A system and method for collecting energy consumption of loads in a building

Abstract

The application discloses a kind of building inner load energy consumption collection system and method, its system is by imitative plant vascular structure collection unit, distributed communication network unit, edge computing processing unit, dynamic collection strategy regulation unit, carbon footprint correlation accounting unit and multidimensional visualization unit composition, its core unit is built-in lightweight ATCN network module and carbon factor dynamic update module, realize the accurate decomposition of building load and carbon emission real-time tracking, it runs in system based on " load type - operating state " two-dimensional dynamic collection implementation method, and low-carbon decision optimization method of maximum entropy inverse reinforcement learning, respectively realizes the goal of accurate configuration of collection resource and building energy efficient low-carbon operation.This system and method are suitable for multiple types of building scene, through multidimensional data synchronous collection, edge high-precision load decomposition, dynamic adaptive collection adjustment and energy consumption-carbon emission linkage control, solve the problems, such as low data precision, load identification fuzzy, poor collection efficiency, carbon emission control disconnection, etc. of traditional technology, provide all-round technical support for building energy fine management and low-carbon transformation.

Description

Technical Field

[0001] This invention relates to the field of building energy management and data acquisition technology. Specifically, it relates to a method and system for collecting energy consumption data of loads within buildings. This method is particularly applicable to various building types, such as commercial buildings, public buildings, and industrial plants, and provides a technical solution for accurately and efficiently collecting, processing, and managing the carbon footprint of energy consumption data for various energy loads, such as air conditioning, lighting, and power equipment. Background Technology

[0002] With the global energy crisis and the advancement of energy conservation and carbon reduction goals, improving the energy efficiency of buildings, as one of the core carriers of energy consumption, has become a key direction for industry development. Building load energy consumption data serves as the core basis for energy management, energy optimization, and carbon footprint accounting; the accuracy, completeness, and efficiency of its collection directly determine the operational efficiency of building energy management systems. However, current building load energy consumption data collection technologies still face many problems that urgently need to be addressed, specifically as follows:

[0003] First, existing data acquisition schemes mostly adopt a single-parameter acquisition mode, resulting in a lack of data dimensions and poor environmental adaptability of the acquisition terminals. Traditional acquisition devices often focus only on the acquisition of electrical parameters such as current and voltage, ignoring environmental and equipment status parameters closely related to load energy consumption, such as temperature, humidity, and operating conditions. This leads to the acquired data failing to fully reflect the true characteristics of load energy consumption. At the same time, the acquisition terminal has a simple structural design, weak anti-interference and resistance to extreme operating conditions, and is prone to data acquisition deviations or equipment failures in environments with high dust, oil pollution, and large temperature and humidity fluctuations.

[0004] Secondly, the load decomposition accuracy is low and the acquisition mode has limitations, resulting in poor adaptability to the edge. Currently, the mainstream acquisition methods are divided into two categories: invasive and non-invasive. Invasive acquisition requires the separate deployment of sensors in the circuits of each terminal device. Although it can obtain energy consumption data of a single device, it has problems such as complex construction, high cost, and significant damage to the original circuits of the building, making it difficult to promote on a large scale in existing buildings. Non-invasive acquisition does not require modification of the original circuits, but existing algorithms are mostly based on simple power threshold judgments or traditional machine learning models. When faced with complex scenarios of multiple loads operating in a building, the load decomposition accuracy is generally less than 75%, and the algorithm has high computing power requirements and large latency, making it unable to adapt to the computing power constraints of building edge computing nodes and difficult to achieve accurate and real-time decomposition of energy consumption data for each independent load.

[0005] Furthermore, the data acquisition strategy lacks dynamic adaptability, resulting in both resource waste and data loss. Existing data acquisition devices mostly adopt a fixed-frequency acquisition mode, failing to consider the differences in the operating characteristics of different types of loads within a building. For intermittent, highly fluctuating loads such as air conditioners and elevators, fixed low-frequency acquisition will miss critical energy consumption data such as equipment start-up and shutdown; while for continuous, stable loads such as lighting and office computers, fixed high-frequency acquisition will generate a large amount of redundant data, which not only increases data transmission and storage costs but also increases the energy consumption of the acquisition device. At the same time, the lack of a resource scheduling mechanism based on load priority makes core load data prone to loss or delay due to resource contention.

[0006] Finally, there is a disconnect between energy consumption data and carbon footprint accounting, resulting in a lack of data support for low-carbon decision-making. Current data collection technologies primarily focus on acquiring energy consumption data itself, lacking a deep integration mechanism with carbon footprint accounting. This prevents the real-time conversion of collected energy consumption data into carbon emission data, requiring manual post-calculation, which leads to low efficiency, large errors, and poor timeliness. Furthermore, the lack of visualized carbon emission data and scientific low-carbon decision-making models hinders dynamic management of carbon quotas and precise retrofitting of high-carbon-emission equipment, failing to meet the real-time decision-making needs of building low-carbon management.

[0007] Against this backdrop, there is an urgent need for a building load energy consumption collection technology that can achieve multi-dimensional synchronous data acquisition, high-precision edge load decomposition, dynamic adaptive acquisition and adjustment, and energy consumption-carbon emission linkage analysis and low-carbon decision-making. This technology aims to address the shortcomings of existing technologies and promote the development of building energy management towards refinement, intelligence, and low-carbonization. Summary of the Invention

[0008] To address the above problems, the technical solution adopted by this invention is as follows:

[0009] Optionally,

[0010] The present invention, by adopting the above-described technical solution, has the following beneficial effects:

[0011] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This invention illustrates a system for collecting energy consumption data within a building load, according to an embodiment of the present invention.

[0014] Figure 2 This invention illustrates a non-intrusive load decomposition method based on edge computing, according to an embodiment of the present invention.

[0015] Figure 3 This invention illustrates a dynamic data acquisition method for load type awareness according to an embodiment of the present invention.

[0016] Figure 4 This invention illustrates a method for real-time correlation and visualization of load consumption and carbon footprint according to an embodiment of the present invention. Detailed Implementation

[0017] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1

[0019] like Figure 1 As shown in the figure, this invention provides a system for collecting load energy consumption within a building. This system simulates the collaborative mechanism of "root absorption – vascular transport – leaf sensing" in plants, constructing a collection system of "self-powered sensing nodes – distributed transmission network – edge collaborative processing" to achieve synchronous collection and accurate analysis of load electrical parameters, environmental parameters, and equipment status parameters within buildings and parks. Specifically, it includes the following modules:

[0020] Module M11: Acquisition unit mimicking plant vascular structures.

