Building life cycle energy consumption real-time management and control method based on edge computing

Abstract

The present application relates to the technical field of building energy consumption management and control, and particularly relates to a building full life cycle energy consumption real-time management and control method based on edge computing, which constructs a benchmark energy consumption library and a dynamic management and control framework, collects data in real time through an edge node, compares the data with benchmark values, forms a deviation matrix, quantifies the influence of external factors, identifies the root cause of the deviation through multi-level attribution, and finally generates and executes a precise control strategy on the edge side based on the analysis results, forms an intelligent closed loop of "comparison- attribution-regulation", realizes a paradigm shift from traditional energy consumption monitoring and control to precise attribution control based on multi-level root cause diagnosis, and thus achieves real-time energy efficiency optimization in a self-adaptive and root cause treatment manner in the full life cycle of a building.

Description

Technical Field

[0001] This invention relates to the field of building energy consumption management technology, and in particular to a method for real-time management of building energy consumption throughout its entire life cycle based on edge computing. Background Technology

[0002] Currently, some technical solutions are attempting to leverage edge computing technology to improve the real-time performance and efficiency of building energy consumption management. These solutions typically use edge computing nodes deployed locally on the building to directly collect operational data from subsystems such as HVAC, lighting, and power, and perform real-time processing and rapid feedback control at the network edge. Compared to the traditional approach of uploading all data to the cloud for processing, this method effectively reduces control latency, alleviates network bandwidth pressure, and can, to some extent, achieve device-level energy efficiency optimization based on simple rules or local models. In addition, some solutions attempt to combine building information model data with operational data in order to provide a reference throughout the building's entire lifecycle. However, existing real-time management methods often rely on preset fixed rules or simple comparisons with static design values ​​to generate control strategies. This leads to the gradual disconnect between the performance evaluation benchmarks on which building energy consumption management systems depend and the actual physical state and operating environment of buildings. As a result, control strategies based on static models or fixed thresholds are in a state of "inaccuracy" for a long time. This not only causes a mismatch between control commands and actual needs, leading to excessive energy consumption or a decline in comfort, but also makes it impossible for the system to achieve continuous and adaptive energy efficiency optimization throughout the entire life cycle of building aging, climate change, and usage pattern evolution. It also makes it impossible to make adaptive benchmark adjustments and strategy iterations based on the actual operating performance of buildings, dynamic changes in the external environment, and equipment performance degradation. Summary of the Invention

[0003] The purpose of this invention is to provide a real-time energy consumption management and control method for the entire life cycle of buildings based on edge computing. By constructing a closed-loop logic of "comparison-attribution-regulation", it realizes a paradigm shift from traditional energy consumption monitoring and control to precise attribution control based on multi-level root cause diagnosis, thereby achieving adaptive and root cause-based real-time energy efficiency optimization throughout the entire life cycle of buildings.

[0004] To achieve the above objectives, this invention provides a method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing, comprising the following steps: Based on historical and design data from each stage of the building's entire life cycle, a multi-dimensional benchmark energy consumption dynamic curve library is constructed, and a dynamic control logic framework with multiple mechanisms is established. During the building operation phase, energy consumption and environmental data of various components within the building are collected in real time through deployed edge computing nodes. The collected energy consumption and environmental data of the building's internal components are compared with the theoretical values ​​under the corresponding conditions in the benchmark energy consumption dynamic curve library to generate a real-time energy consumption deviation matrix. Construct an engine for analyzing the correlation of influencing factors, access and analyze multidimensional external data on meteorology and population in real time, quantify their dynamic correlation with energy consumption deviation, and drive the adjustment of edge-side control strategies; Establish a multi-level energy consumption attribution analysis framework to decompose the total energy consumption deviation into equipment level, system level, behavior level and environmental response level, and identify the root causes of the deviation; Based on the above comparison, analysis and attribution results, and by integrating the various mechanisms in the dynamic control logic framework, targeted real-time control strategies are dynamically generated and executed.

[0005] Specifically, the construction of the multi-dimensional benchmark energy consumption dynamic curve library includes: Based on the design parameters in the building information model, the theoretical baseline energy consumption curves of the building are calculated under different seasons, different time periods and different usage rates. Integrate historical meteorological data, typical year data, and operational data of similar buildings in the building's location to form a comprehensive reference database for comparison; The dynamic curve library supports adaptive iterative updates based on actual building operation feedback and lifecycle stages.

[0006] Specifically, the real-time collection of energy consumption and environmental data within the building through deployed edge computing nodes includes: Edge computing nodes deployed at various levels in the building directly collect sub-metering data on lighting, air conditioning, and power within their respective areas, as well as environmental data on temperature, humidity, illuminance, and the presence of people. The raw data collected at the edge is preprocessed locally, including data cleaning, time alignment and format standardization, to form a structured local data set.

[0007] Specifically, generating the real-time energy consumption deviation matrix includes: Calculate the absolute and relative percentage deviations between the actual energy consumption value and the corresponding theoretical value in the benchmark curve library, using a preset time granularity. The deviation value is spatiotemporally correlated and mapped with the collected internal environmental parameters and external influencing factors to form a multidimensional matrix containing the deviation value, the occurrence time, and the related factors. Based on the persistence, fluctuation range, and scope of impact of the deviation, the deviation type is automatically classified and labeled.

[0008] Specifically, the construction of the influencing factor correlation analysis engine includes: Establish real-time interfaces with multi-source external data services, including meteorology, demographics, calendar events, and energy prices, to acquire structured external data; Using time series-based association rule mining and regression analysis, the correlation coefficients and influence weights of each external variable with the total building energy consumption and the deviations of each component energy consumption are dynamically quantified. Based on the analysis results, a dynamic allocation model for the weights of influencing factors is generated, providing a quantitative basis for adjusting the marginal side strategy.

[0009] Specifically, the establishment of the multi-level energy consumption attribution analysis framework includes: Equipment-level attribution: Analyze the deviation between the actual operating energy efficiency ratio and the rated or designed energy efficiency ratio of key energy-consuming equipment, and correlate it with equipment operating time, load rate and maintenance records; System-level attribution: Analyze the collaborative operating efficiency within and between HVAC and lighting subsystems, and identify system performance losses caused by control logic conflicts and hydraulic imbalances; Behavioral attribution: Based on spatial usage sensor data, analyze the impact of building occupancy patterns and user behavior of unintended device activation on energy consumption patterns; Environmental response attribution: Analyze the performance of passive design of building envelope, shading and natural ventilation under actual climatic conditions and assess its deviation from design expectations.

[0010] The dynamic management and control logic framework includes an adaptive baseline update mechanism and a climate adaptation strategy generation mechanism. The adaptive benchmark update mechanism periodically calibrates the benchmark energy consumption dynamic curve library based on long-term actual operation data, uses machine learning methods to establish an energy consumption prediction model that reflects the personalized operation characteristics of the building, and updates the benchmark reference value accordingly. The climate adaptability strategy generation mechanism predicts the future load trend of buildings based on short-term weather forecasts and historical similar day analysis, and accordingly presets or dynamically adjusts the operating parameters of the HVAC system at the edge.

[0011] The dynamic management and control logic framework also includes a population dynamic response mechanism, a cross-validation diagnostic mechanism, and a life cycle stage adaptive adjustment mechanism. The population dynamic response mechanism: Based on real-time monitoring of population distribution density and flow trends, dynamically adjusts environmental control targets and the working status of energy-consuming equipment in relevant areas to achieve on-demand supply; The cross-validation diagnostic mechanism employs a simulation model based on physical mechanisms and a statistical model based on operational data in parallel to calculate and predict energy consumption. By comparing the consistency of the two results, sensor failures, model inaccuracies, or hidden performance issues are diagnosed. The life cycle stage adaptive adjustment mechanism automatically identifies the life cycle stage of the building based on its years of operation and records of major equipment upgrades and renovations. Different benchmark curves are used at different stages to compare the stringency of the comparison, the aggressiveness of the control strategy, and the compensation algorithm for equipment performance degradation, so that the management logic matches the actual state of the building.

[0012] Specifically, the dynamic generation and execution of targeted real-time control strategies include: Based on the aforementioned deviation matrix, correlation analysis results, and attribution conclusions, a specific set of control instructions is synthesized locally at the edge computing node. The strategy generation takes into account multiple objectives, including real-time energy efficiency optimization, equipment lifespan, and indoor environmental quality. After logical verification, the generated strategy is directly distributed to the field controller or actuator through the edge node to form a low-latency closed-loop control. At the same time, the strategy summary and execution effect are synchronized to the cloud platform for long-term analysis and framework optimization.

