors to form a heterogeneous collection network; Various instruments and sensors collect relevant data according to corresponding standards, all devices access edge computing gateway through LoRaWAN protocol, gateway supports communication, data is stored according to building, building, floor and room four-level granularity, key equipment deploys redundant collection nodes; Data preprocessing module, using time series sliding window and intelligent cleaning combined processing architecture, removing gross errors by pull-in criterion, filling missing data by improved KNN algorithm, processing noise by wavelet threshold denoising algorithm; Establish a data quality evaluation system; Multi-dimensional energy consumption feature extraction module, construct four-dimensional feature system of time domain, frequency domain, space domain and time series, cover all kinds of core features, extract special features for different energy systems; Abnormal diagnosis module based on hybrid AI algorithm, fusion of improved LSTM neural network, gradient boosting tree and knowledge graph construction collaborative diagnosis framework; Input four-dimensional features and environmental parameters, output abnormal probability, abnormal type and abnormal confidence, identify multiple types of abnormalities; Abnormal risk grading module, build a multi-factor coupled risk assessment model, divide the risk level into four levels, adjust the threshold dynamically combined with the energy period, calculate the risk value by weighted summation; Early warning information generation and push module, generate differentiated content according to risk level, early warning information contains core key information, international project supports English and Chinese switching; System self-optimization and iteration module, build a data-driven whole life cycle optimization mechanism, optimize the model through incremental learning and federated learning, establish device health evaluation sub-module and digital twin verification mechanism. 2.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm according to claim 1, characterized in that, Further comprising an energy consumption anomaly degree quantification submodule, the cumulative influence of the energy consumption anomaly in a unit time is calculated by integration, and the formula is Wherein Q is the cumulative influence value of the anomaly, is the anomaly start time, is the anomaly end time, is the actual energy consumption function changing with time, is the normal energy consumption prediction value, is the time weight, is the space weight. 3.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm according to claim 1, characterized in that, A normal energy consumption dynamic prediction sub-module is also included, which fuses historical data and real-time environmental parameters to construct a prediction model, with the formula being wherein is a normal energy consumption prediction value at time t, is a historical same-period energy consumption average value, is a temperature correction coefficient, is a humidity correction coefficient, is an illumination correction coefficient, , , are actual temperature, humidity, and illumination intensity, respectively, = 25℃, = 50% RH, = 500 lux are standard values. 4.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm according to claim 1, characterized in that, Multi-source energy consumption and environmental data acquisition module adopts edge and cloud collaborative architecture, edge nodes are deployed in building weak electricity room, with local data preprocessing ability such as preliminary filtering of outliers and data compression; Sensor supports self-diagnosis; For old building renovation scene, provide wireless attached sensor with magnetic and adhesive double fixation method, installation without wiring; Establish sensor health file. 5.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm according to claim 1, characterized in that, Data preprocessing module introduces time series anomaly pre-detection mechanism, uses isolation forest algorithm to preliminarily screen the original data and identify potential abnormal segments; For seasonal buildings such as shopping malls and hotels, use seasonal and trend decomposition algorithm to separate seasonal, trend and residual components in energy consumption data, and process different components of noise respectively; The missing data filling introduces space-time correlation, when single node data is missing, reference the same period data of adjacent 3 nodes in the same area to calculate the distance, the closer the weight is higher. 6.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm according to claim 1, characterized in that, Multi-dimensional energy consumption feature extraction module increases cross-device correlation features, calculates the energy consumption ratio of central air conditioning and cooling tower, the correlation of water pump flow and pressure, and the linkage rate of lighting system and human body sensor; Introduce equipment operation efficiency features, such as COP of water chiller, boiler thermal efficiency and transformer load rate; For new energy buildings, extract photovoltaic panel power generation capacity and load matching degree, energy storage device charging and discharging efficiency and micro-grid power balance degree features. 7.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm of claim 1, wherein, The abnormal diagnosis module based on the hybrid AI algorithm introduces an explanatory enhancement mechanism, adds a SHAP value analysis layer at the output end of the model, calculates the contribution of each feature to the diagnosis result, and takes positive values as promoting abnormal judgment and negative values as inhibiting. Features with an absolute contribution value greater than 0.2 are listed as key influencing factors. The knowledge graph adopts a dynamic updating mechanism, and new abnormal cases are automatically stored in the database after being confirmed by experts. Associated features are matched to the existing rule base through cosine similarity. For complex abnormalities, i.e., multi-device linkage failures, a causal inference algorithm is used to trace the root cause. 8.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm of claim 1, wherein, The abnormal risk grading module introduces risk spread prediction, constructs a spread model based on device topology relationship and historical data, and the formula is wherein is the influence range after , is the current influence range, and α is the spread coefficient, is the current abnormal energy consumption, is the rated energy consumption of the device. 9.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm of claim 1, wherein, The early warning information generation and pushing module integrates intelligent linkage control functions. After high-risk abnormalities are confirmed, control instructions are automatically sent to the building automation system. The system supports customizing early warning strategies. Early warning information is accessed to the building BIM model, and the abnormal position is highlighted in the three-dimensional view. The surrounding associated devices and evacuation channels are labeled, and the historical energy consumption curve and maintenance records can be viewed by clicking on the device. 10.The building energy consumption anomaly diagnosis and early warning system based on artificial intelligence algorithm of claim 1, wherein, The system self-optimization and iteration module constructs a double closed-loop optimization mechanism. The inner closed loop adjusts the model hyperparameters every hour through real-time data, and the outer closed loop re-trains the model based on full data every month. Reinforcement learning is used to optimize the diagnosis strategy, and the feature weights and decision thresholds are intelligently adjusted. The device health assessment introduces residual life prediction based on LSTM neural network analysis of energy consumption data, operating temperature, and vibration frequency to predict the remaining useful life of the device.

Images