A method and system for optimizing and controlling building energy consumption by sub-item metering

Abstract

This invention discloses a method and system for optimizing and controlling building energy consumption by sub-item metering, relating to the field of building energy consumption management technology. The method includes: collecting multi-source data through a sensor network; obtaining energy consumption sub-sequences through event-level segmentation based on power transition events; determining the energy-consuming unit category and confidence level through an interpretable classification model; obtaining a stable sub-item energy consumption sequence after non-stationarity correction; constructing a sequence prediction model embedded with temperature conditions; performing scale calibration and temporal consistency correction in conjunction with building characteristics to obtain a calibrated sub-item energy consumption prediction model; triggering a multi-strategy control process when the prediction deviation exceeds a threshold, dynamically selecting the optimal control strategy and generating control instructions; and updating the prediction model through transfer learning based on virtual-real consistency feedback. This invention achieves closed-loop optimization from fine sensing and accurate prediction to intelligent control, effectively improving the accuracy of building energy consumption metering and the level of intelligent control.

Description

Technical Field

[0001] This invention relates to the field of building energy consumption management technology, and in particular to a method and system for optimizing and controlling building energy consumption by sub-item metering. Background Technology

[0002] With the acceleration of urbanization and the continuous expansion of building scale, building energy consumption has become a significant component of global energy consumption. Its refined management and optimized control are crucial for achieving the "dual carbon" goal. However, at the data perception level, existing systems largely rely on traditional total metering or coarse-grained itemized metering, making it difficult to accurately capture the independent energy consumption characteristics and dynamic changes of different terminal energy-consuming units within buildings, such as air conditioning, lighting, and office equipment. This weak data foundation leads to subsequent energy consumption analysis and control strategies often remaining at a macro level, unable to support precise management of specific equipment or areas. At the energy consumption prediction level, existing prediction models generally do not adequately consider the response mechanisms of building physical characteristics and external environmental factors, especially temperature changes. Most models treat temperature as a simple linear input, ignoring the complex nonlinear coupling relationship between temperature and building envelope characteristics and internal loads. This results in a significant decrease in prediction accuracy under extreme climate or transitional season conditions, failing to provide a reliable basis for control strategy formulation. At the control execution level, existing systems mostly employ preset, fixed threshold triggering single control strategies (such as simple start-stop control or PID regulation), lacking multi-objective collaborative optimization capabilities. This control method is difficult to dynamically balance multiple objectives such as energy saving, equipment life and personnel comfort. Moreover, when faced with dynamic fluctuations in internal and external building conditions (such as personnel flow, equipment start-up and shutdown, and weather changes), it exhibits poor adaptability and unstable control effects. In fact, improper control may even lead to equipment conflicts or deterioration of comfort.

[0003] Therefore, a method for metering and optimizing building energy consumption is needed to overcome the above limitations, so as to achieve a technological leap from extensive management to refined perception, from single control to intelligent decision-making, and from static strategy to dynamic adaptation. Summary of the Invention

[0004] To address the problems of inefficient perception, distorted prediction, and limited control in existing building energy consumption management processes, this invention proposes a method and system for optimizing and controlling building energy consumption through sub-item metering.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] A method for optimizing and controlling building energy consumption through sub-item metering, the method comprising:

[0007] The system collects fine-grained energy consumption data and occupant activity data from energy-consuming units within the building through a sensor network, and simultaneously acquires outdoor temperature data and building characteristic data.

[0008] Based on the fine-grained energy consumption data and personnel activity data, event-level segmentation processing is performed on the power transition events of the energy consumption signal to obtain energy consumption sub-sequences; an interpretable energy consumption classification model is used to determine the energy-consuming unit category and category confidence of the energy consumption sub-sequences, and non-stationarity correction processing is performed on the energy consumption sub-sequences to obtain stable component energy consumption sequences.

[0009] The sub-item energy consumption sequence is jointly represented with the outdoor temperature data of the corresponding time period to construct a sequence prediction model based on temperature condition embedding. According to the building area and number of floors in the building feature data, the sequence prediction model is scaled and time series consistency corrected according to the error-cost joint optimization criterion to obtain a calibrated sub-item energy consumption prediction model. The calibrated sub-item energy consumption prediction model is used to predict sub-item energy consumption.

[0010] When the deviation between the predicted energy consumption of each item and the target energy consumption reference value exceeds the control threshold, the energy consumption regulation process is triggered: a set of control strategies including rule control, model prediction control and reinforcement learning control strategies are constructed, the occupancy status and uncertainty are inferred by combining the personnel activity data, the control threshold is dynamically adjusted and the optimal control strategy is selected, and the corresponding energy consumption regulation command is generated and executed.

[0011] Collect operational status data after execution and provide feedback on virtual-real consistency. Based on the feedback results, perform transfer learning to update the sub-item energy consumption prediction model and output building energy consumption optimization results.

[0012] As a preferred embodiment of the present invention, the fine-grained energy consumption data includes voltage, current, active power, reactive power, power factor, harmonic components, and equipment switching status of each energy-consuming unit in the building, collected by the energy consumption acquisition module; the energy consumption acquisition module performs timestamp alignment through a unified clock synchronization mechanism and records sampling quality labels in the sampling data frames to reflect sampling packet loss rate, noise level, and equipment calibration status;

[0013] The personnel activity data includes the number of people indoors, personnel density, personnel activity intensity, personnel stay time, personnel entry and exit flow and occupancy status data; the building characteristic data includes building area, number of floors, floor distribution, thermal parameters of building envelope, orientation information, functional zoning identifiers and spatial coordinates of energy consumption acquisition modules.

[0014] As a preferred embodiment of the present invention, the method for obtaining the energy-consuming subsequence specifically includes:

[0015] The fine-grained energy consumption data is preprocessed to suppress noise, and the power change rate of the fine-grained energy consumption data is calculated; an adaptive dual threshold mechanism is adopted to dynamically set the judgment threshold of the power change rate based on local signal characteristics, and when the power change rate exceeds the judgment threshold, a power transition event is marked.

[0016] The sharpness of the time-domain change and the frequency-domain energy distribution change of the power transition event are jointly analyzed. When the joint analysis results meet the set conditions, the start and end boundaries of the power transition event are confirmed, and the initial energy consumption subsequence is segmented from the fine-grained energy consumption data based on the start and end boundaries.

[0017] The initial energy consumption subsequence is subjected to time-series correlation analysis with the synchronously collected personnel activity data. By calculating the spatiotemporal overlap between the intensity of personnel activity and the initial energy consumption subsequence during the duration of the power transition event, false events caused by measurement noise or non-personnel factors are eliminated, and effective energy consumption subsequences are selected.

[0018] The effective energy consumption subsequence is subjected to boundary optimization and segment fusion processing. The segment fusion is based on the power pattern similarity and temporal proximity of the effective energy consumption subsequence, and merges multiple effective energy consumption subsequences that represent the continuous operating state of the same energy-consuming device, and outputs the final energy consumption subsequence after denoising and fusion.

[0019] As a preferred embodiment of the present invention, the method for obtaining the stabilized component energy consumption sequence specifically includes:

[0020] The time-domain statistical features, frequency-domain energy features, and power shape contour features of the energy consumption subsequence are extracted to construct a high-dimensional feature vector. Based on the high-dimensional feature vector, an interpretable energy consumption classification model is used to classify the energy-consuming units. The energy consumption classification model is specifically a gradient boosting decision tree model, which outputs the category prediction results of the energy-consuming units and the corresponding category confidence.