[0013] Specifically, synthesizing a specific set of control instructions locally at the edge computing node includes generating the following control instructions: Operating parameter adjustment instructions: Based on climate adaptability strategies and real-time environmental feedback, dynamically adjust the temperature setpoint, chilled water supply temperature, damper opening, and fresh air ratio of the HVAC system; Equipment mode switching command: Based on the behavioral attribution analysis results and the dynamic response mechanism of personnel, control the dimming mode, zone start / stop of the lighting system, and the intermittent operation or sleep mode of the power equipment. System coordinated control commands: Based on system-level attribution conclusions, generate and issue linkage adjustment commands to eliminate control conflicts between systems and optimize hydraulic or air volume balance; Abnormal operating condition handling instructions: In response to abnormal sensor data or hidden performance degradation of equipment identified through cross-validation diagnostic mechanism, trigger equipment calibration, backup sensor switching or degraded operation contingency plan.

[0014] This invention presents a real-time energy consumption management method for buildings throughout their entire lifecycle based on edge computing. It constructs a multi-dimensional benchmark energy consumption dynamic curve library based on historical and design data from each stage of the building's lifecycle and establishes a dynamic management logic framework with multiple mechanisms. During the building operation phase, it collects real-time energy consumption and environmental data for various components within the building through deployed edge computing nodes. This collected data is compared with theoretical values ​​under corresponding conditions in the benchmark energy consumption dynamic curve library to generate a real-time energy consumption deviation matrix. Simultaneously, it constructs an influencing factor correlation analysis engine to access and analyze multi-dimensional external data such as meteorological and population data in real time. By establishing a dynamic correlation between energy consumption deviation and the overall energy consumption deviation, the adjustment of edge-side control strategies is driven. A multi-level energy consumption attribution analysis framework is established, decomposing the total energy consumption deviation into equipment, system, behavior, and environmental response levels to identify the root causes of the deviation. Based on the above comparison, analysis, and attribution results, and by integrating the various mechanisms in the dynamic control logic framework, targeted real-time control strategies are dynamically generated and executed. By constructing a closed-loop logic of "comparison-attribution-control," a paradigm shift from traditional energy consumption monitoring and control to precise attribution control based on multi-level root cause diagnosis is achieved, thereby realizing adaptive, root cause-based real-time energy efficiency optimization throughout the building's entire life cycle. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is a flowchart of the real-time energy consumption management method for the entire life cycle of a building based on edge computing, as described in this invention.

[0017] Figure 2 This is a flowchart of the real-time energy consumption management method for the entire life cycle of buildings based on edge computing, as described in this invention.

[0018] Figure 3 This is a flowchart of the data processing and benchmark comparison of the present invention.

[0019] Figure 4 This is a flowchart of the root cause diagnosis and dynamic mechanism of the present invention.

[0020] Figure 5 This is a flowchart of the strategy generation and execution control of the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated 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 intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] In the description of this invention, it should be understood that "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] Please see Figures 1 to 5 The present invention provides a method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing, comprising the following steps: S1: Based on historical and design data from each stage of the building's entire life cycle, construct a multi-dimensional benchmark energy consumption dynamic curve library and establish a dynamic control logic framework with multiple mechanisms. S2: During the building operation phase, energy consumption and environmental data of various components within the building are collected in real time through deployed edge computing nodes; S3: Compare the collected energy consumption and environmental data of the building with the theoretical values ​​under the corresponding conditions in the benchmark energy consumption dynamic curve library to generate a real-time energy consumption deviation matrix. S4: Build an influencing factor correlation analysis engine to access and analyze multi-dimensional external data of meteorology and population in real time, quantify its dynamic correlation with energy consumption deviation, and drive the adjustment of edge-side control strategies. S5: Establish a multi-level energy consumption attribution analysis framework to decompose the total energy consumption deviation into equipment level, system level, behavior level and environmental response level, and identify the root causes of the deviation. S6: Based on the above comparison, analysis and attribution results, and by integrating the various mechanisms in the dynamic control logic framework, a targeted real-time control strategy is dynamically generated and executed.

[0024] In this embodiment, taking a smart office building that has been put into operation as an example, the system first generates a theoretical energy consumption benchmark curve library covering 8760 hours a year, distinguishing between weekdays and holidays, and different outdoor temperature and humidity ranges, based on its BIM design model, local ten-year historical meteorological data and energy consumption statistics of similar buildings. It also incorporates multiple intelligent mechanisms such as adaptive updates.

[0025] During operation, edge computing gateways deployed in floor power distribution rooms, air conditioning rooms, and other locations collect data per second from sub-meters for lighting, sockets, chillers, and fresh air units, as well as sensor data on temperature, humidity, CO2 concentration, and the presence of people in each area.

[0026] The system does not simply compare the current total energy consumption with a fixed value. Instead, it accurately retrieves the corresponding theoretical energy consumption value from the benchmark library based on the specific time, outdoor weather conditions, and building reservation occupancy rate, and calculates the real-time deviation.

[0027] Meanwhile, an independent impact factor analysis engine continuously obtains data such as 72-hour forecasts and real-time urban heat maps from meteorological bureaus and population big data platforms to analyze their potential impact on building heating and cooling loads.

[0028] More importantly, when the system detects persistently high energy consumption, it will initiate root cause diagnosis to analyze whether the cause is a decrease in the efficiency of a certain chiller, hydraulic imbalance in the air conditioning water system, some employees leaving the equipment on after get off work, or the failure of the building's west-facing glass curtain wall to meet thermal insulation expectations.

[0029] Ultimately, by integrating all these real-time diagnostic conclusions, the system automatically generates and issues a series of precise control commands at the edge side, such as "while ensuring the comfort of the core area, increase the set temperature of the air conditioning terminal in the west area by 1°C between 2 and 5 pm", "based on the prediction of pedestrian flow, reduce the lighting brightness of the low-usage floors in Building B 15 minutes in advance", and "start the performance compensation algorithm for chiller unit No. 3 and generate a maintenance work order", forming a complete intelligent closed loop from "perceiving abnormalities" to "diagnosing the root cause" to "targeted treatment".

[0030] Furthermore, the construction of the multi-dimensional benchmark energy consumption dynamic curve library specifically includes: Based on the design parameters in the building information model, the theoretical baseline energy consumption curves of the building are calculated under different seasons, different time periods and different usage rates. Integrate historical meteorological data, typical year data, and operational data of similar buildings in the building's location to form a comprehensive reference database for comparison; The dynamic curve library supports adaptive iterative updates based on actual building operation feedback and lifecycle stages.

[0031] In this embodiment, the BIM model of the building is first used to calculate the hourly theoretical energy consumption under standard operating conditions using energy consumption simulation software, which serves as the "design benchmark". The system will further integrate ten years of historical hourly data provided by the local meteorological station to generate a "typical meteorological year benchmark". At the same time, the median of anonymized monthly energy consumption of buildings of the same climate zone and functional type is obtained from the cloud platform to form an "industry reference benchmark". These three sets of benchmarks together constitute the initial "multidimensional benchmark curve library". For example, for the energy consumption of the air conditioning system at 10 am on a weekday, the library not only contains simulated values, but also values ​​corrected based on the historical average temperature for the same period, as well as the energy consumption statistics range of similar buildings during that time period. After the system has been running for a year, it was found that the air conditioning energy consumption of this building at night in summer was consistently lower than the "design benchmark" but met the "industry reference benchmark". After analysis, it was found that this was because the actual shading effect of the building was better than the design expectation. At this time, the adaptive update mechanism was activated. Using the actual operating data of the past year, a machine learning prediction model specifically for this building was trained. This model will gradually replace part of the initial design benchmark, making the theoretical comparison values ​​closer to the actual physical characteristics and usage habits of this building, thereby making the deviation analysis more accurate.

[0032] Adaptive iterative updates do not run continuously; instead, they are intelligently initiated by preset triggering conditions. These triggering conditions are divided into two categories: periodic triggers and event triggers.

[0033] The system pre-sets three time-based periodic update tasks: Short-term update: Performed automatically once a month to capture the slow drift of building energy consumption patterns caused by seasonal changes and variations in usage intensity.

[0034] Mid-term update: Performed automatically once per quarter to accumulate enough data samples to train a more stable predictive model.

[0035] Long-term update: Automatically executed once a year to comprehensively assess the impact of changes in building life cycle stages on energy consumption benchmarks.

[0036] The system will proactively trigger adaptive updates when it detects any of the following situations: The actual operating energy consumption over 30 consecutive days deviated from the current baseline curve by a systematic deviation (mean absolute percentage error) exceeding the preset threshold of 15%.