[0021] In terms of structure, the data acquisition terminal adopts a four-layer structure: "protective skin - biomimetic conduction layer - core sensing layer - energy storage root layer," and uses a dual installation method of rail-mounted and magnetic attachment, allowing for quick fixation to power distribution cabinets, equipment casings, and wall surfaces. The functional design of each layer is as follows:

[0022] (1) Protective skin: mimics the barrier function of plant skin, such as using modified polyamide 66 material, reinforced by adding 4% glass fiber, and coated with polytetrafluoroethylene coating on the surface, which has oil resistance and chemical corrosion resistance, can withstand the temperature range of -40℃~100℃ and IP68 waterproof and dustproof performance, and is suitable for the dusty and oily environment in buildings.

[0023] (2) Bionic transport layer: simulates the transport function of plant vascular bundles, such as using a copper-aluminum composite microchannel array (copper channels transmit electrical energy, aluminum channels conduct heat), with a channel diameter of 60μm and a spacing of 250μm. A silver nanowire conductive film is laid inside the microchannel, with a line impedance ≤0.01Ω to ensure that the signal transmission attenuation is <0.5%;

[0024] (3) Core sensing layer: Simulates the sensing function of plant leaves, integrating a three-unit acquisition unit—the electrical parameter acquisition unit adopts a high-precision power metering chip, combined with a current transformer with adjustable ratio (200:5~2000:5) and a 0-1000V voltage sensor, which can acquire 24 electrical parameters such as active power, reactive power, and power quality (THD≤3%), with a sampling accuracy of 0.05 level and an actual power measurement error ≤±0.01kW; the environmental parameter acquisition unit integrates temperature, humidity and air pressure sensors and air quality sensors (accuracy ±0.1℃ / ±0.3%RH / ±1hPa), and synchronously acquires the temperature, humidity, air pressure and harmful gas concentration around the load; the equipment status acquisition unit adopts a temperature sensor and a high-range vibration sensor (range 0-200g) to monitor the equipment winding temperature and operating vibration frequency in real time;

[0025] (4) Energy Storage Root Layer: Simulating the energy storage function of plant roots, it consists of thermoelectric power generation modules, supercapacitors, and backup power supplies. Thermoelectric modules, such as 56 sets of N-type ( ) and P-type ( The thermoelectric elements are connected in a matrix. The hot end is attached to the device shell through a biomimetic conductive layer, and the cold end is equipped with a heat pipe cooling system. It generates electrical energy by utilizing the temperature difference between the device and the environment (minimum effective temperature difference of 0.8℃). After being regulated by the energy management chip, it outputs 3.3V / 5V dual DC power. The supercapacitor (capacity 3F / 5.5V) stores electrical energy, and the backup battery automatically replenishes power when the temperature difference is insufficient to ensure continuous operation of the terminal.

[0026] For terminal calibration and identification configuration, a standard power calibration device (accuracy class 0.01) is used to perform multi-point calibration of the electrical parameter acquisition unit. By adjusting the chip gain resistor, the measurement error in different power ranges (0.1kW~500kW) is controlled within ±0.1%. The environmental parameter and status parameter acquisition units are calibrated using a constant temperature and humidity chamber and a standard vibration table to ensure that the parameter measurement deviation meets industrial-grade requirements. After calibration, each terminal is assigned a unique MAC address and physical identifier (e.g., "P1 location A-05-12" represents device 5 and terminal 12 at location P1), and the parameters are written via the Bluetooth configuration module.

[0027] Module M12: Terminal distributed deployment and network networking.

[0028] Regarding the deployment plan, it follows the principle of "load power classification and separate deployment for key equipment": for example, two terminals are deployed for each high-power equipment above 50kW (one at the power input end and one on the equipment body); one terminal is deployed for each equipment between 10kW and 50kW; and for dispersed loads below 10kW, one terminal is deployed for every 30㎡. During terminal installation, ensure that the thermal conductivity of the hot end of the thermoelectric module in contact with the equipment is ≥4.0W / (m·K), and that there is good air convection at the heat dissipation end to improve thermoelectric conversion efficiency.

[0029] In terms of building a multi-level communication network, a four-level communication architecture of "terminal-LoRa gateway-edge computing node-cloud platform" is adopted: the acquisition terminal has a built-in communication module, operates at a frequency of 433MHz, and has a transmission distance of up to 2000 meters. Spread spectrum communication technology is used to improve anti-interference capabilities; from the LoRa gateway to the central node, the edge computing node is connected via industrial Ethernet; the edge computing node establishes communication with the cloud energy management platform through a dedicated fiber optic line with a bandwidth of ≥100Mbps to ensure real-time data transmission.

[0030] In terms of network initialization and link optimization, network configuration commands are issued through the cloud platform, and each terminal automatically completes network registration and frequency band matching. Edge computing nodes inspect the communication quality of all terminals. For terminals with signal strength below -85dBm, optimization is carried out by adjusting the gateway antenna angle, adding signal repeaters, or optimizing the installation position to ensure that the communication success rate between the terminal and the gateway is ≥99.8%.

[0031] Module M13: Multi-source data collaborative acquisition and real-time preprocessing.

[0032] Regarding the data acquisition mode and cycle settings, a dual-mode approach of "power-level acquisition + event triggering" is adopted: for example, the acquisition cycle for electrical parameters of loads above 50kW is set to 100ms, and the acquisition cycle for environmental and status parameters is set to 500ms; the acquisition cycle for electrical parameters of loads from 10kW to 50kW is set to 200ms, and the acquisition cycle for other parameters is set to 1 second; the acquisition cycle for all parameters of loads below 10kW is uniformly set to 5 seconds. When the equipment status acquisition unit detects a sudden change in vibration frequency ≥20Hz (a precursor to equipment failure), vibration >20Hz, a sudden change in current ≥20A, or a temperature exceeding the threshold (e.g., temperature >120℃), the high-frequency acquisition mode is automatically triggered, and the acquisition cycle for all parameters is shortened to 50ms. After continuous acquisition for 20 seconds, the normal mode is restored to ensure the capture of abnormal load data.

[0033] In terms of multi-source data synchronous acquisition, multi-parameter synchronous acquisition is achieved through the microcontroller built into the terminal, and hardware trigger signals are used to synchronize each acquisition unit to ensure that the timestamp deviation of electrical parameters, environmental parameters, and status parameters is ≤1ms. For example, when acquiring motor load, the active power, power factor, stator temperature, operating vibration frequency, and ambient temperature of the motor are recorded synchronously, providing a complete data chain for analyzing the correlation between motor operating status and energy consumption.

[0034] In terms of data preprocessing and anomaly labeling, the collected data is first preprocessed by the microcontroller: the Kalman filter algorithm is used to filter out high-frequency noise in the electrical parameters, and the moving average method is used to smooth the environmental parameters; a threshold model is established based on the rated parameters of the equipment and historical operating data. When the collected data exceeds the threshold range (such as the motor power exceeding the rated value by 110%), an "anomaly identifier" is marked, and information such as the time of anomaly occurrence and the magnitude of parameter change is recorded.