[0021] During the inference process of the gradient boosting decision tree model, concept drift detection is performed based on the model prediction result data stream. When the detected distribution change exceeds a set threshold, the incremental learning process of the gradient boosting decision tree model is triggered, and the gradient boosting decision tree model is updated online using newly arrived high-confidence samples.

[0022] Based on the comparison results between the category confidence and the preset confidence threshold, a similarity retrieval based on dynamic time warping is initiated for energy consumption subsequences with low confidence. The most similar template sequence is matched from the pre-constructed historical energy consumption event label library to correct the category label and update the category confidence.

[0023] Based on the energy consumption subsequences after classification and confidence updates, they are summarized according to the energy-consuming unit category to generate a preliminary sub-item energy consumption sequence. A directed graph of equipment interlocking is constructed to represent the causal relationship between different energy-consuming unit categories. Logical consistency correction is performed on the preliminary sub-item energy consumption sequence according to the directed graph of equipment interlocking to eliminate redundant calculations or conflict allocations caused by equipment linkage control.

[0024] The non-stationarity correction process is applied to the logically corrected energy consumption sequence to obtain a stable energy consumption sequence.

[0025] As a preferred embodiment of the present invention, the construction of the sequence prediction model based on temperature condition embedding includes:

[0026] Multi-time-scale feature extraction is performed on the stabilized energy consumption sequence to obtain energy consumption pattern feature vectors with daily cycle, weekly cycle and seasonal trend.

[0027] For the outdoor temperature data, a temporal convolutional network is used to extract the deep features of the historical data sequence, and the temperature change rate and cumulative temperature load index are calculated in parallel to form a multi-scale temperature temporal feature vector.

[0028] The energy consumption pattern feature vector is fused with the multi-scale temperature time-series feature vector through cross-attention to generate a temperature-condition-aware energy consumption representation. The expression is:

[0029] ;

[0030] In the formula, This is a query matrix for energy consumption characteristics. This is the temperature feature key value matrix. Scaling factor For activation function, For the thermal parameters of the building envelope, For building area, and To modulate network parameters;

[0031] Energy consumption characterization based on the aforementioned temperature conditions A sequence prediction model is constructed with dilated causal convolution as the encoder and gated recurrent unit as the decoder. At each time step of the decoder, the temperature condition context vector is received by mapping the multi-scale temperature time-series feature vector through a gated fusion mechanism.

[0032] The sequence prediction model employs a phased strategy during the training phase, including encoder-decoder pre-training under fixed temperature scenarios, end-to-end fine-tuning with the introduction of temperature-conditional context vectors, and an adversarial training strategy under extreme temperature scenarios.

[0033] As a preferred embodiment of the present invention, the obtained calibrated sub-item energy consumption prediction model specifically includes:

[0034] The sequence prediction model is treated as a differentiable whole, and a scale calibration network is connected after the output layer. The scale calibration network uses the building area from the building feature data. With the number of floors as a conditional input, a set of calibration parameters is dynamically generated. The calibration parameters include at least a scaling factor matrix and a bias vector.

[0035] Minimize the following joint loss function using the error-cost joint optimization criterion. To jointly optimize the calibration parameters of the sequence prediction model and the scale calibration network:

[0036] ;

[0037] In the formula, This is the mean square error loss term between the predicted energy consumption value and the actual energy consumption value; The equipment control cost penalty is calculated based on the deviation between the predicted energy consumption and the target energy consumption, as well as the frequency of changes in the equipment's start-up and shutdown status. This is the physical constraint loss term based on the building heat balance equation; This is a time-series consistency loss term; These are hyperparameters used to balance the weights of the various loss terms;

[0038] Among them, the physical constraint loss term The calculation formula is:

[0039] ;

[0040] Temporal consistency loss term The calculation formula is:

[0041] ;

[0042] In the formula, Indicates at time The sum of the predicted values ​​of each component energy consumption, Indicates the overall heat transfer coefficient of the building envelope. Indicates time The temperature difference between indoors and outdoors To predict the time step; Indicates time Internal heat gain related to personnel activity data; This indicates the total number of energy-consuming unit categories; The model represents the first Each energy-consuming unit category at time The predicted values ​​of energy consumption by item;

[0043] By minimizing the joint loss function The building features, equipment control costs, physical constraints, and temporal consistency constraints are permanently embedded into the model parameters, resulting in a calibrated sub-item energy consumption prediction model for sub-item energy consumption prediction.

[0044] As a preferred embodiment of the present invention, based on the sub-item energy consumption prediction results output by the calibrated sub-item energy consumption prediction model, the deviation between the predicted energy consumption of each energy-consuming unit and the target energy consumption reference value is calculated. When the deviation exceeds the control threshold for dynamic adjustment, rule control, model prediction control and reinforcement learning control strategies are triggered in parallel to generate a candidate control instruction set.

[0045] A multi-objective evaluation function is constructed that includes energy consumption deviation, equipment control cost, and comfort impact, and the candidate control instruction set is comprehensively scored.

[0046] The optimal control strategy is selected based on the scoring results, and the regulation instructions of the optimal control strategy are prospectively simulated and verified using the calibrated sub-item energy consumption prediction model. When the simulation results show that the system state exceeds the limit or the equipment instructions conflict, the instruction correction mechanism is activated.

[0047] Execute the control commands that have passed the verification.

[0048] As a preferred embodiment of the present invention, the step of collecting and executing the operational status data and providing virtual-real consistency feedback, and updating the sub-item energy consumption prediction model through transfer learning based on the feedback results, includes:

[0049] The system collects operational status data after the execution of energy consumption control commands, compares the virtual and real data for consistency with the predicted data, calculates three indicators: energy consumption sequence fitting degree, peak deviation rate, and trend consistency, and calculates a comprehensive consistency index by weighting. When the comprehensive consistency index is lower than a set threshold, the system triggers a transfer learning update of the calibrated sub-item energy consumption prediction model.

[0050] The transfer learning update employs a dual-path hybrid update strategy:

[0051] Path 1 aligns and corrects the feature distribution of historical data and newly collected operational status data based on the maximum mean difference criterion;

[0052] Path 2 uses an elastic weight consolidation algorithm to evaluate the importance of the parameters of the sub-item energy consumption prediction model to historical data, and imposes constraints on the update magnitude of highly important parameters when adapting to new data;

[0053] A phased, incremental update process is established, first updating the calibration parameters of the scale calibration network, and then fine-tuning all the parameters of the sub-item energy consumption prediction model.

[0054] Energy consumption control instructions are regenerated based on the updated sub-item energy consumption prediction model, and the building energy consumption optimization results are output.

[0055] A system for optimizing and controlling building energy consumption through sub-item metering, the system comprising:

[0056] The data acquisition module is used to collect fine-grained energy consumption data and personnel activity data of energy-consuming units in the building through a sensor network, and simultaneously acquire outdoor temperature data and building characteristic data.

[0057] The data processing and analysis module is used to perform event-level segmentation processing on the power transition events of the energy consumption signal based on the fine-grained energy consumption data and personnel activity data to obtain energy consumption sub-sequences; to determine the energy-consuming unit category and category confidence of the energy consumption sub-sequences using an interpretable energy consumption classification model; and to perform non-stationarity correction processing on the energy consumption sub-sequences to obtain stable component energy consumption sequences.

[0058] The prediction model construction and calibration module is used to jointly represent the sub-item energy consumption sequence with the outdoor temperature data of the corresponding time period, construct a sequence prediction model based on temperature condition embedding, and perform scale calibration and time series consistency correction on the sequence prediction model according to the building area and number of floors in the building feature data and the error-cost joint optimization criterion to obtain a calibrated sub-item energy consumption prediction model, and use the calibrated sub-item energy consumption prediction model to predict sub-item energy consumption.