[0037] Significant changes to a building include, but are not limited to: replacement or major repair of major energy-consuming equipment (such as chillers and boilers), energy-saving renovation of the building envelope, partial adjustment of the building's function (such as converting office floors into laboratories), or a long-term significant deviation of the population density from the design value.

[0038] The multi-level attribution analysis framework consistently identifies "environmental response-level attribution" or "equipment-level attribution" indicating a systematic deviation between the actual performance of the building and the design expectations, and this deviation persists for more than ninety days.

[0039] The life cycle stage adaptive adjustment mechanism determines that a building has entered a new stage (such as from "initial operation" to "stable operation" or "equipment aging"), at which point the baseline curve needs to be updated to match the energy consumption characteristics of the new stage.

[0040] When any of the above triggering conditions are met, the system executes the following process to train a dedicated machine learning prediction model: The first step is historical data preparation. The system extracts operational data from a continuous period of the past from the time-series database as a training sample set. The start and end times of the sample set are determined according to the trigger type: periodic updates typically use data from the past 90 days to one year; event-driven updates use data from 180 days before the event as the baseline period, and data from more than 30 days after the event as the validation period. The extracted data fields include: outdoor meteorological parameters (temperature, humidity, solar radiation intensity) at each time point (at a granularity of 15 minutes or one hour), building internal operational parameters (average temperature, humidity, personnel density, operating status and load rate of major equipment in each area), and the corresponding total building energy consumption and the energy consumption values ​​of each component.

[0041] The second step is feature engineering. The system converts the raw data into feature vectors usable by the prediction model. Each sample corresponds to a point in time, and its features include: time-related features (hours, day of the week, whether it is a holiday, season), meteorological features (current outdoor temperature, temperature trend over the past hour, cumulative cooling hours or heating hours for the day), and status-related features (current population density level, average load rate of major equipment, difference between the average indoor temperature and the set temperature). Simultaneously, to capture the time-series dependence of energy consumption, the system also constructs autoregressive features, i.e., the actual energy consumption values ​​at the same time point on the previous time and the same time point on the previous day of the same type.

[0042] The third step is model training and validation. The system uses the gradient boosting tree algorithm from ensemble learning as the basic prediction model. During training, historical data is randomly divided into a training set (80%) and a validation set (20%). The model takes feature vectors as input and the actual energy consumption value at the corresponding time point as the output target, and minimizes the prediction error through iterative optimization. After training, the system uses the validation set to calculate the model's prediction accuracy, including mean absolute percentage error (MASE) and root mean square error (RMSE). If the validation accuracy meets the preset requirements (e.g., MSE is less than 10%), the model passes acceptance.

[0043] The fourth step is the fusion and replacement of baseline curves. The newly trained dedicated model does not completely replace all the original baseline curves, but adopts a gradual fusion strategy. The system weights and fuses the predicted values ​​of the new model with the original design baseline curves and industry reference baseline curves. The fusion weights are dynamically adjusted based on the model's performance on the validation set: the higher the prediction accuracy of the new model, the greater its weight; conversely, the original baselines remain dominant. For situations where there are significant changes in the building life cycle stages, the system will significantly increase the weight of the new model, allowing it to more quickly dominate the baseline reference values. The fused new baseline curves take effect immediately after synchronization with edge nodes and are used for subsequent real-time deviation comparisons. At the same time, the old model is archived and saved for future retrospective analysis.

[0044] Through the aforementioned activation mechanism that combines periodic and event-based triggering, and the complete process of extracting features from historical data, training gradient boosting tree models, and progressively merging and replacing them, the adaptive iterative update of this invention can calibrate the benchmark energy consumption curve library in a timely and accurate manner at different stages of the building's entire life cycle, ensuring that it always reflects the actual performance characteristics of the building. This ensures that subsequent energy consumption comparisons, attribution analyses, and management strategy generation are always based on a scientific and dynamic reference benchmark.

[0045] Furthermore, the real-time collection of energy consumption and environmental data within the building through deployed edge computing nodes specifically includes: Edge computing nodes deployed at various levels in the building directly collect sub-metering data on lighting, air conditioning, and power within their respective areas, as well as environmental data on temperature, humidity, illuminance, and the presence of people. The raw data collected at the edge is preprocessed locally, including data cleaning, time alignment and format standardization, to form a structured local data set.

[0046] In this embodiment, when used in a large commercial complex, edge computing nodes are deployed in a hierarchical manner: nodes deployed at the substation level are responsible for collecting data on the building's total power consumption and main power circuits; nodes deployed in the low-voltage rooms on each floor collect data on the energy consumption of lighting, sockets, and air conditioning on that floor; and nodes deployed in the air conditioning room are specifically responsible for collecting detailed operating parameters (current, frequency, water temperature, etc.) of chillers, cooling towers, and water pumps.

[0047] These nodes are directly connected to smart meters and sensors in the field via industrial bus or IoT protocol, and collect data in cycles of 1-5 minutes. The raw data collected may contain noise such as instantaneous communication errors and range jumps.

[0048] Edge nodes immediately perform preprocessing locally: for example, threshold judgment and moving average filtering are used to clean up obvious outlier data points; electricity meter data and temperature data from different brands and different collection periods are aligned and resampled according to a unified timestamp; finally, the processed data is packaged into a standardized JSON or time-series data format.

[0049] This edge preprocessing significantly reduces the amount of data and redundant information uploaded to the cloud, alleviates network bandwidth pressure, and provides a high-quality, low-latency local data source for subsequent real-time comparison and analysis.

[0050] Furthermore, the generation of the real-time energy consumption deviation matrix specifically includes: Calculate the absolute and relative percentage deviations between the actual energy consumption value and the corresponding theoretical value in the benchmark curve library, using a preset time granularity. The deviation value is spatiotemporally correlated and mapped with the collected internal environmental parameters and external influencing factors to form a multidimensional matrix containing the deviation value, the occurrence time, and the related factors. Based on the persistence, fluctuation range, and scope of impact of the deviation, the deviation type is automatically classified and labeled.

[0051] In this embodiment, the system uses a 15-minute analysis cycle. At the end of each cycle, the edge server located in the building management layer will perform a comparison calculation. For example, the actual total energy consumption in the current cycle is 850 kWh, while the theoretical value found in the benchmark database based on the current time (Wednesday afternoon 2 pm), outdoor temperature (32°C), and building reservation occupancy rate (80%) is 780 kWh.

[0052] The system calculated an absolute deviation of +70 kWh and a relative deviation of approximately +9%. Subsequently, the system mapped this deviation value to a series of related factors: internal factors (such as an average indoor temperature of 26.5℃ and high population density in the core area) and external factors (solar radiation intensity and humidity at this moment). All this information constitutes a "vector" in the deviation matrix. This matrix uses a preset time granularity (e.g., 15 minutes) as its time axis, with each row representing a deviation event vector within an analysis period. The dimensions of this vector include at least: Deviation quantification dimension: Store the absolute deviation (kWh) and relative deviation percentage (%) of the total energy consumption and the energy consumption of each item such as lighting, air conditioning, and power relative to the benchmark value within this period.

[0053] Spatiotemporal attribute dimension: Spatial location information such as the specific timestamp of the occurrence of the correlation deviation, the building area or system to which it belongs, and the equipment identifier.

[0054] Internal environment correlation dimension: Maps internal state parameters that may be related to deviations within this period, including but not limited to real-time sensor data such as regional average temperature, humidity, illuminance, CO2 concentration, personnel density, and personnel mobility.

[0055] External factors are associated with the following dimensions: mapping the outdoor meteorological parameters (temperature, humidity, solar radiation intensity, wind speed and direction), air quality index, calendar event types (weekdays / holidays, large events), real-time electricity price levels, and gridded population heat index, etc., acquired in the same period.

[0056] Deviation feature label dimensions: Based on the duration, fluctuation amplitude, trend of change (stable, increasing, decreasing) and spatial distribution range of the deviation, the deviation type label is automatically generated through built-in classification logic, such as "instantaneous fluctuation", "persistent deviation", "divergent anomaly", "regional anomaly", etc.

[0057] The system continuously tracks changes in the deviation vector: if a deviation of +9% lasts for only one cycle, it may be marked as "instantaneous fluctuation"; if it persists for four consecutive cycles (1 hour) with a stable amplitude, it is marked as "persistent deviation"; if the deviation amplitude continues to increase, it may be marked as "divergent anomaly". This matrix-based management approach enables the system to clearly understand the patterns, trends, and accompanying characteristics of energy consumption anomalies, providing structured input for subsequent attribution.