[0035] Regarding data encryption and transmission, pre-processed data is encrypted using the AES-256 encryption algorithm. Normal data is transmitted to the LoRa gateway in a "batch upload of 20 data sets" mode. Abnormal data marked with "abnormal identifier" adopts a priority transmission mechanism and is immediately uploaded separately with higher transmission priority than other data to ensure rapid feedback of abnormal information.

[0036] Module M14: Edge Node Data Processing and Cloud Platform Integration Applications

[0037] In terms of data reception and time calibration, after receiving data uploaded by each terminal, the edge computing node uses its own clock module (with an accuracy of 1ms) as a benchmark to uniformly calibrate the timestamps of the data from each terminal, eliminating time deviations caused by communication delays and ensuring that the load data collected by different terminals in the same area have time consistency.

[0038] In terms of multi-source data fusion analysis, an improved DS evidence theory is used to fuse multi-dimensional data of the same load: a three-level state identification framework of "normal operation - inefficient operation - fault warning" is established, the support of each collected parameter for different states is calculated (e.g., when the motor vibration frequency exceeds the standard, the support of "fault warning" is 0.88), and the information of each parameter is fused through evidence combination rules to obtain the final operating state of the load, with an identification accuracy of 98.6%.

[0039] In terms of data storage and cloud synchronization, edge computing nodes store the merged data in a four-level classification of "region-location-device-terminal". They also cache the full amount of data for the past 60 days locally and upload it to the cloud platform in a hierarchical manner according to data type: fault warning data is uploaded in real time, high-power device data is uploaded every 1 minute, and low- and medium-power device data is uploaded every 5 minutes, thus optimizing data transmission resources.

[0040] In terms of cloud platform application and terminal parameter optimization, the cloud platform visualizes the received data, generates equipment energy consumption trend charts, energy consumption comparison charts, and fault warning dashboards, which can be viewed by managers through the web and APP. By analyzing the correlation between load energy consumption fluctuation patterns and activities, the platform sends collection parameter optimization instructions to edge nodes. For example, during peak periods (8:00-20:00), the terminal collection cycle is shortened, and during off-peak periods (20:00-8:00), it is extended. After dynamic optimization, the data transmission volume is reduced by 70%, and the terminal energy consumption is reduced by 45%.

[0041] Module M15: Anomaly Early Warning and Full-Lifecycle Data Traceability

[0042] Regarding multi-level early warning push, when an edge computing node detects an abnormal load state, it immediately initiates multi-level early warnings through the cloud platform: Level 1 early warning (equipment failure) is pushed simultaneously to equipment maintenance personnel and responsible persons via SMS, telephone, and APP pop-up; Level 2 early warning (inefficient operation) is pushed via SMS and APP; Level 3 early warning (parameter fluctuation) is stored in the system log. The early warning information includes terminal identifier, device name, abnormal parameters, and handling suggestions (such as "P1 location A-05-12 terminal: motor vibration frequency exceeds the standard, it is recommended to stop and check").

[0043] In terms of data traceability and analysis, the cloud platform supports querying historical data by multiple dimensions such as terminal identifier, device type, and time range. It can generate energy consumption reports, parameter change curves, and status diagnostic reports for any time period, with data storage time not less than 3 years. Managers can analyze equipment energy consumption patterns through data traceability. For example, if abnormal standby energy consumption of a motor is found, and it is found to be a contactor fault, the monthly average energy savings after repair can be quantitatively compared, verifying the practicality and energy-saving value of this method.

[0044] This embodiment achieves synchronous acquisition of multi-dimensional load data through an integrated terminal design mimicking plant vascular bundles, and enhances the continuity of data acquisition by combining it with thermoelectric self-powered technology. Relying on a power-level and event-triggered acquisition mode, it reduces resource consumption while ensuring data integrity. Through multi-source data fusion at edge nodes and linkage with the cloud platform, it achieves accurate identification of load status and refined energy consumption management. This method is adaptable to building and park loads of different power levels, has high acquisition accuracy and strong anti-interference capabilities, and can effectively support energy optimization scheduling and equipment operation and maintenance decisions, possessing broad application prospects.

[0045] Example 2

[0046] like Figure 2As shown, the non-intrusive load decomposition method based on edge computing in this embodiment of the invention is an extension of the core data processing in Embodiment 1. The total load power, power factor, harmonic characteristics and environmental parameters output by the acquisition terminal are encrypted and transmitted to the edge computing node. The total load is then accurately split into individual devices through this optimization algorithm.

[0047] High-precision decomposition is achieved through the following workflow: "Data preprocessing (including formula correction) - Feature extraction (including formula optimization) - ATCN (Attention-Enhanced Temporal Convolutional Network, abbreviated as ATCN, a lightweight attention-enhanced temporal convolutional network) load decomposition (including parameter configuration) - Two-factor correction (including correction formula)". Specifically, the steps include:

[0048] Step S21: Data Acquisition and Enhancement Preprocessing

[0049] (1) Data decryption and format alignment

[0050] The edge node receives the encrypted data from the aforementioned acquisition terminal, decrypts it using AES-256, and extracts the core data item: total power. (Sampling frequency f = 50Hz, t is a timestamp), power factor 3rd harmonic amplitude Ambient temperature Align non-standard timestamp data to a unified timeline using the following formula to ensure data synchronization:

[0051]

[0052] in, The sampling interval is... This is a rounding function. This is the aligned standard timestamp. In this embodiment, the value of t ranges from 0 to 86400s (daily data), and the aligned data volume is 4,320,000 records.

[0053] (2) Outlier correction and noise suppression

[0054] Outliers are handled using the "3σ criterion + linear interpolation". The outlier determination formula is as follows:

[0055]

[0056] in, for The mean of the sliding window (window size N=50, corresponding to 1 second of data). This represents the standard deviation of the window. If... If the outlier determination formula is satisfied, then it is determined that... outlier outliers The corrected normal data is obtained by linear interpolation using the formula below. :

[0057]

[0058] in, 'This is the timestamp of the outlier t and the timestamp of the previous normal data' The time difference ensures that the correction value conforms to the trend of data time series changes.

[0059] Noise suppression uses db4 wavelet thresholding for noise reduction. The calculation is performed according to the formula to ensure a balance between noise reduction effect and data integrity:

[0060]

[0061] in, Here, N represents the noise estimate, and N is the window size. In this embodiment... After noise reduction, the signal-to-noise ratio of the data is improved to over 45dB.

[0062] (3) Data normalization

[0063] To eliminate dimensional differences, the temperature, power, and other data are normalized using the following formula, mapping them to the [0,1] interval:

[0064]

[0065] Where x is the original data (e.g., ... , ), , This represents the global maximum / minimum value of the data. The temperature normalization parameter in this embodiment is: , Power normalization parameters: , (Maximum load of the park).