[0059] The energy consumption control module is used to trigger the energy consumption control process when the deviation between the predicted energy consumption of each item and the target energy consumption reference value exceeds the control threshold. It constructs a set of control strategies that includes rule control, model prediction control and reinforcement learning control strategies, infers the occupancy status and uncertainty based on the personnel activity data, dynamically adjusts the control threshold and selects the optimal control strategy, and generates corresponding energy consumption control instructions for execution.

[0060] The feedback learning and optimization module is used to collect the running status data after execution and provide virtual-real consistency feedback. Based on the feedback results, the module performs transfer learning to update the sub-item energy consumption prediction model and outputs the building energy consumption optimization results.

[0061] The beneficial effects of this invention are as follows: By using event-level segmented processing and interpretability classification, the energy consumption metering granularity is refined from the overall building to specific equipment categories, providing a reliable data foundation for accurate energy efficiency diagnosis and optimization; by embedding temperature conditions and using building physics-based scale calibration, the prediction model can not only capture the influence of external environmental factors, but also inherently reflect the thermal characteristics and scale patterns of different buildings, effectively avoiding prediction bias caused by individual building differences; by using multi-strategy parallel decision-making and forward-looking safety verification, the limitations of single control methods in terms of complexity, adaptability, or safety are overcome, achieving dynamic and precise control of energy consumption while ensuring indoor comfort and equipment safety; based on the transfer learning mechanism of virtual-real consistency feedback, the prediction and control model can continuously self-correct using operating data, effectively adapting to long-term drift factors such as equipment performance degradation and changes in energy consumption habits, ensuring the effectiveness and stability of the system throughout its entire life cycle. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the modular structure of the system in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0064] like Figure 1 As shown, this is an embodiment of the present invention. This embodiment provides a method for optimizing and controlling the metering of building energy consumption sub-items. Through multi-source data acquisition, event-based energy consumption segmentation, temperature-condition-embedded energy consumption prediction, and multi-strategy control, it realizes real-time metering, prediction, and optimized control of energy consumption of each energy-consuming unit in the building. It is generally run on a building energy management system (BEMS) or a building automation system (BAS).

[0065] This method mainly includes the following steps:

[0066] S1: Collect fine-grained energy consumption data and personnel activity data of energy-consuming units in the building through a sensor network, and simultaneously acquire outdoor temperature data and building characteristic data.

[0067] In one embodiment, energy consumption acquisition modules (such as smart meters or sensors) are installed in lighting circuits, socket circuits, air conditioning systems, hot water supply systems, and other major equipment circuits. Each module periodically collects parameters such as voltage, current, active power, reactive power, power factor, harmonic components (obtained by fast Fourier transform analysis of current or voltage waveforms, used to identify nonlinear loads, such as the start-up, shutdown, and operating status of computers and frequency converters), equipment switch status (relay and circuit breaker status directly read through dry contacts or communication protocols, providing the most direct equipment start-up and shutdown information), and sampling timestamps, i.e., fine-grained energy consumption data. Each energy consumption acquisition module is timestamped through a unified clock synchronization mechanism (such as a Network Time Protocol (NTP) server or GPS clock source inside the building) and records a sampling quality label in the sampling data frame. This label is used to reflect the sampling packet loss rate (the number of data packets lost per unit time, reflecting the reliability of the communication link), noise level (usually expressed as the signal-to-noise ratio (SNR), calculated by analyzing the fluctuation of the sampling signal, reflecting the working status of the sensor and the degree of electromagnetic interference), and equipment calibration status (a flag bit that records the last calibration time of the acquisition module; if the calibration validity period has expired, it is marked as "pending calibration" to remind maintenance personnel).

[0068] Personnel activity data includes indoor personnel number, personnel density, personnel activity intensity, personnel stay time, personnel entry and exit flow and occupancy status data. It is acquired by integrating various Internet of Things sensors in the building, such as infrared sensors, human body induction detectors, access control systems, people flow counters, carbon dioxide concentration sensors and wireless terminal detection devices deployed in main rooms or areas, to identify personnel entry, exit and activity intensity.

[0069] For example, the number of people indoors is obtained by counting infrared counters installed at the entrance of an area or by a camera-based facial recognition / target counting system; in specific functional areas (such as open-plan offices), the density of people is estimated by analyzing the number of mobile device MAC addresses captured by Wi-Fi probes or the data change trends of carbon dioxide (CO2) concentration sensors; the frequency and amplitude of human movement are detected by passive infrared (PIR) sensors deployed in the room, or the intensity of human activity is quantified by analyzing changes in ambient sound levels in decibels; the dwell time of people is calculated by analyzing the duration of their presence within the monitoring range of a certain sensor; at main entrances, the number of people entering and leaving per unit time is counted by a two-way counting access control system to obtain the flow of people in and out; the occupancy status data is a comprehensive judgment result, usually in the form of "occupied", "idle" or "transitional", which is inferred from the above data by simple rules (e.g., "occupied" if the number of people > 0) or a lightweight machine learning model.

[0070] Outdoor temperature data includes meteorological parameters such as dry-bulb temperature, wet-bulb temperature, relative humidity, solar radiation intensity, and wind speed for the corresponding time period. These parameters are obtained through temperature sensors installed on the exterior of the building or by connecting to a local meteorological service interface.

[0071] Building feature data consists of static or semi-static parameters extracted from Building Information Modeling (BIM), architectural drawings, or operation and maintenance management systems. These parameters include: building area, number of floors, floor distribution, thermal parameters of the building envelope (such as the heat transfer coefficient of walls and windows, solar heat gain coefficient, etc., which directly affect the building's air conditioning load), orientation information, functional zoning labels (labels that logically divide the building's internal space, such as "Office-East Zone," "Meeting Room A," "Lobby," etc., used to associate energy consumption and personnel data with spatial location), and spatial coordinates of the energy consumption acquisition module (recording the specific installation location of each smart meter or sensor within the building, such as floor and room number). This data is then linked with the energy consumption acquisition module to achieve a one-to-one mapping between energy consumption data and the building's spatial structure.

[0072] All the time-series data (including fine-grained energy consumption data, personnel activity data, and outdoor temperature data) are synchronized using a unified timestamp to ensure that data from different sources can be aligned and used together in the time dimension. The system automatically checks the integrity and time matching of the data; if time drift or data loss exists, it can be supplemented through interpolation or time calibration.

[0073] S2: Based on fine-grained energy consumption data and personnel activity data, the energy consumption subsequence is obtained by performing event-level segmentation processing through power transition events of energy consumption signals; an interpretable energy consumption classification model is used to determine the energy-consuming unit category and category confidence of the energy consumption subsequence, and the non-stationarity correction processing of the energy consumption subsequence is performed to obtain the stabilized sub-item energy consumption sequence.

[0074] Building energy consumption data is continuously changing, including various "events" caused by equipment start-ups, shutdowns, and status changes. Direct analysis from continuous data is very difficult and easily affected by noise and spurious signals. This embodiment provides a method for automatically and accurately segmenting high-quality energy consumption subsequences from continuous energy consumption data. The core of this method lies in ensuring that each subsequence corresponds to a real and independent energy consumption event through multi-step detection, verification, spurious removal, and fusion. Specifically, the method for obtaining energy consumption subsequences includes:

[0075] Preprocessing fine-grained energy consumption data to suppress noise, for example, filtering the active power time series in the fine-grained energy consumption data (as the core analysis object), such as using a moving average filter or a median filter, to smooth random fluctuations and suppress noise;

[0076] At each sampling point, the power difference between it and the previous sampling point is calculated and divided by the sampling time interval to obtain the instantaneous power change rate.