[0058] Furthermore, the construction of the influencing factor correlation analysis engine specifically includes: Establish real-time interfaces with multi-source external data services, including meteorology, demographics, calendar events, and energy prices, to acquire structured external data; Using time series-based association rule mining and regression analysis, the correlation coefficients and influence weights of each external variable with the total building energy consumption and the deviations of each component energy consumption are dynamically quantified. Based on the analysis results, a dynamic allocation model for the weights of influencing factors is generated, providing a quantitative basis for adjusting the marginal side strategy.

[0059] In this embodiment, the influencing factor correlation analysis engine runs as a cloud or regional edge service, and obtains detailed weather forecasts (temperature, humidity, wind speed, cloud cover), holiday information, and the gridded population activity index of the area for the next 24-72 hours from the public data platform through API on a regular basis.

[0060] The engine maintains a time-series database that stores historical external data and building energy consumption data. Every week, the engine runs a correlation analysis. For example, by analyzing data from the past month, it may find that: "The total building energy consumption during the weekday morning rush hour (8:00-10:00) has a correlation coefficient of 0.85 with the outdoor temperature of the previous hour and a correlation coefficient of 0.7 with the real-time subway passenger flow"; "During peak electricity price periods, the energy consumption of lighting sockets tends to decrease proactively."

[0061] Based on these analyses, the engine dynamically generates and distributes a weighted model. In summer, the weight of the temperature factor is increased; on days with large events, the weight of the population factor is significantly increased. When the engine predicts that a severe convective weather event will cause a sharp drop in temperature tomorrow afternoon, it will notify the local edge nodes of this information and the "change in temperature factor weight" in advance. When the edge nodes formulate the pre-cooling strategy for tomorrow afternoon, they will refer to this adjusted weighted model to more accurately assess the impact of the temperature drop on the load reduction, thereby generating a more reasonable equipment operation plan.

[0062] The core function of this influencing factor correlation analysis engine is to dynamically correlate multi-source external data with building energy consumption deviation data and generate a quantitative weight model that can be used for edge-side strategy adjustments. Its specific workflow consists of three stages: data preprocessing, association rule mining and regression analysis, and dynamic weight allocation.

[0063] The engine first converts the raw data obtained from various external interfaces into a unified format. Meteorological data is resampled at an hourly granularity, and parameters such as hourly temperature, humidity, wind speed, cloud cover, and solar radiation are organized into a continuous time series. Holiday information is encoded as a binary variable (whether it is a holiday or not) and a holiday type weight (e.g., statutory holidays, adjusted workdays, ordinary weekends). The population activity index is derived from gridded mobile signaling or heat map data. The engine aggregates the population density values ​​of the grid where the building is located and its neighboring grids by hour, and simultaneously calculates the actual monitored personnel density data inside the building. The two are combined to form two related variables: "external population activity pressure" and "internal actual occupancy rate". In addition, calendar events (such as large conferences, exhibitions, and promotional activities) are also extracted from public calendars or building management systems as discrete event variables. All external variables are aligned with the building energy consumption deviation data (total energy consumption deviation and deviations of each component) collected from edge nodes using the same timestamp to form a multidimensional time series dataset for analysis.

[0064] The engine performs a deep analysis weekly, using data from the past four to eight weeks as the analysis window. The analysis process is divided into two phases.

[0065] Phase 1: Association Rule Mining. The engine employs a frequent pattern mining method based on time series analysis to identify the temporal correlation between changes in external variables and changes in energy consumption deviation. Specifically, the engine discretizes continuous data into state labels. For example, outdoor temperature changes are categorized into five states: "sudden rise," "gradual rise," "stable," "gradual fall," and "sudden fall"; population density is categorized into five levels: "extremely low," "low," "medium," "high," and "extremely high"; and energy consumption deviation is categorized into five intervals: "negatively large," "negatively small," "zero," "positively small," and "positively large." Subsequently, the engine scans the time series and calculates the probability of a certain pattern appearing in the energy consumption deviation state within one to three time points after a certain external state occurs. When this probability is significantly higher than a random distribution, an association rule is mined. For example, the engine might find: "Under the condition of a sudden rise in outdoor temperature and high population density, the probability of a positively large deviation in air conditioning energy consumption within the following hour is 85%." These rules are stored in the engine's knowledge base to help explain the causes of the deviation.

[0066] Phase Two: Regression Analysis. The engine simultaneously employs multivariate regression analysis to quantify the marginal contribution of each external variable to energy consumption deviation. Using the total energy consumption deviation or individual energy consumption deviations at each time point as the dependent variable, and external variables at the same time point and their lags as independent variables (e.g., current temperature, temperature change in the previous hour, current population density, holiday markers, etc.), the engine constructs a linear or nonlinear regression model. The regression coefficient for each independent variable is calculated using the least squares method or regularized regression algorithm. The absolute value of this coefficient reflects the strength of the variable's influence on energy consumption deviation, while the sign of the coefficient indicates the direction of influence (positive promotion or negative inhibition). For example, the regression analysis might output: a positive and large coefficient for the temperature variable, indicating that increased temperature significantly increases energy consumption deviation; while a negative coefficient for holiday markers, indicating that holidays reduce energy consumption deviation.

[0067] Based on the regression coefficients obtained from the above regression analysis and the confidence levels in the association rules, the engine generates a set of quantifiable influence factor weights. These weights include, but are not limited to, the following types: Temperature factor weight: This reflects the degree of influence of outdoor air temperature on building heating and cooling loads and corresponding energy consumption deviations. The weight is quantified by taking the absolute value of the regression coefficient of the temperature variable (and its lagged terms) in the regression analysis, normalizing it, and mapping it to the interval between zero and one. This weight is typically higher during the summer cooling season and automatically decreases during the spring and autumn transition seasons.

[0068] Humidity factor weight: Reflects the impact of air humidity on the latent heat load of the air conditioning system. Its quantification method is similar to that of the temperature factor, taking the absolute value of the regression coefficient of the humidity variable and normalizing it.

[0069] Population density factor weight: This reflects the impact of the degree of population concentration inside and outside the building on energy consumption. This weight consists of two parts: a weighted average of the regression coefficient of the external population activity index and the regression coefficient of the actual internal occupancy rate. A higher weight value indicates that greater attention should be paid to the driving effect of changes in population distribution on energy consumption when adjusting strategies.

[0070] Time factor weights include the weights for the impact of holidays and the weights for the impact of time periods. The weight for the impact of holidays is the absolute value of the regression coefficient of the holiday indicator variable; the weight for the impact of time periods is calculated hourly to reflect the differences in the impact of external factors on energy consumption at different times.

[0071] Calendar event weights: When there is a specific event (such as a large event or an important meeting), the regression coefficient of the dummy variable corresponding to the event is added as a temporary weight to the relevant factor.

[0072] Association rule confidence weight: For special combinations discovered through association rule mining (such as the simultaneous occurrence of high temperature and high humidity), the engine generates a combination factor weight, the value of which is equal to the condition confidence of the rule. When the same external combination conditions are detected in actual operation, this combination weight will be activated.

[0073] All individual weights are normalized before output, ensuring the sum of all factor weights is a fixed value (e.g., 100). The engine encapsulates the final weight vector (e.g., at the current time: temperature factor weight 0.35, humidity factor weight 0.15, population density factor weight 0.30, holiday factor weight 0.10, time period factor weight 0.10) in JSON format and distributes it to each edge computing node via a message queue. When generating control strategies, edge nodes adjust the priority of different objective functions based on these weights. For example, when the population density factor weight increases significantly, the strategy will more aggressively adjust fresh air and lighting to follow population movement; when the temperature factor weight dominates, the strategy will prioritize pre-cooling or pre-heating in response to temperature changes.

[0074] Through the complete process of data preprocessing, association rule mining and regression analysis, and dynamic weight allocation and quantification, the influencing factor association analysis engine can automatically update the weight model weekly, ensuring that the edge control strategy always keeps pace with the changing trends of the external environment.

[0075] Furthermore, the establishment of the multi-level energy consumption attribution analysis framework specifically includes: Equipment-level attribution: Analyze the deviation between the actual operating energy efficiency ratio and the rated or designed energy efficiency ratio of key energy-consuming equipment, and correlate it with equipment operating time, load rate and maintenance records; System-level attribution: Analyze the collaborative operating efficiency within and between HVAC and lighting subsystems, and identify system performance losses caused by control logic conflicts and hydraulic imbalances; Behavioral attribution: Based on spatial usage sensor data, analyze the impact of building occupancy patterns and user behavior of unintended device activation on energy consumption patterns; Environmental response attribution: Analyze the performance of passive design of building envelope, shading and natural ventilation under actual climatic conditions and assess its deviation from design expectations.