[0066] Step S22: Multi-dimensional feature extraction calculation and dimensionality reduction

[0067] (1) Quantization calculation of time-domain features

[0068] Based on preprocessing The four core time-domain features are calculated using the following formula:

[0069] 1) Short-term power mean:

[0070]

[0071] 2) Power change rate:

[0072]

[0073] 3) Peak power:

[0074]

[0075] 4) Operating cycle:

[0076]

[0077] In this embodiment, the motor load Peak load can reach 75kW / ms, air conditioning load The feature differentiation is significant.

[0078] (2) Frequency domain and coupling characteristic calculation

[0079] The frequency domain characteristics are calculated using Fast Fourier Transform (FFT), and the fundamental frequency is... 3rd harmonic distortion rate Calculate using the formula:

[0080]

[0081] in This represents the fundamental frequency amplitude. Combining power factor and temperature, a coupling characteristic is constructed:

[0082]

[0083]

[0084] This ultimately results in a 22-dimensional feature vector. .

[0085] (3) Feature dimensionality reduction

[0086] Principal component analysis (PCA) was used for dimensionality reduction. The core parameter was the feature covariance matrix Σ (22×22 dimensions). The principal component contribution rate was calculated using the following formula:

[0087]

[0088] in Let Σ be the k-th eigenvalue, sorted in descending order. Take the cumulative contribution rate. The first 8 principal components form the dimensionality-reduced feature vector. (8-dimensional), reducing computational load by 63.6%.

[0089] Step S23: Lightweight ATCN Network Load Decomposition

[0090] (1) Network structure and core parameters

[0091] The Lightweight Attention-Enhanced Temporal Convolutional Network (ATCN) is adapted to the computing power characteristics of edge computing nodes. It consists of 4 layers with the following parameter configuration:

[0092] 1) Input layer: Dimension = 8 (corresponding to...) ), Input sequence length L=60 (corresponding to 1.2s time series data);

[0093] 2) Deep convolutional layer: Dilated causal convolution is used, with kernel size k=3, dilation coefficient d=[1,2,4] (multi-scale capture of temporal features of the load), number of kernels K=16 (lightweight design), activation function is GELU, output dimension=16;

[0094] 3) Attention Enhancement Layer: Employs Channel Attention (CAM) to weight the output features of deep convolutional layers. The formula for calculating the attention weights is:

[0095]

[0096] in This represents the c-th channel feature output by the deep convolutional layer; GlobalAvgPool is a global average pooling operation, and its calculation logic is as follows for the temporal feature dimension (length = 60) in this embodiment:

[0097]

[0098] That is, the global mean is calculated for the 60-dimensional temporal features of each channel, compressing the temporal dimension to 1 dimension, which preserves the global information of the channel features and reduces the computational cost of the attention layer; the MLP is a 2-layer perceptron (hidden layer dimension = 8). The channel attention weights are set, and the dropout rate is 0.1 (to prevent overfitting).

[0099] 4) Output layer: The fully connected layer's dimensions are mapped from 16 to 8 (corresponding to 8 types of loads), the activation function is Sigmoid, and the output is the predicted power value for each load. (k=1,2,...,8).

[0100] (2) Core computing logic of the network

[0101] The formula for calculating dilated causal convolution in ATCN is as follows:

[0102]

[0103] in For the i-th weight of the c-th convolutional kernel, The bias term is d, and the dilation coefficient is d, which ensures that the convolution depends only on historical temporal features, which conforms to the temporal causality of load decomposition.

[0104] Feature output formula after attention enhancement:

[0105]

[0106] (3) Model training and loss function

[0107] Using the 7 days of measured data (30,240,000 data points) collected earlier, the dataset was divided into training and testing sets in a 7:3 ratio. The mean square error (MSE) between the predicted and actual load power values ​​was used as the loss function, calculated as follows:

[0108]

[0109] Where M is the sample size. The actual power of class k loads (calibrated via intrusive data acquisition). Trained using the AdamW optimizer, with an initial learning rate of... Weight decay coefficient attenuation coefficient , After 150 training rounds, the loss converged to 0.0012 (convergence speed improved by 30% compared to LSTM).

[0110] (4) Real-time decomposition output

[0111] Will Input the trained ATCN model and output the predicted power for 8 types of loads. Satisfying power balance constraints:

[0112]

[0113] Step S24: Two-factor correction improves decomposition accuracy

[0114] (1) Correction of abnormality level

[0115] Calculate the degree of anomalousness score of the decomposition results using the following formula:

[0116]

[0117] in , This represents the historical mean and standard deviation of load type k for the same period (the same time period over the past 7 days). When corrected according to the formula:

[0118]

[0119] in This is a sign function, ensuring that the correction value is within a reasonable range.

[0120] (2) Power balance correction

[0121] Calculate the deviation between the total decomposed power and the actual power. The correction amount is allocated according to the formula:

[0122]

[0123] in As power factor weights, ensure that after correction .

[0124] Step S25: Performance Verification and Output

[0125] To accurately verify the algorithm's adaptability in office building scenarios, a medium-sized office building (8000㎡, 350 employees per day) within a smart industrial park was selected as the test subject. This building's load exhibits typical characteristics of "operating from morning to evening, dynamic fluctuations, and a high proportion of low-power equipment." The test period covered a full work week (5 workdays + 2 weekend days). The algorithm's performance was verified through "real load calibration + multi-dimensional indicator comparison," while simultaneously analyzing its application value in relation to the actual energy management needs of office buildings.

[0126] (1) Office building test environment and load configuration

[0127] 1) Hardware and data acquisition configuration

[0128] Edge computing nodes: adopt high-performance edge gateway architecture commonly used in office scenarios (e.g., 4 cores 1.2GHz, 2GB DDR4 memory), simulating the computing power constraints of actual office building edge devices, without additional GPU acceleration modules;

[0129] Data acquisition terminal deployment: Following the "Bionic structural coupling acquisition method" mentioned above, data acquisition terminals are deployed in the main power distribution box of the building and the power distribution boxes of the three floors. The sampling frequency is 50Hz, and the total power, A / B / C phase voltage and current, power factor, 3rd / 5th / 7th harmonics and indoor temperature (3 monitoring points per floor) data are collected simultaneously.

[0130] True value calibration: Intrusive high-precision power sensors (accuracy ±0.05kW) are deployed in the independent power supply circuits of 8 core loads to directly collect the true power of each load as the benchmark data for algorithm verification.

[0131] 2) Core load types and characteristics of office buildings

[0132] The load on office buildings is mainly composed of "comfort load + office equipment load," and the specific configuration and operating characteristics are shown in the table below, covering various operating conditions such as steady state, intermittent operation, and dynamic fluctuations:

[0133]

[0134] 3) Test data size and distribution

[0135] Data collection period: 14 days (including 5 working days and 2 weekends, covering sunny days, cloudy days, and days with a high temperature of 35℃);

[0136] Total data: 60,480,000 records (sampling frequency of 50Hz), of which the first 7 days (30,240,000 records) are the training set and the last 7 days (30,240,000 records) are the test set;

[0137] Key test scenarios include 12,800 start-ups and shutdowns of office equipment, 3,200 frequency conversion adjustments of air conditioners, and 1,500 start-ups and shutdowns of elevators, accounting for 5.8% of the test set data.