[0077] An adaptive dual-threshold mechanism is adopted, meaning the threshold is not a fixed value, but is dynamically calculated based on the local characteristics of the signal (such as the signal standard deviation within a recent time window, i.e., the local signal-to-noise ratio). The threshold can be set as follows:

[0078] Upper threshold = baseline noise level + k1 × local standard deviation (used to detect power increase events, such as device startup).

[0079] Lower threshold = Baseline noise level - k2 × Local standard deviation (used to detect power drop events, such as device shutdown);

[0080] Where k1 and k2 are coefficients set based on experience;

[0081] When the rate of change of power exceeds the upper threshold or falls below the lower threshold, a potential power transition event is marked at that point;

[0082] Relying solely on the rate of change may lead to misjudgment; therefore, it is necessary to perform joint verification of marked events using multiple features, jointly analyzing the time-domain sharpness of the power transition event and the frequency-domain energy distribution change. Time-domain sharpness is calculated by taking the second derivative of the power curve near the event point; a sharp transition will exhibit a distinct pulse peak. Frequency-domain energy distribution change involves performing a short-time Fourier transform (STFT) on the signal within two short time windows before and after the event point, comparing the distribution changes of its spectral energy (especially harmonic components). When motor-type equipment starts up, low-frequency harmonic energy typically increases significantly. Only when the joint analysis results meet set conditions (e.g., at least one of the two features, time-domain sharpness and frequency-domain energy distribution change, exceeds its respective verification threshold, or is comprehensively scored according to set weights), is it finally confirmed as a genuine power transition event. The start and end boundaries of the event are precisely determined as the sampling points where the change begins and ends, and an initial energy consumption subsequence is segmented from the fine-grained energy consumption data based on these boundaries.

[0083] The initial energy consumption subsequence (i.e., power data between two boundary points) is subjected to time-series correlation analysis with synchronously collected personnel activity data. By calculating the spatiotemporal overlap between the intensity of personnel activity and the initial energy consumption subsequence during the duration of the power transition event, false events caused by measurement noise or non-personnel factors are eliminated, and effective energy consumption subsequences are selected. For example, for an energy consumption event that occurs in a conference room, it is checked whether the "number of people" in the conference room is greater than 0 and whether the "intensity of personnel activity" has significantly increased during its duration. A quantitative "overlap" score is calculated as the spatiotemporal overlap. If the spatiotemporal overlap score of an energy consumption subsequence is extremely low (e.g., a small power fluctuation detected when no one is around at night), it is judged as a false event caused by measurement noise or non-personnel factors (such as a water dispenser that is turned on at a time) and eliminated. The remaining subsequences are the effective energy consumption subsequences that are highly correlated with personnel activity.

[0084] The effective energy consumption subsequences undergo boundary optimization and fragment fusion processing. Boundary optimization involves fine-tuning the cutting points, such as aligning them with the nearest local extrema to make the boundaries more accurate. For example, an air conditioner running at constant power for a long time may be mistakenly segmented into multiple subsequences due to slight fluctuations in the power grid. Fragment fusion aims to identify and merge these "pseudo-events." Based on the power shape similarity of the effective energy consumption subsequences (calculating whether the power curves of these adjacent subsequences are similar, for example, using dynamic time warping algorithms to calculate the distance between sequences, the smaller the distance, the more similar they are) and temporal proximity (checking whether the time interval between these subsequences is very short, such as less than 1 minute), when two subsequences are both similar in shape and temporally proximity, they are considered to represent the continuous operating state of the same energy-consuming device and are thus merged into a longer, more representative energy consumption subsequence. The final output is a high-quality energy consumption subsequence that has been denoised, de-spoofed, and fused.

[0085] Furthermore, methods for obtaining the stabilized component energy consumption sequence include:

[0086] Feature extraction is performed on the energy consumption subsequences obtained through event-level segmentation. Specifically, the following three types of features are calculated and extracted from each energy consumption subsequence (i.e., the power data segment before and after a power transition event) to form a high-dimensional feature vector:

[0087] Time-domain statistical characteristics: including but not limited to the mean, standard deviation, skewness, kurtosis, maximum value, minimum value, and the magnitude (power difference between the end point and the start point) and duration of the energy consumption subsequence;

[0088] Frequency domain energy characteristics: Perform fast Fourier transform on the energy consumption subsequence to calculate its amplitude in typical power frequency harmonics (such as 50Hz / 60Hz and their harmonics), as well as the energy ratio of the low frequency band (0-5Hz) to the high frequency band (above 5Hz). These characteristics help to identify the unique spectral fingerprints of different types of equipment such as motors and lighting.

[0089] Power shape profile characteristics: By calculating the derivative curve of the energy consumption subsequence, the steepness (slope) of its rising / falling edge and the fluctuation after power stabilization are extracted to describe the overall shape of power change;

[0090] The high-dimensional feature vectors are input into an interpretable energy consumption classification model for energy-consuming unit category classification. In this embodiment, the energy consumption classification model is specifically a gradient boosting decision tree model, which consists of multiple decision trees and achieves high-precision classification through ensemble learning. Its interpretability is reflected in the fact that after the model is trained, the total number of times each feature is used to split nodes in all decision trees or the average information gain can be calculated, thereby obtaining the feature importance ranking. This is one implementation method of the feature contribution visualization and interpretation module. The model outputs a category prediction result (such as "air conditioner", "lighting", "office equipment", etc.) for each input energy consumption subsequence. At the same time, the category confidence can be estimated by the proportion of votes obtained by the sample in the final ensemble voting, or by the weighted average of the purity of the training samples covered by the leaf nodes of each decision tree.

[0091] To ensure the classification model can adapt to changes in data distribution caused by equipment aging or altered energy consumption habits, this embodiment introduces a concept drift detection and online update mechanism. During online inference, the system continuously monitors the data stream of the model's prediction results. For example, after processing 100 new energy consumption subsequences, the KL divergence between the predicted class distribution of this batch of new data and the class distribution of the data in the early stages of model training is calculated. When the calculated KL divergence value exceeds a pre-set threshold (e.g., 0.1), the system determines that a significant concept drift has occurred. At this point, the system automatically triggers an incremental learning process: samples with high prediction confidence (i.e., high-confidence samples) are selected from the newly arrived data, along with their class labels (which may be confirmed through subsequent retrieval and correction), to form a new training set. This set is then used to incrementally update the existing gradient boosting decision tree model, adjusting the model parameters to adapt to the new data distribution.

[0092] For low-confidence energy consumption subsequences predicted by the model (i.e., subsequences with confidence below a preset threshold, such as 0.7), the system initiates a similarity retrieval based on dynamic time warping for auxiliary correction. Dynamic time warping effectively measures the similarity between two sequences of different lengths that exhibit scaling or delay on the time axis. The system calculates the DTW distance between the low-confidence subsequence to be corrected and each template sequence in a pre-built historical energy consumption event annotation library (a database containing standard event power patterns, annotated by experts or composed of high-confidence historical data during system initialization). Finally, the system retrieves the template sequence with the smallest DTW distance, assigns its category label to the current low-confidence energy consumption subsequence, and updates the confidence corresponding to that category to a higher value (such as the baseline confidence of the original template sequence).

[0093] After obtaining the results of all energy consumption subsequences after classification and confidence updates, the system generates a preliminary sub-item energy consumption sequence. This is achieved by summing according to the energy-consuming unit category: at each sampling time t, the power values ​​contributed by all energy consumption subsequences classified into the same category k are summed to obtain the total energy consumption value of that category at time t. Arranging the energy consumption values ​​of all categories at all times in chronological order constitutes the preliminary sub-item energy consumption sequence.