[0076] In this embodiment, when the real-time energy consumption deviation matrix shows that there is a continuous high energy consumption in the office area during the weekend daytime, the multi-level attribution analysis framework is triggered.

[0077] The term "occupancy pattern" refers to the temporal state and changing patterns of human presence in various areas within a building, as described by spatial usage sensor data. This pattern provides a quantitative characterization of human behavior and includes the following five dimensions of characteristic parameters: First, there is the status dimension. This refers to whether a certain area is occupied by people at a specific time. The occupancy status is divided into four levels: "vacant," "low occupancy," "medium occupancy," and "high occupancy." This level is dynamically determined based on the ratio of the number of people monitored in real time within the area to the area's designed capacity.

[0078] Second, the time distribution dimension. This refers to the distribution characteristics of occupancy status along the 24-hour timeline and the 7-day timeframe, including time-series indicators such as the initial occupancy time, continuous occupancy period, frequency of occupancy interruptions, and the last occupancy end time. For example, a typical occupancy pattern in office areas is characterized by continuous occupancy from 9:00 AM to 6:00 PM on weekdays, with a low-occupancy intermittent period from 12:00 PM to 1:00 PM.

[0079] Third, the spatial distribution dimension. This refers to the correlation characteristics of occupancy status between different areas, including the synchronicity of occupancy in adjacent areas, the temporal transmission relationship of occupancy along personnel flow paths, and the degree of coupling of occupancy status between different functional areas. For example, the occupancy pattern of meeting rooms is temporally correlated with the occupancy rate of adjacent office areas.

[0080] Fourth, the behavioral type dimension. Based on the rate of change and duration of occupancy status, occupancy patterns are classified into "stable stay pattern," "rapid passage pattern," "intermittent use pattern," and "abnormal clustering pattern." Stable stay pattern refers to people staying in the same area for a long time; rapid passage pattern refers to people passing through an area in a short time (such as a corridor); intermittent use pattern refers to people frequently entering and exiting (such as a tea room); abnormal clustering pattern refers to people density exceeding twice the historical statistical limit.

[0081] Fifth, the predictive pattern dimension. By performing time-series analysis on historical occupancy data, the system extracts periodic occupancy patterns, including weekday patterns, weekend patterns, holiday patterns, and special event patterns. These predictive patterns are used to compare with real-time occupancy data to identify behavioral deviations.

[0082] In practical applications, edge computing nodes fuse multi-source data from infrared sensors, millimeter-wave radar, WiFi probes, and access control card swipe records to generate occupancy pattern vectors for each region at a granular five-minute interval. Behavioral attribution analysis compares the actual occupancy pattern of the current time period with typical patterns of the same region on similar historical days at the same time, calculating the pattern difference. When the difference exceeds a preset threshold, it is determined to be an abnormal behavior. Further analysis is then conducted to determine whether the device's operating status matches the current occupancy pattern (e.g., an air conditioner still running at full power in "idle" mode), thereby quantifying the contribution of behavior to energy consumption deviations.

[0083] In summary, the “occupancy pattern” in this method has a clearly defined technical connotation, namely, the quantitative existence status and changing patterns of people in different areas of the building in terms of time, space and behavior type, which is the core input feature of behavioral attribution analysis.

[0084] First, device-level attribution was performed: the system retrieved the operating data of the air conditioning unit that ran on the weekend, calculated its actual COP (coefficient of performance), and found that although the value had decreased slightly, it was still within a reasonable range, ruling out the possibility of serious deterioration of the unit.

[0085] Next, system-level attribution was conducted: analysis of the air conditioning water system revealed that although the main unit was running, most of the terminal electric two-way valves were forcibly closed due to being in weekend mode, resulting in chilled water circulating at a low flow rate, low pump efficiency, and instances of bypass valves being opened by mistake, causing a systemic waste of "oversized power for a small load".

[0086] Then, behavioral attribution was performed: combining access control and sensor logs, it was found that some meeting rooms were booked for use on weekends, but the BA system still executed the "weekend low power mode" for the entire floor, failing to respond judiciously to this local demand.

[0087] Finally, environmental response-level attribution was conducted: the inspection revealed that the west corridor received a large amount of solar radiation on weekend afternoons due to the lack of artificial lighting and the failure of the curtains to close automatically, which led to a local temperature increase and indirectly increased the cooling load of the air conditioning in the adjacent area.

[0088] Through this four-level step-by-step analysis, the system accurately identified multiple root causes, such as "coarse system control strategy," "improper water system adjustment," and "lack of shading control," rather than simply attributing it to "power consumption of the air conditioning unit."

[0089] The multi-level energy consumption attribution analysis framework is not continuously running, but is triggered by the following three conditions: First, deviation-driven triggering. When the real-time energy consumption deviation matrix detects a persistent deviation (the deviation amplitude is stable and exceeds the threshold for more than four consecutive analysis cycles) or divergent anomaly (the deviation amplitude increases for three consecutive cycles) in a certain region or system, the system automatically initiates attribution analysis for the energy consumption entity associated with the anomaly.

[0090] Second, periodic health checks are triggered. Every day at dawn, the system performs a rapid attribution scan under low load conditions on major energy-consuming devices, subsystems, and typical functional areas to detect potential performance degradation or hidden faults.

[0091] Third, event triggering. When a building undergoes equipment maintenance, changes in control strategies, adjustments to usage functions, or drastic fluctuations in external weather conditions, the system proactively triggers a full-level attribution analysis of the affected area to assess the energy efficiency impact of the changes.

[0092] Upon triggering, the system executes attribution analysis sequentially in a fixed order: "device level → system level → behavior level → environmental response level." Only when the previous level cannot explain all the biases will the next level be executed. If the attribution at a certain level can explain more than 90% of the total bias, subsequent levels can be selectively executed to save computational resources.

[0093] The goal of equipment-level attribution is to identify whether the performance degradation of specific energy-consuming equipment (such as chillers, pumps, fans, air compressors, etc.) constitutes the main source of energy consumption deviations. Its automated implementation process is as follows: First, the system obtains the sequence of operating parameters of the target equipment during the deviation period from the edge nodes, including equipment power, current, frequency, inlet and outlet temperatures, pressure, flow rate, operating time, and load rate. At the same time, the system extracts rated parameters from the equipment nameplate or initial commissioning records, and extracts the performance benchmarks of the equipment under similar operating conditions (such as typical energy efficiency ratio and cooling capacity per unit power) from the database of historical normal operating periods.

[0094] Secondly, the system calculates the actual energy efficiency index of the equipment under the current operating conditions. For example, for chillers, it calculates the ratio of actual cooling capacity to input power; for water pumps, it calculates the ratio of actual flow rate to power; and for fans, it calculates the ratio of actual air volume to power. Then, it compares the actual value with the rated value under the current operating conditions and the historical benchmark value, and calculates the percentage of relative deviation.

[0095] Secondly, the system employs statistical process control methods to determine whether the deviation is caused by the equipment itself. Specifically, it constructs a normal distribution model using the equipment's energy efficiency ratio (EER) data under the same load rate and outdoor temperature conditions over the past ninety days, and calculates the degree of deviation of the current measured EER from this distribution. When the deviation exceeds three times the standard deviation, it is determined to be an equipment-level anomaly. Simultaneously, the system analyzes the time-series trends of the equipment's operating parameters, using the moving average method to detect whether there are monotonic changes with continuous decreases or increases, thereby distinguishing between sudden failures and slow aging.

[0096] Finally, the device-level attribution outputs two quantitative results: first, the percentage contribution of device performance degradation to the total energy consumption deviation (equal to the actual energy efficiency ratio deviation rate multiplied by the proportion of device energy consumption to total energy consumption); and second, specific anomaly type codes, such as "clogged condenser in chiller leading to decreased heat exchange efficiency," "pump impeller wear causing flow-power curve shift," and "increased fan bearing friction leading to increased power consumption." These anomaly types are automatically identified by pattern matching between actual parameter patterns and a pre-stored fault feature library.

[0097] When device-level attribution cannot explain all the biases, the system moves to system-level attribution. This level focuses on the coordination efficiency within or between subsystems composed of multiple devices (such as chilled water systems, cooling water systems, and air systems).

[0098] For HVAC systems, the automated implementation of system-level attribution includes the following steps: The first step is for the system to collect synchronous operating parameters of all related equipment within the subsystem. Taking the chilled water system as an example, it is necessary to simultaneously acquire the evaporator inlet and outlet temperatures of the chiller unit, the frequency and flow rate of the chilled water pump, the opening status of each terminal valve, and the readings of the differential pressure sensor.