[0138] (2) Verification of core performance indicators in office building scenarios

[0139] Based on the load characteristics of office buildings, "decomposition accuracy (with a focus on low-power loads), real-time performance, and robustness in office scenarios" were selected as the core verification indicators, and comparative tests were conducted with traditional LSTM and standard TCN algorithms.

[0140] 1) Decomposition accuracy index

[0141] The three main metrics used are mean absolute error (MAE), root mean square error (RMSE), and load category accuracy.

[0142]

[0143]

[0144]

[0145] Where M is the number of test set samples (30,240,000) and K is the number of load categories (8). The actual power value collected by the invasive sensor.

[0146] Overall accuracy comparison

[0147] Test results show that the ATCN algorithm significantly outperforms traditional algorithms in office building scenarios. Specific data is as follows:

[0148]

[0149] Low power load accuracy verification

[0150] Low-power loads (lighting, office computers, security equipment) account for up to 35% of office building data, but traditional algorithms generally have low accuracy in identifying them. This algorithm significantly improves the accuracy of low-power load identification through feature coupling and attention enhancement design:

[0151]

[0152] Accuracy performance under typical working conditions

[0153] Air conditioner inverter regulation mode: ATCN algorithm accuracy is 95.3%, LSTM is 75.8%, and standard TCN is 88.5%, because ATCN's multi-scale dilated convolution can accurately capture the power change characteristics of the inverter process;

[0154] Centralized start-up and shutdown of office equipment (e.g., 30 minutes before work): ATCN algorithm accuracy is 94.7%, LSTM is 70.2%, and standard TCN is 86.9%. The attention mechanism effectively distinguishes the superimposed power characteristics of multiple device start-ups and shutdowns.

[0155] High-power impact conditions in elevators: ATCN algorithm achieves an accuracy of 97.1%, LSTM 82.4%, and standard TCN 91.3%. The causal convolutional structure ensures the identification of the temporal correlation of impact loads.

[0156] 2) Real-time indicators

[0157] Edge gateways in office buildings typically share computing power with systems such as security and lighting, and have stringent requirements for algorithm latency (≤100ms). The test results are as follows:

[0158] Single inference latency: ATCN algorithm average 70ms, maximum 78ms; LSTM average 132ms, maximum 155ms; standard TCN average 95ms, maximum 110ms; ATCN fully meets the real-time requirements of office scenarios, and has better latency stability.

[0159] Daily data processing efficiency: For office buildings with 4,320,000 data entries per day, the ATCN algorithm completes the full decomposition, correction, and result encapsulation in just 12 minutes, while LSTM takes 28 minutes and the standard TCN takes 18 minutes, significantly improving the data processing efficiency of edge nodes.

[0160] Computing power consumption: The average CPU utilization rate of the edge gateway during the ATCN algorithm operation is 35%, LSTM is 68%, and the standard TCN is 52%. The low computing power consumption avoids resource conflicts with other office systems.

[0161] 3) Robustness verification in office scenarios

[0162] Three robustness test scenarios were designed to address the harsh operating conditions commonly found in office buildings:

[0163]

[0164] Tests show that the proposed algorithm has significantly better anti-interference capabilities than traditional algorithms under complex working conditions in office buildings, meeting the stability requirements of practical engineering applications.

[0165] (3) Scene output results and application adaptation

[0166] Specific tests in office building scenarios demonstrate that this lightweight ATCN algorithm possesses core advantages of "high accuracy, low latency, and high robustness," achieving an overall decomposition accuracy of 96.8%, with a more than 20% improvement in low-power load decomposition accuracy, and an inference latency of ≤78ms. It fully adapts to the computing power constraints and real-time management needs of edge computing nodes in office buildings. The algorithm's output can be directly integrated with existing office building energy and building control systems, providing reliable data support for refined energy management and equipment operation and maintenance optimization, demonstrating significant engineering application value.

[0167] After testing and verification of its applicability in office building scenarios, this algorithm can directly achieve refined energy consumption statistics, implementation of energy-saving optimization strategies, and early warning of equipment operation and maintenance based on the decomposition results. For example, by using the power fluctuation characteristics of water dispensers, it can detect the heating tube failure of 3 aging water dispensers 5 days in advance, avoiding equipment damage and safety hazards; by using the abnormal power impact of elevators, it can predict the wear problem of the braking system of 1 elevator, reducing maintenance costs.

[0168] In summary, this edge computing-based non-intrusive load decomposition algorithm replaces the traditional LSTM with a lightweight ATCN network and clarifies the computational logic of each stage using quantization formulas, constructing a complete technical system of "feature quantization - lightweight decomposition - two-factor correction". This algorithm solves the problems of difficult deployment, low accuracy, and high latency of traditional non-intrusive decomposition algorithms at the edge. Together with the previously mentioned acquisition method, it forms a collaborative architecture of "data acquisition - edge intelligent processing," providing core technical support for refined energy management and equipment operation and maintenance optimization in buildings and parks. Furthermore, the parameters of each formula can be flexibly adjusted according to different building scenarios (commercial, industrial, and residential), demonstrating significant engineering application value and technological innovation.

[0169] Example 3

[0170] like Figure 3 As shown in the figure, a load type-aware dynamic data acquisition strategy method according to an embodiment of the present invention includes load feature database construction, accurate load status identification, dynamic data acquisition mode matching, priority scheduling, and strategy optimization, specifically including the following steps:

[0171] Step S31: Construction and initialization of load characteristic database

[0172] This step provides a benchmark for dynamic data acquisition. Based on historical data and industry standards, a feature database covering all types of building loads is constructed, and the basic parameter configuration of the acquisition terminal is completed to ensure the stability of the system's initial operation.

[0173] (1) Construction of a multidimensional load characteristic database

[0174] The database adopts a "categorized storage-index association" structure, with tables created according to the eight load categories defined above (central air conditioning, industrial motors, office equipment, etc.). Each load category stores six core characteristic parameters. Examples of characteristic parameters for some key loads are shown in Table 1 below. The parameter values ​​were measured and calibrated using the invasive sensors mentioned above.

[0175] Table 1 Typical Load Core Characteristic Parameters

[0176]

[0177] The database supports dynamic updates. Edge computing nodes collect daily load operation data. When the average value of a certain type of load characteristic parameters (such as the operating cycle) deviates from the data in the database by more than 10% for 7 consecutive days, the update mechanism is automatically triggered to ensure that the characteristic parameters are consistent with the actual operating status of the load.