[0094] Next, the system performs logical consistency correction based on device interlock relationships. First, a directed graph of device interlocks is constructed: The event sequences of different categories of energy-consuming units in historical data are analyzed, and methods such as Granger causality tests are used to uncover the sequential triggering relationships between them. For example, if the occurrence of the "lighting" event statistically significantly precedes the occurrence of the "air conditioning" event, an edge is added to the directed graph from the "lighting" node to the "air conditioning" node, forming a causal relationship. Then, the preliminary energy consumption sequence is corrected based on this directed graph of device interlocks. For example, if the system detects energy consumption of "air conditioning" during a period when "lighting" is not turned on, but this is impossible according to the directed graph of device interlocks, then this "air conditioning" energy consumption is attributed to misclassification or noise, and it is removed from the "air conditioning" category or merged into its true source category, thereby eliminating redundant calculations or conflict allocations.

[0095] The non-stationarity correction process is applied to the logically corrected energy consumption sequence. Since building energy consumption sequences typically contain long-term trends, seasonal patterns (such as daily or weekly cycles), and random fluctuations, this embodiment employs a seasonal trend decomposition method to decompose the sequence into a trend term, a seasonal term, and a residual term. Subsequently, a generalized autoregressive conditional heteroscedasticity model is used to analyze the residual term. If fluctuation clustering (i.e., variance changing over time) is detected, a variance stabilization transformation (such as a logarithmic transformation) is performed on it. Simultaneously, a polynomial fitting is applied to the trend term to smooth or correct its long-term variations. Finally, the sequence after removing the seasonal term and undergoing trend correction and residual stabilization is recombined to obtain the stabilized energy consumption sequence, which exhibits better statistical properties, facilitating accurate learning by subsequent prediction models.

[0096] S3: Combine the sub-item energy consumption sequence with the outdoor temperature data of the corresponding time period to construct a sequence prediction model based on temperature condition embedding. According to the building area and number of floors in the building feature data, perform scale calibration and time consistency correction on the sequence prediction model according to the error-cost joint optimization criterion to obtain the calibrated sub-item energy consumption prediction model. Use the calibrated sub-item energy consumption prediction model to predict sub-item energy consumption.

[0097] In one embodiment, a sequence prediction model based on temperature condition embedding is constructed, including:

[0098] Multi-time-scale feature extraction is performed on the stabilized energy consumption sequence to obtain energy consumption pattern feature vectors with daily, weekly, and seasonal trends.

[0099] Daily cycle characteristics: Extract the average energy consumption value at each hour within 24 hours to form a 24-dimensional daily cycle pattern vector. For example, calculate the average energy consumption at the same time every day through a sliding window.

[0100] Weekly cycle characteristics: Extract the average energy consumption value of each day (Monday to Sunday) within a week to form a 7-dimensional weekly cycle pattern vector;

[0101] Seasonal trend characteristics: The long-term trend of energy consumption is extracted using the moving average method (such as the 30-day moving average), and the annual cycle component is extracted through Fourier transform.

[0102] For outdoor temperature data, a temporal convolutional network is used to extract deep features of the historical data sequence, and the temperature change rate and cumulative temperature load index are calculated in parallel to form a multi-scale temperature temporal feature vector.

[0103] Temporal Convolutional Network Feature Extraction: A TCN network with dilated causal convolutions is used to process historical temperature sequences (such as temperature data from the previous 72 hours) to automatically learn deep patterns of temperature changes. The dilation factors of the TCN are set to 1, 2, 4, and 8 to capture dependencies at different time scales.

[0104] Temperature change rate: Calculates the temperature difference between adjacent sampling points, reflecting the degree of temperature change;

[0105] Cumulative temperature load: Calculates the cumulative number of hours when the temperature exceeds a set threshold (e.g., 26°C), reflecting the impact of sustained high or low temperatures.

[0106] By fusing energy consumption pattern feature vectors with multi-scale temperature time-series feature vectors through cross-attention, a temperature-condition-aware energy consumption representation is generated. The expression is:

[0107] ;

[0108] In the formula, The energy consumption feature query matrix is ​​obtained by linearly transforming the energy consumption pattern feature vectors. The temperature feature key value matrix is ​​obtained by transforming the multi-scale temperature time series feature vectors through two different linear transformations. This is a scaling factor used to prevent the gradient vanishing problem caused by excessively large dot products; For activation function, These are the thermal parameters of the building envelope (such as the heat transfer coefficient of the wall, the window-to-wall ratio, etc.). For building area, and To modulate network parameters;

[0109] Energy consumption characterization based on temperature conditions A sequence prediction model is constructed. The encoder uses an extended causal convolutional network with a kernel size of 3 and an exponentially increasing expansion factor (1, 2, 4, 8, ...) to ensure that the receptive field covers a sufficiently long historical sequence. The decoder uses a gated recurrent unit. In addition to the hidden state of the previous time step, the input at each time step also receives a temperature conditional context vector obtained by mapping multi-scale temperature time-series feature vectors through a gated fusion mechanism. During the inference phase, the model is forward-propagated multiple times (e.g., 50 times), and some neurons are randomly dropped each time (Monte Carlo Dropout). The uncertainty is estimated by the variance of the prediction results.

[0110] The sequence prediction model employs a phased strategy during the training phase, which mainly consists of three phases:

[0111] Pre-training phase: Train the encoder-decoder basic structure under a fixed temperature scenario (such as selecting historical data with gradual temperature changes) to enable the model to initially learn the basic pattern of energy consumption sequence;

[0112] End-to-end fine-tuning: Introducing temperature-conditional embeddings to fine-tune the entire model on the complete dataset, enabling the model to learn to adjust its predictions according to temperature changes;

[0113] Adversarial training: Mix extreme temperature scenario samples (such as heat waves and cold waves) into the training data, and add targeted gradient perturbations during the training process to improve the robustness of the model under critical conditions.

[0114] Furthermore, a calibrated partial energy consumption prediction model is obtained, the method of which includes:

[0115] The sequence prediction model is treated as a differentiable whole, and a scale calibration network is connected after the output layer. The scale calibration network is a lightweight neural network module that incorporates building area from the building feature data. After standardizing the input vector with the number of floors, the vector is concatenated to form the input vector. Structurally, it contains two fully connected layers. The first layer uses the ReLU activation function, and the second layer has no activation function. The outputs are scaling factor matrices. With bias vector The scaling network output is multiplied element-wise (scaled) with the main model output, and then the bias is added, i.e.: ,in This represents the final energy consumption prediction after scale calibration. This is the initial prediction output of the sequence prediction model.

[0116] Minimize the following joint loss function using the error-cost joint optimization criterion. The calibration parameters of the sequence prediction model and the scale calibration network are jointly optimized.