[0099] The second step involves calculating key indicators of subsystem coordination efficiency. For example, calculating the "delivery coefficient" of the chilled water system, which is the ratio of the actual delivered cooling capacity to the pump input power; calculating the "water system balance index," which is the ratio of the most unfavorable terminal pressure difference to the design pressure difference, and the dispersion of valve openings in each branch. When the balance index is below a threshold or the valve opening dispersion is too large, it indicates a hydraulic imbalance.

[0100] The third step involves the system detecting control logic conflicts. This is achieved by analyzing the sequence of commands issued by different controllers to determine if any contradictions exist. For example, the system might simultaneously receive commands to "increase the chilled water supply temperature setpoint" and "decrease the chilled water supply temperature setpoint," or a positive feedback loop might form between the cooling tower fan start / stop control and the cooling water bypass valve adjustment. The system identifies such conflicts by comparing the command timestamps with the actuator feedback status.

[0101] The fourth step is for the system to output system-level attribution results: the percentage contribution of system coordination efficiency loss to total energy consumption deviation (calculated by comparing the actual coordination index with the design optimal index, and then converting it into equivalent energy consumption), as well as specific system anomaly type codes, such as "severe imbalance of chilled water system branches leading to increased water pump energy consumption", "mismatch between chiller unit and cooling tower fan start-stop logic causing excessive cooling water temperature fluctuations", "failure of fresh air and return air valve linkage logic leading to over-cooling", etc.

[0102] Behavioral attribution analysis assesses the impact of user behavior (such as unexpected device start-up and shutdown, and abnormal space occupancy patterns) on energy consumption deviations. This level is entirely automated and relies on sensor data, requiring no manual observation.

[0103] The specific implementation steps are as follows: First, the system acquires personnel presence data for each functional area from space occupancy sensors, access control systems, WiFi probes, or infrared counters, generating occupancy rate time series in five-minute increments. Simultaneously, it obtains actual operating status sequences from equipment loops affected by user behavior, such as lighting, air conditioning terminals, and office equipment.

[0104] Secondly, the system compares the actual operating status of the equipment with the expected operating status based on occupancy rate. The rule for generating the expected status is: in normal mode, the equipment should start within two minutes after personnel enter the area and shut down or enter energy-saving mode within fifteen minutes after personnel leave. The system automatically identifies behavioral deviation patterns such as "continuous operation of equipment during unattended periods," "failure to start equipment in a timely manner during occupied periods," and "abnormal equipment start-stop frequency (too frequent or inactive for extended periods)" through a sliding window matching algorithm.

[0105] Secondly, the system performs cluster analysis on equipment operation data during non-working hours (such as nighttime, weekends, and holidays). Using unsupervised algorithms such as DBSCAN or K-means, historical equipment operation patterns during non-working hours are divided into several typical clusters, such as the "completely off cluster," the "minor overtime lighting cluster," and the "abnormally fully on cluster." If the current time period's pattern is classified into the "abnormally fully on cluster," it is considered a behavioral-level deviation.

[0106] Finally, the system quantifies the contribution of behavioral-level deviations to energy consumption. The calculation method is as follows: subtract the theoretical operating time (based on the average operating time of similar historical occupancy periods) from the actual equipment operating time, multiply by the equipment's rated power to obtain the excess energy consumption attributed to the behavior, and then divide by the total energy consumption deviation to obtain the percentage contribution of the behavior. Output exception type codes include: "Lights left on after get off work," "Air conditioning running throughout the entire area on weekends, not just in the overtime area," and "Meeting room empty but equipment running continuously at full power," etc.

[0107] Environmental response attribution is used to assess the deviation between the performance of passive design elements of a building (envelope, shading, natural ventilation, etc.) and design expectations under actual climatic conditions.

[0108] The automation implementation process is as follows: The first step is for the system to extract the design parameters of the building envelope (wall heat transfer coefficient, window solar heat gain coefficient, and geometric parameters of the shading device) from the BIM model, and to obtain the outdoor climate data (temperature, solar radiation, wind speed and direction) from the weather station.

[0109] The second step involves establishing a simplified physical model of the building's passive performance. This model does not perform a full energy consumption simulation; instead, it calculates only the response amplitude and delay time of indoor temperature relative to outdoor temperature and solar radiation under conditions without active heating or cooling. For example, the system uses the nighttime rate of indoor temperature decrease to inversely calculate the thermal inertia parameters of the actual building envelope.

[0110] The third step involves comparing the measured passive response parameters with the design values. Specific comparison items include: the difference between the actual maximum indoor temperature and the design prediction value under the same outdoor conditions; the ratio of the actual number of air changes to the design natural ventilation potential during the natural ventilation period; and the deviation between the measured radiation intensity transmitted through the glass and the theoretical value of the design shading coefficient after the shading device is activated.

[0111] The fourth step involves the system automatically identifying possible causes of the deviation. For example, if the actual indoor temperature is significantly higher than the design prediction value in the summer afternoon, and an abnormally high glass surface temperature is detected at the same time, the system determines it as "deterioration of glass shading performance" or "external shading device not activated." If the indoor temperature drops slowly at night, it is determined as "the actual thermal insulation performance of the building envelope is lower than the design standard." If the indoor and outdoor temperature difference is too small during natural ventilation, it is determined as "the ventilation opening area or location does not conform to the design."

[0112] Step 5, Environmental Response Attribution Output: The percentage contribution of passive performance deviation to total energy consumption deviation (calculated by calculating the additional active cooling or heating energy consumption required to maintain indoor comfort), and specific anomaly type codes, such as "the actual value of the shading coefficient of the west-facing glass curtain wall is 30% higher than the design value", "the heat transfer coefficient increases due to moisture in the roof insulation layer", and "the actuator of the natural ventilation window fails to open according to the preset strategy".

[0113] Through the automated implementation of the above-mentioned attribution methods at each level, the system can locate the root cause of energy consumption deviations step by step without relying on human judgment, and output quantified contribution and standardized anomaly type codes, providing accurate decision-making basis for subsequent edge-side strategy generation.

[0114] Furthermore, the dynamic management and control logic framework includes an adaptive baseline update mechanism and a climate adaptation strategy generation mechanism; The adaptive benchmark update mechanism periodically calibrates the benchmark energy consumption dynamic curve library based on long-term actual operation data, uses machine learning methods to establish an energy consumption prediction model that reflects the personalized operation characteristics of the building, and updates the benchmark reference value accordingly. The climate adaptability strategy generation mechanism predicts the future load trend of buildings based on short-term weather forecasts and historical similar day analysis, and accordingly presets or dynamically adjusts the operating parameters of the HVAC system at the edge.

[0115] In this embodiment, the adaptive baseline update mechanism runs once a quarter. It collects all actual operating data from the past quarter and uses algorithms such as gradient boosting trees to train a predictive model with "date type, time, outdoor temperature and humidity, and personnel density" as features and "total building energy consumption" as the objective. The machine learning method used by the adaptive baseline update mechanism is a gradient boosting decision tree model based on ensemble learning, specifically using the LightGBM or XGBoost algorithm framework.

[0116] The model learns the building's real "personality," which may differ from the original design. After training, the energy consumption values ​​predicted by the new model will serve as an important component of the "personalized baseline curve" for the next quarter, replacing some outdated theoretical design values ​​and making the "normal" standards more in line with the current situation.

[0117] The climate adaptation strategy generation mechanism runs daily. Every morning, the edge server obtains the weather forecast for the day and the next day, and searches for "historically similar days" with similar meteorological conditions in the historical database.

[0118] By combining personalized baseline curves with actual load curves from similar days, the system predicts the hourly building cooling and heating load for the next 24 hours. For example, if the forecast shows that the load peak will occur between 14:00 and 16:00 today, the system can appropriately lower the temperature setpoint of public areas around 10:00 in the morning (e.g., by 0.5°C within the comfort range) to "pre-cool" the building by utilizing its thermal inertia. This reduces the start-up and shutdown time or high-load operation time of the air conditioning unit during peak hours, smooths the load curve, and achieves peak shifting and valley filling.

[0119] The "dynamic management and control logic framework" refers to a set of configurable, collaborative, and adaptive logical rules and decision-making scheduling modules embedded in edge computing nodes and cloud collaboration platforms. This framework is not a single algorithm or fixed process, but rather a top-level logical architecture used to organize and coordinate the operation of various energy consumption management and control mechanisms. Specifically, the framework includes at least the following five core mechanisms: an adaptive baseline update mechanism, a climate adaptation strategy generation mechanism, a population dynamic response mechanism, a cross-validation diagnostic mechanism, and a life-cycle stage adaptive adjustment mechanism.