[0178] (2) Initialization configuration of data acquisition terminal parameters

[0179] Based on the importance of the area where the load is located and the load type, the parameters of 128 data acquisition terminals were initialized (corresponding to the number of terminals deployed in the aforementioned test building). The specific configuration rules are as follows:

[0180] Basic parameters: Default sampling frequency for all terminals Sampling interval The data cache capacity is set to 2000 records to ensure that at least 16 minutes of data can be cached when the network is disconnected;

[0181] Priority Identification: Priority is distinguished by the physical identifier of the terminal (referencing the aforementioned terminal identification rules). The terminal identifier prefix for core area loads (data center UPS, fire emergency equipment) is set to "C1" (Level 1 priority), the main air conditioner and elevator in the office area is set to "C2" (Level 2 priority), and the lighting in public areas and non-core office equipment is set to "C3" (Level 3 priority).

[0182] Communication parameters: Configure the terminal communication frequency band to 433MHz and the data transmission baud rate to 9600bps, matching the LoRa gateway parameters mentioned above (referencing the aforementioned network configuration standard), ensuring an initial communication success rate of ≥99.5%.

[0183] Step S32: Real-time identification of load type and operating status

[0184] After receiving real-time data uploaded by the acquisition terminal, the edge computing node accurately locates the type and operating status of the current load through a three-level identification mechanism of "feature extraction - similarity matching - status determination". The identification delay is ≤100ms, which meets the real-time control requirements.

[0185] (1) Real-time feature vector extraction

[0186] Real-time power data collected by edge computing nodes Preprocessing is performed to extract 4-dimensional core feature vectors. The calculation methods and physical meanings of each feature are as follows:

[0187] Instantaneous power : Real-time load power directly collected, unit kW, accuracy ±0.01kW (referring to the aforementioned electrical parameter collection accuracy).

[0188] Power change rate This reflects the severity of power fluctuations. The calculation formula is shown below, with units of kW / ms. It is a core indicator for identifying load start-up, shutdown, and impact.

[0189]

[0190] In the formula, (1 / 10 of the basic sampling interval, improving the accuracy of the rate of change calculation), P(t-Δt) is the power value at the previous high-precision sampling time. When the power surge is ≥5kW / ms, it is considered a power surge.

[0191] Short-term power peak This reflects the maximum short-term energy consumption of the load, taking the maximum power value within 5 seconds (100 sampling points). The calculation formula is shown below, used to distinguish between impulsive and steady-state loads:

[0192]

[0193] In the formula, , where i is the sampling point index, covering the complete short-term power fluctuation cycle.

[0194] Volatility coefficient This reflects the stability of the load power. The calculation formula is shown below. It is dimensionless and used to determine the steady-state / fluctuation state.

[0195]

[0196] In the formula, The standard deviation of power data within 10 seconds. The average power data over 10 seconds. A value ≤5% indicates stable load operation. A value greater than 15% indicates significant fluctuations.

[0197] (2) Load type similarity matching

[0198] The cosine similarity algorithm is used to compare the real-time extracted feature vector F with the standard feature vectors of various load types in the load feature database. The matching and similarity calculation are as follows:

[0199]

[0200] In the formula, For the i-th component of the real-time feature vector, This represents the i-th component of the standard feature vector in the database. When the load is determined to be of the standard load type, if there are multiple load types with a similarity of ≥0.85, the load type corresponding to the maximum similarity is selected and marked as "mixed load" (such as a scenario where computers and printers are running simultaneously in an office area).

[0201] Example verification: The real-time feature vector of a certain load, F=[10.8kW, 3.2kW / ms, 11kW, 8.5%], was collected and compared with the standard feature vector of an elevator. =[11kW, 4kW / ms, 12kW, 10%] Calculation yields Sim=0.92≥0.85, indicating that the load is an elevator, and the identification is accurate.

[0202] (3) Multi-dimensional determination of operating status

[0203] Based on load type and real-time characteristics, a three-dimensional state determination model of "start-stop-steady-abnormal" is constructed. All state determinations must meet the "continuous triggering + delayed confirmation" mechanism to avoid misjudgment caused by instantaneous interference.

[0204] Step S33: Dynamic Acquisition Pattern Matching and Execution

[0205] The edge computing node automatically matches the corresponding acquisition mode based on the identification result of "load type-operation status" and sends a sampling frequency adjustment command to the acquisition terminal through the industrial bus (command transmission delay ≤ 50ms, referring to the terminal control mechanism of Example 1) to achieve precise allocation of acquisition resources.

[0206] (1) Intermittent high-fluctuation load pattern

[0207] The core requirement for this type of load is "capturing start-stop impacts + refining steady-state data," employing a three-stage approach of "event triggering + high-frequency supplementary sampling + steady-state frequency reduction." The specific execution logic is as follows:

[0208] High-frequency event triggering: When a "start / stop / impact" state is detected, a high-frequency acquisition command of f=10Hz (sampling interval 100ms) is immediately issued, and the terminal's "data pre-storage" function is triggered simultaneously—automatically retrieving historical data from 1 second before the trigger (stored at 500ms intervals) to ensure a complete record of the power change curve before and after the event. For example, when an elevator starts, a complete 5-second data chain is formed from 1 second before start (t-1s) to 4 seconds after start (t+4s), containing 50 high-frequency sampling points, which can clearly reflect the start-up impact process.

[0209] Fluctuating operation in the mid-frequency band: When a "fluctuating operation" state is detected (such as air conditioner inverter adjustment), a mid-frequency acquisition command of f=5Hz (sampling interval 200ms) is issued, which is 2.5 times higher than the base frequency, ensuring the capture of gradual power change characteristics. If the value drops below 5% and remains below 5% for 10 seconds, the fluctuation is considered to have ended, and the system will automatically revert to the base frequency.

[0210] Steady-state operation base segment: When no event is triggered, the base frequency of f=2Hz (sampling interval 500ms) is maintained. At this time, the data redundancy rate is ≤10%, which reduces the amount of data by 80% compared with fixed high-frequency acquisition.

[0211] In office building air conditioning systems, this mode reduced the daily data collection volume from 1.728 million records in the fixed 10Hz mode to 345,600 records, a reduction of 80%, while achieving a 100% capture rate for start-stop impact data.

[0212] (2) Continuous stable load mode

[0213] The core requirement for this type of load is "balance accuracy and efficiency," employing a "fluctuation coefficient adaptive" mode based on real-time calculations. The sampling frequency is dynamically adjusted, and the specific correspondence is as follows:

[0214] when ≤5% (steady-state operation, such as continuous lighting): f=0.5Hz (sampling interval 2000ms), the daily data collection volume is only 43,200 records, which is 75% less than the base frequency;

[0215] When 5% < ≤15% (slight fluctuations, such as the start-up and shutdown of some office computers): f=2Hz (base frequency) to ensure that fluctuation data is not missed;

[0216] when >15% (drastic fluctuations, such as a concentrated surge of computer startups in an office area): f=5Hz (intermediate frequency sampling), continuing until A callback will occur after the value drops to ≤15%.