[0117] ;

[0118] In the formula, As the basic loss term, mean squared error loss is used to calculate the mean squared error loss term between the predicted energy consumption value and the actual energy consumption value. The equipment control cost penalty item is calculated based on two factors: energy consumption deviation cost (the absolute difference between predicted energy consumption and target energy consumption) and equipment operation cost (the number of equipment status changes multiplied by the unit operation cost coefficient). This is the physical constraint loss term based on the building heat balance equation; This is a time-series consistency loss term; These are hyperparameters used to balance the weights of the various loss terms;

[0119] Among them, the physical constraint loss term The calculation formula is:

[0120] ;

[0121] Temporal consistency loss term The calculation formula is:

[0122] ;

[0123] In the formula, Indicates at time The sum of the predicted values ​​of each component energy consumption, The overall heat transfer coefficient of the building envelope is calculated using building energy consumption simulation software. Indicates time The temperature difference between indoors and outdoors To predict the time step (e.g., 1 hour); Indicates time Internal heat gain related to personnel activity data, such as approximately 100W of heat generation per person, and equipment heat generation estimated based on rated power; This indicates the total number of energy-consuming unit categories; The model represents the first Each energy-consuming unit category at time The predicted values ​​of energy consumption by item;

[0124] The model training employs the gradient descent algorithm, with the optimization objective being to minimize the joint loss function. The optimal tradeoff coefficients are determined through grid search, with a typical value being... , , A warm-up learning rate strategy was adopted, with the initial learning rate set to 0.001 and decaying every 10 epochs.

[0125] After the above joint optimization, the complete model parameters, including the scale calibration network, are saved as the final component energy consumption prediction model. In practical applications, simply inputting real-time data into this model will directly output component energy consumption prediction results that conform to physical laws, are scaled reasonably, and are self-consistent.

[0126] S4: When the deviation between the predicted sub-item energy consumption and the target energy consumption reference value exceeds the control threshold, the energy consumption regulation process is triggered: Construct a set of control strategies including rule control, model prediction control and reinforcement learning control strategies, combine personnel activity data to infer the occupancy status and uncertainty, dynamically adjust the control threshold and select the optimal control strategy, and generate corresponding energy consumption regulation instructions for execution.

[0127] In one specific embodiment, step S4 includes:

[0128] Based on the occupancy status and its uncertainty inferred from personnel activity data, the control threshold is dynamically adjusted. Specifically, when the personnel occupancy status is certain (such as office working hours), a strict energy consumption control threshold is adopted; when the occupancy status is uncertain (such as holidays or nighttime), the control threshold is appropriately relaxed to reduce unnecessary control actions.

[0129] Based on the sub-item energy consumption prediction results output by the calibrated sub-item energy consumption prediction model, the deviation between the predicted energy consumption of each energy-consuming unit and the target energy consumption reference value is calculated. When the deviation exceeds the control threshold for dynamic adjustment, rule control, model prediction control and reinforcement learning control strategies are triggered in parallel to generate a candidate control instruction set.

[0130] Rule-based control strategy: Based on a preset "if-then" rule base, such as "If the indoor temperature is >26℃ and there are people, then turn on the air conditioner";

[0131] Model predictive control strategy: Using a calibrated partial energy consumption prediction model as a constraint, the optimal control sequence is solved in the rolling time domain to minimize the energy consumption deviation at multiple future times;

[0132] Reinforcement learning control strategy: Based on deep Q-network, the state space is energy consumption and environmental parameters, and the action space is device control commands. The control strategy is optimized through long-term rewards.

[0133] A multi-objective evaluation function is constructed, which includes energy consumption deviation (the absolute difference between predicted energy consumption and target value), equipment control cost (the frequency of equipment status changes multiplied by the unit action cost coefficient) and comfort impact (the degree to which indoor environmental parameters deviate from the comfort range, such as temperature and humidity). The weighted summation method is used to calculate the comprehensive score of each candidate strategy in the candidate control instruction set. The weight coefficients are determined by optimization through expert experience or historical data. The strategy with the highest score is selected as the optimal control strategy.

[0134] Using a calibrated partial energy consumption prediction model, the system response under the selected optimal control strategy is simulated over a future period (e.g., 2 hours). A correction mechanism is triggered when the simulation results show the following:

[0135] System status exceeding limits: such as indoor temperature exceeding the comfort range, equipment power exceeding limits, etc.

[0136] Equipment command conflict: such as receiving conflicting commands for heating and cooling at the same time;

[0137] Conflicting instructions can be adjusted using rule-based reasoning or optimization algorithms, such as adjusting the execution timing, modifying parameter settings, or replacing them with alternative instructions.

[0138] The verified control commands are sent to each energy-consuming device for execution through building automation systems (such as BACnet and Modbus protocols), while recording key information such as command content and execution time.

[0139] S5: Collect operational status data after execution and provide feedback on virtual-real consistency. Based on the feedback results, perform transfer learning to update the sub-item energy consumption prediction model and output the building energy consumption optimization results.

[0140] The actual operating status data after the execution of control commands is collected through a sensor network, including:

[0141] Energy consumption data: actual power and electricity consumption of each energy-consuming unit;

[0142] Environmental data: indoor and outdoor temperature, humidity, and illuminance;

[0143] Equipment status: equipment operating mode, set parameters, fault status;

[0144] The actual data and the predicted data are compared over time to calculate consistency indices, including:

[0145] Energy consumption sequence fit: the correlation coefficient between the actual energy consumption sequence and the predicted sequence;

[0146] Peak deviation rate: The relative error between the actual peak value and the predicted peak value;

[0147] Trend consistency: The degree to which the actual trend matches the predicted trend;

[0148] The three consistency indicators are weighted to calculate the comprehensive consistency index. When the comprehensive consistency index is lower than the set threshold (e.g., 0.8), the transfer learning update process is automatically started. The transfer learning update adopts a dual-path hybrid update strategy.

[0149] Path 1: Feature Distribution Alignment

[0150] The Maximum Mean Discrepancy (MMD) criterion is used to measure the difference in distribution between historical data and newly collected data in the feature space.

[0151] By minimizing the MMD loss, the parameters of the model's feature extraction layer are adjusted to make the feature distribution of the new and old data more consistent.

[0152] The specific implementation uses the Gaussian kernel function to calculate MMD, and adds an MMD penalty term to the model training objective.

[0153] Path Two: Consolidation of Elastic Weights

[0154] Evaluating the importance of model parameters to historical data: During model training, the importance of parameters is quantified by calculating the second derivative of the loss function with respect to each parameter (Fisher information matrix);

[0155] Constraining the update of important parameters: When adapting to new data, apply update constraints to parameters that are of high importance, and the updated parameter values ​​should not deviate too much from the original values;

[0156] Specifically, this is achieved by adding an Elastic Weight Consolidation (EWC) regularization term to the loss function, in the form of: ,in To reinforce the regularization loss value for elastic weights; This is the EWC regularization strength hyperparameter, used to control the degree to which history is retained, and is usually set to [10, 1000]. For the first The diagonal elements of the Fisher information matrix for each parameter are used to quantify the importance of that parameter to historical tasks. For the current model, the first The values ​​of each parameter are the real-time parameter values ​​during the transfer learning update process; For reference model number The values ​​of the parameters are the original parameter values ​​saved before the start of transfer learning.

[0157] Establish a phased, incremental update process:

[0158] Phase 1: Rapid Adaptation of Scale Calibration Network

[0159] The backbone parameters of the fixed sequence prediction model are updated only, along with the calibration parameters of the scale calibration network.

[0160] Using newly acquired operational status data, the network parameters are optimized and calibrated using the gradient descent method.

[0161] At this stage, the learning rate is set relatively high (e.g., 0.01) to achieve rapid adaptation.

[0162] Phase Two: Full Model Fine-Tuning

[0163] Unfreeze all parameters of the sub-item energy consumption prediction model and fine-tune them end-to-end together with the scaling calibration network;

[0164] Use a small learning rate (e.g., 0.001) to avoid overfitting and performance oscillations;

[0165] A dual-path update strategy is applied simultaneously during training to ensure a balance between knowledge retention and new knowledge learning.