[0120] The framework's functional positioning is as follows: First, as the scheduling hub for data and control flows, it is responsible for receiving output results from the deviation matrix, correlation analysis engine, and attribution analysis framework, and determining which mechanism(s) to invoke in response based on preset priorities and triggering conditions. Second, as a coordination layer between various mechanisms, when multiple mechanisms are triggered simultaneously and may generate conflicting control commands (e.g., a climate adaptation strategy requiring an increase in temperature setpoints while a population dynamic response mechanism requires a decrease), the framework arbitrates according to its built-in conflict resolution rules to determine the weight or execution order of each mechanism's command. Third, as an environment container for strategy generation, it provides multi-objective constraints (such as energy efficiency priority, equipment lifespan priority, or comfort priority) and parameter configuration interfaces for each mechanism in the synthesis of control command sets.

[0121] In the method execution flow, the framework's working logic is as follows: After the real-time energy consumption deviation matrix is ​​generated, the influencing factor correlation analysis engine and the attribution analysis framework output the external influence weights and deviation root cause information, respectively. This information is sent to the dynamic control logic framework. The framework first identifies the main type and severity level of the current deviation, and then consults the internally preset rule mapping table to determine the mechanism that should be prioritized. For example, when the attribution conclusion points to equipment aging and the life cycle stage is determined to be "equipment aging stage", the framework prioritizes calling the performance degradation compensation algorithm in the life cycle stage adaptive adjustment mechanism; when the deviation matrix shows that there is a predicted load peak in the future period, the framework prioritizes calling the climate adaptability strategy generation mechanism for pre-cooling or pre-heating planning. The preliminary control suggestions produced by each mechanism are returned to the framework, which integrates and resolves conflicts to form the final unified control instruction set, which is then issued to the field actuators.

[0122] The framework's configuration features dynamic evolution. The rule mapping table, conflict resolution priorities, and activation thresholds for each mechanism stored internally can all be adjusted based on periodic updates from the cloud platform. Simultaneously, the framework records the context and execution effect of each strategy decision, gradually optimizing its scheduling strategy for various mechanisms through closed-loop feedback. For example, if historical records show that prioritizing the population dynamic response mechanism over climate adaptation strategies achieves better energy-saving results under a certain deviation mode, the framework will automatically adjust the mechanism invocation order under that mode using reinforcement learning.

[0123] In summary, the "dynamic management and control logic framework" is the core logical entity in this invention, distinguishing it from traditional fixed-rule control. Specifically, it refers to a comprehensive decision-making and scheduling system embedded in the edge-cloud collaborative system, used to organize, schedule, coordinate, and execute multiple mechanisms, including adaptive baseline updates, climate adaptation strategy generation, population dynamic response, cross-validation diagnosis, and lifecycle stage adaptive adjustments. This framework receives deviation matrices, correlation analysis results, and attribution conclusions, and dynamically generates unified real-time management and control strategies according to preset rule mapping and conflict resolution logic, supporting self-optimization based on execution performance.

[0124] Furthermore, the dynamic management and control logic framework also includes a population dynamic response mechanism, a cross-validation diagnostic mechanism, and a life cycle stage adaptive adjustment mechanism; The population dynamic response mechanism: Based on real-time monitoring of population distribution density and flow trends, dynamically adjusts environmental control targets and the working status of energy-consuming equipment in relevant areas to achieve on-demand supply; The cross-validation diagnostic mechanism employs a simulation model based on physical mechanisms and a statistical model based on operational data in parallel to calculate and predict energy consumption. By comparing the consistency of the two results, sensor failures, model inaccuracies, or hidden performance issues are diagnosed. The life cycle stage adaptive adjustment mechanism automatically identifies the life cycle stage of the building based on its years of operation and records of major equipment upgrades and renovations. Different benchmark curves are used at different stages to compare the stringency of the comparison, the aggressiveness of the control strategy, and the compensation algorithm for equipment performance degradation, so that the management logic matches the actual state of the building.

[0125] In this embodiment, in a large airport terminal, the population dynamic response mechanism dynamically manages the lighting and air conditioning of different concourses based on real-time data provided by the video passenger flow analysis system.

[0126] When a flight delay causes a large number of passengers to gather in the waiting area, the system automatically increases the fresh air volume and cooling capacity of the area; when passenger flow drops sharply late at night, it turns off the lighting and air conditioning in most non-main passageways.

[0127] The cross-validation diagnostic mechanism operates regularly: a simplified white-box model based on building physics parameters (such as calculating energy consumption using building envelope parameters and equipment rated power) and a black-box statistical model based on historical data (such as a time series prediction model) simultaneously predict energy consumption for the next period.

[0128] Under normal circumstances, the results of the two models should be consistent. If a sudden increase in the predicted value of the black box model is found while the predicted value of the white box model remains unchanged, it is discovered upon investigation that a signal transmission failure of a certain electricity meter is causing the data anomaly. Conversely, if the white box model continues to overestimate energy consumption due to outdated equipment nameplate parameters (such as a decrease in the actual efficiency of the water pump), the system will trigger a warning of "equipment performance degradation" and recommend updating the model parameters.

[0129] The life cycle stage adaptive adjustment mechanism manages strategies from a macro perspective: for newly commissioned "youthful" buildings, the control strategy is relatively robust, with strict benchmark comparisons to identify design or construction legacy issues; for "middle-aged" buildings that have been in operation for more than ten years, the system automatically relaxes the temperature fluctuation range of non-critical areas and introduces an efficiency compensation coefficient for aging refrigeration units in the control algorithm, allowing them to operate at slightly higher set temperatures to protect the equipment. At the same time, the benchmark curve is updated to reflect the reasonable decline in equipment energy efficiency, making energy-saving targets more scientific and achievable.

[0130] Furthermore, the dynamic generation and execution of targeted real-time control strategies specifically includes: Based on the aforementioned deviation matrix, correlation analysis results, and attribution conclusions, a specific set of control instructions is synthesized locally at the edge computing node. The strategy generation takes into account multiple objectives, including real-time energy efficiency optimization, equipment lifespan, and indoor environmental quality. After logical verification, the generated strategy is directly distributed to the field controller or actuator through the edge node to form a low-latency closed-loop control. At the same time, the strategy summary and execution effect are synchronized to the cloud platform for long-term analysis and framework optimization.

[0131] In this embodiment, once the edge server completes the analysis of the current "high energy consumption of lighting in the parking lot area at night", it immediately starts policy synthesis locally.

[0132] The attribution results indicate that the problem was primarily caused by malfunctions in some areas' "always-on" lights and sensors. Within milliseconds, the strategy generation module integrated three constraints—"turning off unnecessary lighting to save energy" (energy efficiency target), "avoiding frequent switching to extend lamp life" (equipment life target), and "ensuring safe illuminance in main passageways" (environmental quality target)—to generate a specific set of instructions. For areas with sensor malfunctions, change their lighting mode from "automatic sensing" to "timed off"; For normal areas, increase the sensor sensitivity by one level; Generate a repair work order.

[0133] These instructions are first logically validated by the local "rule engine" to prevent contradictory instructions such as "sending both light on and light off simultaneously". After the validation is passed, the instructions are directly sent to the corresponding lighting controllers for execution via the BACnet / IP protocol. The entire closed loop from analysis to execution is completed within seconds. At the same time, the summary of this event (problem, attribution, instructions, expected energy saving) and the actual energy consumption change data after execution are compressed, packaged, and asynchronously uploaded to the cloud historical database to enrich the case library and train a more accurate attribution model.

[0134] Furthermore, the specific synthesis of the control instruction set locally on the edge computing node specifically includes generating the following control instructions: Operating parameter adjustment instructions: Based on climate adaptability strategies and real-time environmental feedback, dynamically adjust the temperature setpoint, chilled water supply temperature, damper opening, and fresh air ratio of the HVAC system; Equipment mode switching command: Based on the behavioral attribution analysis results and the dynamic response mechanism of personnel, control the dimming mode, zone start / stop of the lighting system, and the intermittent operation or sleep mode of the power equipment. System coordinated control commands: Based on system-level attribution conclusions, generate and issue linkage adjustment commands to eliminate control conflicts between systems and optimize hydraulic or air volume balance; Abnormal operating condition handling instructions: In response to abnormal sensor data or hidden performance degradation of equipment identified through cross-validation diagnostic mechanism, trigger equipment calibration, backup sensor switching or degraded operation contingency plan.