[0217] In office lighting system applications, this mode reduces the amount of data during steady-state periods by 75%, while achieving a 99.8% capture rate for fluctuating data during peak computer usage periods from 8:00 to 9:00 AM, thus achieving a balance between efficiency and accuracy.

[0218] (3) Impact load pattern

[0219] The core requirement for this type of load is "anticipated impact + full coverage", and it adopts a "historical prediction + pre-triggering + high frequency throughout" mode. The specific execution process is as follows:

[0220] Impact Time Prediction: Edge computing nodes construct a prediction model using a BP neural network based on historical operational data of impact loads stored in the database. The input features are "time period - ambient temperature - number of operations in the previous hour", and the output is "impact probability". When the probability is ≥85%, it is determined to be a high-risk impact period. For example, historical data of electric heating load shows that the impact probability is 92% from 11:30 to 12:00, so this period is set as a high-risk period.

[0221] Pre-triggered high-frequency acquisition: During high-risk impact periods, edge computing nodes send a pre-triggered command of f=10Hz to the corresponding terminal 3 seconds in advance, and the terminal enters the "high-frequency standby" state to ensure that acquisition can be carried out immediately when the impact occurs.

[0222] Impact full coverage: After an impact is detected, the frequency is maintained at f=10Hz until 5 seconds after the impact ends, ensuring complete recording of the rising edge, peak value and falling edge data of the impact. After the impact ends, the frequency is switched back to the base frequency.

[0223] In electric heating load applications, this mode increases the impact data capture rate from 78% in the fixed mode to 99.6%, completely solving the problem of missing key data for impact loads.

[0224] Step S34: Load Priority Scheduling and Data Transmission Optimization

[0225] When multiple terminals simultaneously trigger high-frequency data collection or abnormal events, the system allocates communication and computing resources through a priority scheduling mechanism to ensure that core load data is processed first and to avoid data loss caused by resource contention.

[0226] (1) Dynamic classification and identifier binding of load priority

[0227] Priority classification is not fixed and needs to be dynamically adjusted based on load operation status and building energy demand. The specific implementation process is as follows:

[0228] Basic Priority Classification: Continuing the aforementioned three-level classification of "C1-C3", the basis for classifying the load at each level and the core protection objectives are clearly defined. C1 level (core load) includes data center UPS, fire emergency power supply, etc., with the protection objective of "zero data loss and zero transmission delay"; C2 level (important load) includes office area central air conditioning, passenger elevators, production workshop main motors, etc., with the protection objective of "no omission of critical event data and transmission delay ≤500ms"; C3 level (general load) includes public area lighting, toilet exhaust fans, ordinary sockets, etc., with the protection objective of "complete statistical data and minimal resource consumption".

[0229] Priority dynamic adjustment mechanism: Based on the device status acquisition capability of the acquisition terminal in the biomimetic structural coupling multi-source load data acquisition method of the aforementioned embodiment, when a C3 level load experiences an abnormal state (such as a sudden change in lighting circuit current ≥20A, determined to be a short circuit precursor), the terminal automatically sends a "temporary priority upgrade request" to the edge node. The edge node completes the status verification within 100ms and temporarily upgrades it to C2 level, initiating high-frequency acquisition and priority transmission; after the anomaly is resolved, it automatically reverts to C3 level within 30s to avoid long-term resource occupation. For example, if a common socket in an office area suddenly overloads, after the temporary priority upgrade, the short circuit fault data is completely captured, providing accurate basis for operation and maintenance.

[0230] Priority identifier bound to terminal: The priority identifier is permanently bound to the terminal's physical ID via the terminal's built-in EEPROM storage chip, while reserving 16 bytes of erasable and rewritable space to store "temporary priority" states. When the edge node issues a data acquisition command, the first 3 bits of the command frame explicitly identify the priority (001=C1, 010=C2, 100=C3). After receiving the command, the terminal distinguishes the command priority through hardware level signals. C1-level commands directly trigger an interrupt response and are executed first.

[0231] (2) Implementation of priority-based resource acquisition scheduling

[0232] The resources collected include the terminal's sampling clock resources and the edge node's receiving buffer resources. Scheduling must follow the principle of "core resources prioritized, idle resources shared," and the specific technical means are as follows:

[0233] Terminal sampling clock scheduling: The terminal adopts a dual-clock source design (high-precision crystal oscillator + ordinary crystal oscillator). A 10MHz high-precision crystal oscillator is used for C1-level load acquisition to ensure a sampling interval error ≤1μs; a 1MHz ordinary crystal oscillator is used for C2 and C3-level load acquisition. When multiple priority acquisition tasks occur concurrently, a clock switching control module enables rapid switching within 50ns. For example, if C1-level fire-fighting equipment and C3-level lighting are triggered for acquisition simultaneously, the terminal first completes the 10Hz high-frequency sampling of the fire-fighting equipment, and then performs the 0.5Hz sampling of the lighting, ensuring no sampling timing conflicts.

[0234] Edge node cache resource allocation: Edge nodes are configured with 2GB DDR4 memory, divided into 128 independent cache areas. Among them, C1 level has 32 cache areas (2MB each) exclusively, and adopts the "circular overwrite + exception locking" strategy - normal data is stored in a circular overwrite manner, and data marked as abnormal is automatically locked and overwritten. C2 level occupies 64 cache areas (1MB each), and adopts "batch storage + timeout cleanup". C3 level shares 32 cache areas (512KB each), and adopts the "first-in, first-out" strategy to ensure that C1 level cache resources are not squeezed out.

[0235] (3) Priority optimization of data transmission links

[0236] Based on the aforementioned "terminal-LoRa gateway-edge node" communication architecture, priority transmission is achieved from three dimensions: link bandwidth allocation, transmission protocol optimization, and anomaly handling.

[0237] Dynamic bandwidth allocation: The total bandwidth of the LoRa gateway is set to 80kbps, using a "fixed guarantee + dynamic sharing" allocation mode—C1 level always occupies 30kbps of bandwidth to ensure a transmission channel is still available in extreme cases; C2 and C3 levels share the remaining 50kbps bandwidth. When the data volume of C2 level surges, C3 level bandwidth can be temporarily occupied (up to 20kbps), resulting in a delay in C3 level data transmission. For example, during peak commuting hours (8:00-9:00), when the data volume of C2 level air conditioning increases, the gateway automatically adjusts the bandwidth allocation, reducing the C2 level transmission latency from 200ms to 80ms.