[0166] The updated sub-item energy consumption prediction model is redeployed into the prediction system. Energy consumption control instructions are regenerated based on the new model, and the energy consumption control process continues to be executed. The building energy consumption optimization results are output, which mainly include the following key performance indicators:

[0167] Energy saving rate: the percentage of energy saved relative to the baseline operating conditions;

[0168] Comfort retention rate: The percentage of time that indoor environmental parameters remain within the comfortable range;

[0169] Equipment operating efficiency: The ratio of the actual operating efficiency of the equipment to its rated efficiency.

[0170] like Figure 2 As shown, another embodiment of the present invention provides a building energy consumption sub-metering optimization and control system, mainly comprising:

[0171] The data acquisition module is used to collect fine-grained energy consumption data and personnel activity data of energy-consuming units in the building through a sensor network, and simultaneously acquire outdoor temperature data and building characteristic data.

[0172] The data processing and analysis module is used to perform event-level segmentation processing on power transition events of energy consumption signals based on fine-grained energy consumption data and personnel activity data to obtain energy consumption sub-sequences; it uses an interpretable energy consumption classification model to determine the energy-consuming unit category and category confidence of the energy consumption sub-sequences, and performs non-stationarity correction processing on the energy consumption sub-sequences to obtain stable component energy consumption sequences.

[0173] The prediction model construction and calibration module is used to jointly represent the sub-item energy consumption sequence with the outdoor temperature data of the corresponding time period, construct a sequence prediction model based on temperature condition embedding, and perform scale calibration and temporal consistency correction on the sequence prediction model according to the error-cost joint optimization criterion based on the building area and number of floors in the building feature data, to obtain the calibrated sub-item energy consumption prediction model, and use the calibrated sub-item energy consumption prediction model to predict sub-item energy consumption.

[0174] The energy consumption control module is used to trigger the energy consumption control process when the deviation between the predicted energy consumption of each item and the target energy consumption reference value exceeds the control threshold. It constructs a set of control strategies that includes rule control, model prediction control and reinforcement learning control strategies, infers the occupancy status and uncertainty by combining personnel activity data, dynamically adjusts the control threshold and selects the optimal control strategy, and generates corresponding energy consumption control instructions for execution.

[0175] The feedback learning and optimization module is used to collect operational status data after execution and provide feedback on virtual-real consistency. Based on the feedback results, it performs transfer learning to update the sub-item energy consumption prediction model and outputs the building energy consumption optimization results.

[0176] In summary, this invention systematically solves key problems in building energy consumption sub-item metering and optimization control, such as rough perception, inaccurate prediction, rigid control, and model degradation, by constructing a complete technical chain of "refined perception - accurate prediction - intelligent control - closed-loop evolution". It provides a complete, efficient, and self-learning solution for achieving in-depth, continuous, and reliable improvement of building energy efficiency.

[0177] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A building energy consumption sub-metering optimization control method, characterized in that, The method includes: The system collects fine-grained energy consumption data and occupant activity data from energy-consuming units within the building through a sensor network, and simultaneously acquires outdoor temperature data and building characteristic data. Based on the fine-grained energy consumption data and personnel activity data, event-level segmentation processing is performed on the power transition events of the energy consumption signal to obtain energy consumption sub-sequences; an interpretable energy consumption classification model is used to determine the energy-consuming unit category and category confidence of the energy consumption sub-sequences, and non-stationarity correction processing is performed on the energy consumption sub-sequences to obtain stable component energy consumption sequences. The sub-item energy consumption sequence is jointly represented with the outdoor temperature data of the corresponding time period to construct a sequence prediction model based on temperature condition embedding. According to the building area and number of floors in the building feature data, the sequence prediction model is scaled and time series consistency corrected according to the error-cost joint optimization criterion to obtain a calibrated sub-item energy consumption prediction model. The calibrated sub-item energy consumption prediction model is used to predict sub-item energy consumption. The construction of the sequence prediction model based on temperature condition embedding includes: Multi-time-scale feature extraction is performed on the stabilized energy consumption sequence to obtain energy consumption pattern feature vectors with daily cycle, weekly cycle and seasonal trend. For the outdoor temperature data, a temporal convolutional network is used to extract the deep features of the historical data sequence, and the temperature change rate and cumulative temperature load index are calculated in parallel to form a multi-scale temperature temporal feature vector. The energy consumption mode feature vector is cross-attention fused with a multi-scale temperature time sequence feature vector to generate a temperature condition-aware energy consumption representation The expression is: ； wherein, is the energy consumption feature query matrix, is the temperature feature key-value matrix, is the scaling factor, is the activation function, is the building envelope thermal parameter, is the building area, and is the modulation network parameter; Energy consumption characterization based on the aforementioned temperature conditions A sequence prediction model is constructed with dilated causal convolution as the encoder and gated recurrent unit as the decoder. At each time step of the decoder, the temperature condition context vector is received by mapping the multi-scale temperature time-series feature vector through a gated fusion mechanism. The sequence prediction model adopts a phased strategy during the training phase, including encoder-decoder pre-training under fixed temperature scenarios, end-to-end fine-tuning with the introduction of temperature-conditional context vectors, and an adversarial training strategy under extreme temperature scenarios. When the deviation between the predicted energy consumption of each item and the target energy consumption reference value exceeds the control threshold, the energy consumption regulation process is triggered: a set of control strategies including rule control, model prediction control and reinforcement learning control strategies are constructed, the occupancy status and uncertainty are inferred by combining the personnel activity data, the control threshold is dynamically adjusted and the optimal control strategy is selected, and the corresponding energy consumption regulation command is generated and executed. Collect operational status data after execution and provide feedback on virtual-real consistency. Based on the feedback results, perform transfer learning to update the sub-item energy consumption prediction model and output building energy consumption optimization results.

2. The method for optimizing and controlling building energy consumption by sub-item metering according to claim 1, characterized in that, The fine-grained energy consumption data includes the voltage, current, active power, reactive power, power factor, harmonic components, and equipment switching status of each energy-consuming unit in the building, collected by the energy consumption acquisition module. The energy consumption acquisition module uses a unified clock synchronization mechanism to align timestamps and records sampling quality labels in the sampling data frames to reflect the sampling packet loss rate, noise level, and device calibration status. The personnel activity data includes the number of people indoors, personnel density, personnel activity intensity, personnel stay time, personnel entry and exit flow and occupancy status data; the building characteristic data includes building area, number of floors, floor distribution, thermal parameters of building envelope, orientation information, functional zoning identifiers and spatial coordinates of energy consumption acquisition modules.

3. The method for optimizing and controlling building energy consumption through sub-item metering according to claim 1, characterized in that, The specific methods for obtaining energy-consuming subsequences include: The fine-grained energy consumption data is preprocessed to suppress noise, and the power change rate of the fine-grained energy consumption data is calculated; an adaptive dual threshold mechanism is adopted to dynamically set the judgment threshold of the power change rate based on local signal characteristics, and when the power change rate exceeds the judgment threshold, a power transition event is marked. The sharpness of the time-domain change and the frequency-domain energy distribution change of the power transition event are jointly analyzed. When the joint analysis results meet the set conditions, the start and end boundaries of the power transition event are confirmed, and the initial energy consumption subsequence is segmented from the fine-grained energy consumption data based on the start and end boundaries. The initial energy consumption subsequence is subjected to time-series correlation analysis with the synchronously collected personnel activity data. By calculating the spatiotemporal overlap between the intensity of personnel activity and the initial energy consumption subsequence during the duration of the power transition event, false events caused by measurement noise or non-personnel factors are eliminated, and effective energy consumption subsequences are selected. The effective energy consumption subsequence is subjected to boundary optimization and segment fusion processing. The segment fusion is based on the power pattern similarity and temporal proximity of the effective energy consumption subsequence, and merges multiple effective energy consumption subsequences that represent the continuous operating state of the same energy-consuming device, and outputs the final energy consumption subsequence after denoising and fusion.