[0135] In this embodiment, during a specific midday scenario, the edge nodes synthesized the following hybrid instruction set: Operating parameter adjustment command: As the weather forecast shows that the cloud cover will increase and the solar radiation will weaken in the afternoon, the system will automatically increase the air conditioning temperature setpoint of the west wing of the building facing the curtain wall from 24℃ to 25℃. At the same time, based on the real-time return air CO2 concentration, the fresh air valve opening will be dynamically adjusted from 40% to 30%.

[0136] Device mode switching command: Based on the data from the personnel sensor, if it is detected that the meeting in the large lecture hall has ended and all personnel have left, the system will switch the lighting in the area from "lecture mode" to "all off mode" and switch the air conditioner from "cooling mode" to "ventilation + humidity control" sleep mode.

[0137] System coordinated control command: The system-level attribution found that the frequency of the chilled water pump and the opening of the terminal valve were mismatched, resulting in an excessive pressure difference. The system generated a command to reduce the frequency of the primary pump by 5Hz while ensuring the flow rate at the most unfavorable terminal, and simultaneously fine-tune the balancing valves of each branch to bring the system hydraulics back to the efficient balance point.

[0138] Abnormal operating condition handling instructions: Cross-validation mechanism alarm: The reading of a temperature sensor in a certain zone deviates from the statistical values ​​of multiple neighboring sensors by more than 2°C and is judged as "suspected drift".

[0139] The system immediately triggered a response command: The faulty sensor is temporarily disabled in the control logic, and a virtual value calculated based on nearby sensors is used instead. Generate a high-priority calibration work order in the management interface; If the environmental control requirements for the area are extremely high, the backup wireless temperature and humidity sensor will be activated and connected to the system.

[0140] All these instructions are generated and executed quickly locally, ensuring precise, timely, and autonomous control over complex building energy systems.

[0141] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing, characterized in that: Includes the following steps: Based on historical and design data from each stage of the building's entire life cycle, a multi-dimensional benchmark energy consumption dynamic curve library is constructed, and a dynamic control logic framework with multiple mechanisms is established. During the building operation phase, energy consumption and environmental data of various components within the building are collected in real time through deployed edge computing nodes. The collected energy consumption and environmental data of the building's internal components are compared with the theoretical values ​​under the corresponding conditions in the benchmark energy consumption dynamic curve library to generate a real-time energy consumption deviation matrix. Construct an engine for analyzing the correlation of influencing factors, access and analyze multidimensional external data on meteorology and population in real time, quantify their dynamic correlation with energy consumption deviation, and drive the adjustment of edge-side control strategies; Establish a multi-level energy consumption attribution analysis framework to decompose the total energy consumption deviation into equipment level, system level, behavior level and environmental response level, and identify the root causes of the deviation; Based on the above comparison, analysis and attribution results, and by integrating the various mechanisms in the dynamic control logic framework, targeted real-time control strategies are dynamically generated and executed. The generation of the real-time energy consumption deviation matrix specifically includes: Calculate the absolute and relative percentage deviations between the actual energy consumption value and the corresponding theoretical value in the benchmark curve library, using a preset time granularity. The deviation value is spatiotemporally correlated and mapped with the collected internal environmental parameters and external influencing factors to form a multidimensional matrix containing the deviation value, the occurrence time, and the related factors. Based on the persistence, fluctuation range, and scope of impact of the deviation, the deviation type is automatically classified and labeled.

2. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The construction of the multi-dimensional benchmark energy consumption dynamic curve library specifically includes: Based on the design parameters in the building information model, the theoretical baseline energy consumption curves of the building are calculated under different seasons, different time periods and different usage rates. Integrate historical meteorological data, typical year data, and operational data of similar buildings in the building's location to form a comprehensive reference database for comparison; The dynamic curve library supports adaptive iterative updates based on actual building operation feedback and lifecycle stages.

3. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The real-time collection of energy consumption and environmental data within the building through deployed edge computing nodes specifically includes: Edge computing nodes deployed at various levels in the building directly collect sub-metering data on lighting, air conditioning, and power within their respective areas, as well as environmental data on temperature, humidity, illuminance, and the presence of people. The raw data collected at the edge is preprocessed locally, including data cleaning, time alignment and format standardization, to form a structured local data set.

4. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The construction of the influencing factor correlation analysis engine specifically includes: Establish real-time interfaces with multi-source external data services, including meteorology, demographics, calendar events, and energy prices, to acquire structured external data; Using time series-based association rule mining and regression analysis, the correlation coefficients and influence weights of each external variable with the total building energy consumption and the deviations of each component energy consumption are dynamically quantified. Based on the analysis results, a dynamic allocation model for the weights of influencing factors is generated, providing a quantitative basis for adjusting the marginal side strategy.

5. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The establishment of the multi-level energy consumption attribution analysis framework specifically includes: Equipment-level attribution: Analyze the deviation between the actual operating energy efficiency ratio and the rated or designed energy efficiency ratio of key energy-consuming equipment, and correlate it with equipment operating time, load rate and maintenance records; System-level attribution: Analyze the collaborative operating efficiency within and between HVAC and lighting subsystems, and identify system performance losses caused by control logic conflicts and hydraulic imbalances; Behavioral attribution: Based on spatial usage sensor data, analyze the impact of building occupancy patterns and user behavior of unintended device activation on energy consumption patterns; Environmental response attribution: Analyze the performance of passive design of building envelope, shading and natural ventilation under actual climatic conditions and assess its deviation from design expectations.

6. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The dynamic management and control logic framework includes an adaptive baseline update mechanism and a climate adaptation strategy generation mechanism; The adaptive benchmark update mechanism periodically calibrates the benchmark energy consumption dynamic curve library based on long-term actual operation data, uses machine learning methods to establish an energy consumption prediction model that reflects the personalized operation characteristics of the building, and updates the benchmark reference value accordingly. The climate adaptability strategy generation mechanism predicts the future load trend of buildings based on short-term weather forecasts and historical similar day analysis, and accordingly presets or dynamically adjusts the operating parameters of the HVAC system at the edge.

7. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The dynamic management and control logic framework also includes a population dynamic response mechanism, a cross-validation diagnostic mechanism, and a life cycle stage adaptive adjustment mechanism; The population dynamic response mechanism: Based on real-time monitoring of population distribution density and flow trends, dynamically adjusts environmental control targets and the working status of energy-consuming equipment in relevant areas to achieve on-demand supply; The cross-validation diagnostic mechanism employs a simulation model based on physical mechanisms and a statistical model based on operational data in parallel to calculate and predict energy consumption. By comparing the consistency of the two results, sensor failures, model inaccuracies, or hidden performance issues are diagnosed. The life cycle stage adaptive adjustment mechanism automatically identifies the life cycle stage of the building based on its years of operation and records of major equipment upgrades and renovations. Different benchmark curves are used at different stages to compare the stringency of the comparison, the aggressiveness of the control strategy, and the compensation algorithm for equipment performance degradation, so that the management logic matches the actual state of the building.

8. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 1, characterized in that, The dynamic generation and execution of targeted real-time control strategies specifically include: Based on the aforementioned deviation matrix, correlation analysis results, and attribution conclusions, a specific set of control instructions is synthesized locally at the edge computing node. The strategy generation takes into account multiple objectives, including real-time energy efficiency optimization, equipment lifespan, and indoor environmental quality. After logical verification, the generated strategy is directly distributed to the field controller or actuator through the edge node to form a low-latency closed-loop control. At the same time, the strategy summary and execution effect are synchronized to the cloud platform for long-term analysis and framework optimization.

9. The method for real-time management and control of building energy consumption throughout its entire lifecycle based on edge computing as described in claim 8, characterized in that, The specific synthesis of the control instruction set locally at the edge computing node includes generating the following control instructions: Operating parameter adjustment instructions: Based on climate adaptability strategies and real-time environmental feedback, dynamically adjust the temperature setpoint, chilled water supply temperature, damper opening, and fresh air ratio of the HVAC system; Equipment mode switching command: Based on the behavioral attribution analysis results and the dynamic response mechanism of personnel, control the dimming mode, zone start / stop of the lighting system, and the intermittent operation or sleep mode of the power equipment. System coordinated control commands: Based on system-level attribution conclusions, generate and issue linkage adjustment commands to eliminate control conflicts between systems and optimize hydraulic or air volume balance; Abnormal operating condition handling instructions: In response to abnormal sensor data or hidden performance degradation of equipment identified through cross-validation diagnostic mechanism, trigger equipment calibration, backup sensor switching or degraded operation contingency plan.

Images