4. The method for optimizing and controlling building energy consumption through sub-item metering according to claim 1, characterized in that, The methods for obtaining the stabilized component energy consumption sequence specifically include: The time-domain statistical features, frequency-domain energy features, and power shape contour features of the energy consumption subsequence are extracted to construct a high-dimensional feature vector. Based on the high-dimensional feature vector, an interpretable energy consumption classification model is used to classify the energy-consuming units. The energy consumption classification model is specifically a gradient boosting decision tree model, which outputs the category prediction results of the energy-consuming units and the corresponding category confidence. During the inference process of the gradient boosting decision tree model, concept drift detection is performed based on the model prediction result data stream. When the detected distribution change exceeds a set threshold, the incremental learning process of the gradient boosting decision tree model is triggered, and the gradient boosting decision tree model is updated online using newly arrived high-confidence samples. Based on the comparison results between the category confidence and the preset confidence threshold, a similarity retrieval based on dynamic time warping is initiated for energy consumption subsequences with low confidence. The most similar template sequence is matched from the pre-constructed historical energy consumption event label library to correct the category label and update the category confidence. Based on the energy consumption subsequences after classification and confidence updates, they are summarized according to the energy-consuming unit category to generate a preliminary sub-item energy consumption sequence. A directed graph of equipment interlocking is constructed to represent the causal relationship between different energy-consuming unit categories. Logical consistency correction is performed on the preliminary sub-item energy consumption sequence according to the directed graph of equipment interlocking to eliminate redundant calculations or conflict allocations caused by equipment linkage control. The non-stationarity correction process is applied to the logically corrected energy consumption sequence to obtain a stable energy consumption sequence.

5. The method for optimizing and controlling building energy consumption by sub-item metering according to claim 1, characterized in that, The obtained calibrated sub-item energy consumption prediction model specifically includes: The sequence prediction model is treated as a differentiable whole, and a scale calibration network is connected after the output layer. The scale calibration network uses the building area from the building feature data. With the number of floors as a conditional input, a set of calibration parameters is dynamically generated. The calibration parameters include at least a scaling factor matrix and a bias vector. Minimize the following joint loss function using the error-cost joint optimization criterion. To jointly optimize the calibration parameters of the sequence prediction model and the scale calibration network: ； In the formula, This is the mean square error loss term between the predicted energy consumption value and the actual energy consumption value; The equipment control cost penalty is calculated based on the deviation between the predicted energy consumption and the target energy consumption, as well as the frequency of changes in the equipment's start-up and shutdown status. This is the physical constraint loss term based on the building heat balance equation; This is a time-series consistency loss term; These are hyperparameters used to balance the weights of the various loss terms; Among them, the physical constraint loss term The calculation formula is: ； Temporal consistency loss term The calculation formula is: ； In the formula, Indicates at time The sum of the predicted values ​​of each component energy consumption, Indicates the overall heat transfer coefficient of the building envelope. Indicates time The temperature difference between indoors and outdoors To predict the time step; Indicates time Internal heat gain related to personnel activity data; This indicates the total number of energy-consuming unit categories; The model represents the first Each energy-consuming unit category at time The predicted values ​​of energy consumption by item; By minimizing the joint loss function The building features, equipment control costs, physical constraints, and temporal consistency constraints are permanently embedded into the model parameters, resulting in a calibrated sub-item energy consumption prediction model for sub-item energy consumption prediction.

6. The method for optimizing and controlling building energy consumption by sub-item metering according to claim 5, characterized in that, Based on the sub-item energy consumption prediction results output by the calibrated sub-item energy consumption prediction model, the deviation between the predicted energy consumption of each energy-consuming unit and the target energy consumption reference value is calculated. When the deviation exceeds the control threshold for dynamic adjustment, rule control, model prediction control and reinforcement learning control strategies are triggered in parallel to generate a candidate control instruction set. A multi-objective evaluation function is constructed that includes energy consumption deviation, equipment control cost, and comfort impact, and the candidate control instruction set is comprehensively scored. The optimal control strategy is selected based on the scoring results, and the regulation instructions of the optimal control strategy are prospectively simulated and verified using the calibrated sub-item energy consumption prediction model. When the simulation results show that the system state exceeds the limit or the equipment instructions conflict, the instruction correction mechanism is activated. Execute the control commands that have passed the verification.

7. The method for optimizing and controlling building energy consumption through sub-item metering according to claim 6, characterized in that, The process of collecting and executing operational status data and providing virtual-real consistency feedback, followed by updating the sub-item energy consumption prediction model through transfer learning based on the feedback results, includes: The system collects operational status data after the execution of energy consumption control commands, compares the virtual and real data for consistency with the predicted data, calculates three indicators: energy consumption sequence fitting degree, peak deviation rate, and trend consistency, and calculates a comprehensive consistency index by weighting. When the comprehensive consistency index is lower than a set threshold, the system triggers a transfer learning update of the calibrated sub-item energy consumption prediction model. The transfer learning update employs a dual-path hybrid update strategy: Path 1 aligns and corrects the feature distribution of historical data and newly collected operational status data based on the maximum mean difference criterion; Path 2 uses an elastic weight consolidation algorithm to evaluate the importance of the parameters of the sub-item energy consumption prediction model to historical data, and imposes constraints on the update magnitude of highly important parameters when adapting to new data; A phased, incremental update process is established, first updating the calibration parameters of the scale calibration network, and then fine-tuning all the parameters of the sub-item energy consumption prediction model. Energy consumption control instructions are regenerated based on the updated sub-item energy consumption prediction model, and the building energy consumption optimization results are output.

8. The control system of the building energy consumption sub-item metering optimization control method according to any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to collect fine-grained energy consumption data and personnel activity data of energy-consuming units in the building through a sensor network, and simultaneously acquire outdoor temperature data and building characteristic data. The data processing and analysis module is used to perform event-level segmentation processing on the power transition events of the energy consumption signal based on the fine-grained energy consumption data and personnel activity data to obtain energy consumption sub-sequences; to determine the energy-consuming unit category and category confidence of the energy consumption sub-sequences using an interpretable energy consumption classification model; and to perform non-stationarity correction processing on the energy consumption sub-sequences to obtain stable component energy consumption sequences. The prediction model construction and calibration module is used to jointly represent the sub-item energy consumption sequence with the outdoor temperature data of the corresponding time period, construct a sequence prediction model based on temperature condition embedding, and perform scale calibration and time series consistency correction on the sequence prediction model according to the building area and number of floors in the building feature data and the error-cost joint optimization criterion to obtain a calibrated sub-item energy consumption prediction model, and use the calibrated sub-item energy consumption prediction model to predict sub-item energy consumption. The energy consumption control module is used to trigger the energy consumption control process when the deviation between the predicted energy consumption of each item and the target energy consumption reference value exceeds the control threshold. It constructs a set of control strategies that includes rule control, model prediction control and reinforcement learning control strategies, infers the occupancy status and uncertainty based on the personnel activity data, dynamically adjusts the control threshold and selects the optimal control strategy, and generates corresponding energy consumption control instructions for execution. The feedback learning and optimization module is used to collect the running status data after execution and provide virtual-real consistency feedback. Based on the feedback results, the module performs transfer learning to update the sub-item energy consumption prediction model and outputs the building energy consumption optimization results.

Images