[0238] Differentiated configuration of transmission protocols: C1 level data uses UDP protocol + custom check header. The check header includes terminal ID, timestamp, and data type (1-byte identifier). Only AES-256 encryption is performed before transmission, and no data compression is performed to ensure transmission speed; C2 level data uses UDP + LZ4 lightweight compression with a compression ratio of 3:1 to balance speed and bandwidth; C3 level data uses TCP protocol + ZIP compression with a compression ratio of 5:1 to ensure data transmission reliability and reduce bandwidth consumption.

[0239] Fault tolerance handling for transmission anomalies: When C1-level data transmission fails, "multi-link backup" is immediately triggered - first, retransmission is performed via LoRa (retransmission interval 10ms, up to 3 times). If retransmission fails, it automatically switches to the 4G module, with 4G transmission latency ≤100ms. After C2-level data transmission fails, it is retransmitted once within 500ms. If it fails, it is marked as "to be retransmitted" and retransmitted in batches when idle. C3-level data transmission failure is not retransmitted immediately. Retransmission is performed uniformly at 2:00 AM every day to reduce the interference of retransmission on the core link.

[0240] Step S35: Dynamic Optimization and Effect Evaluation of the Strategy

[0241] Establish a closed-loop optimization mechanism of "data feedback - parameter iteration - effect evaluation" to ensure that the strategy adapts to changes in load operation status in the long term, and verify the effectiveness of the strategy through quantitative indicators.

[0242] (1) Feedback data collection and analysis for strategy optimization

[0243] Edge nodes automatically perform data statistical analysis from 1:00 AM to 3:00 AM daily (off-peak load period) to generate a "Strategy Optimization Feedback Report." Core data includes:

[0244] Accuracy feedback data acquisition: Ten typical terminals from each of levels C1-C3 were selected. The acquired data was compared with the invasive measurement data (from the aforementioned Example 2). Power measurement error and event capture deviation were calculated. Terminals with errors exceeding the standard were marked as "to be calibrated". Power measurement error:

[0245]

[0246]

[0247] In the formula, Power is collected by the terminal. For invasive power measurement; Capture event timestamps for the terminal. This refers to the actual time the event occurred. When... or At that time, terminal parameter calibration is triggered.

[0248] Load characteristic change data: The average daily operating cycle, peak power, and fluctuation coefficient of various loads are statistically analyzed and compared with the initial characteristic database. For example, the average daily operating cycle of central air conditioning in summer is shortened from 300s to 220s, and the fluctuation coefficient is increased from 18% to 25%. These changes will serve as the core basis for optimizing the data collection mode.

[0249] Resource consumption statistics: Through the power metering module of the terminal energy storage root layer, CPU utilization, memory utilization, and bandwidth utilization are statistically analyzed through the system logs of the edge nodes. When the average daily energy consumption of the C3-level terminal is greater than 1kWh or the CPU utilization of the edge node is continuously greater than 75%, resource optimization is triggered.

[0250] (2) Implementation of a layered and progressive strategy

[0251] Based on feedback data, a layered optimization strategy of "terminal calibration - parameter update - mode optimization" is adopted. The optimization instructions are sent to the terminal through the edge node, and the execution process does not affect normal data acquisition.

[0252] Terminal hardware parameter calibration (error triggering): For terminals marked "to be calibrated", the edge node issues a calibration command, and the terminal automatically adjusts the gain resistor and offset resistor of the power metering chip. For example, if the power measurement error of a C1-level UPS terminal reaches 0.8%, after the calibration command is issued, the terminal adjusts the gain resistor from 10kΩ to 10.2kΩ and the offset resistor from 1kΩ to 0.98kΩ. After calibration, the error is reduced to 0.2%, meeting the accuracy requirements.

[0253] Dynamic parameter updates (daily): Edge nodes update the terminal's data collection threshold parameters based on load characteristic changes. For example, if the central air conditioning fluctuation coefficient rises to 25%, its "fluctuating operation" judgment threshold is adjusted from... >15% adjusted to >12%, and at the same time, the sampling frequency during the fluctuation period was increased from 5Hz to 6Hz to ensure that no fluctuation data was missed; for the winter electric heating load, based on historical data, its peak start time was predicted to be brought forward from 11:00 to 10:30, and the pre-trigger time was extended from 3 seconds to 5 seconds to improve the impact data capture rate.

[0254] Data Acquisition Mode Iterative Optimization (Executed Weekly): For scenarios where resource consumption exceeds limits, the data acquisition mode logic is optimized. For example, a C3-level lighting terminal has an average daily energy consumption of 1.2 kWh. Analysis of feedback data revealed that the terminal was still acquiring data at 0.5 Hz after the lights were turned off at night (22:00-6:00). After optimization, the terminal automatically switches to "sleep acquisition" mode at night, reducing the acquisition frequency to 0.01 Hz (acquiring data once every 100 seconds). The terminal's average daily energy consumption decreased to 0.7 kWh, a reduction of 41.7%.

[0255] (3) Multi-dimensional effect evaluation and verification

[0256] A three-dimensional evaluation system of "technology-economy-application" is established, with an evaluation cycle of 30 days. All evaluation data is obtained through the technical means described in the previous examples to ensure objectivity.

[0257] Technical Indicator Assessment: The core assessment indicators and results are as follows:

[0258] Critical event capture rate: 100% for C1 level load, 99.6% for C2 level, and 98.2% for C3 level, an improvement of 2.1-3.5 percentage points compared to before optimization;

[0259] Data transmission success rate: C1 level 100%, C2 level 99.8%, C3 level 99.5%, retransmission rate reduced by 60% compared to before optimization;

[0260] Resource utilization: The average CPU utilization of edge nodes decreased from 78% to 52%, and the average bandwidth utilization of LoRa gateways decreased from 85% to 62%.

[0261] Economic indicators assessment: The average monthly total energy consumption of 128 terminals decreased from 128kWh to 89.6kWh, which, based on an industrial electricity price of RMB 1.2 / kWh, resulted in an average monthly saving of RMB 46.08 in electricity costs; the communication traffic between edge nodes and the cloud decreased from 120GB to 54GB per month, resulting in an average monthly saving of RMB 408 in communication costs; and the terminal calibration cycle was extended from 3 months to 6 months, resulting in an annual saving of RMB 1200 in operation and maintenance costs, demonstrating significant economic benefits.

[0262] Application Value Assessment: The optimized collected data provides higher quality input for the edge computing-based non-intrusive load decomposition algorithm in the implementation example, improving the decomposition accuracy from 96.8% to 97.5%. Through precise data collection, it was discovered that a C2-level air conditioner was in an "inefficient operation" state (COP value dropped from 3.2 to 2.5). After the maintenance personnel cleaned the filter, the COP value recovered, resulting in an average monthly power saving of 320kWh, verifying the application value of the strategy.

[0263] (4) Mechanism for ensuring continuous optimization effect

[0264] To avoid parameter drift after optimization, a "real-time monitoring - threshold alarm - manual review" mechanism is established: edge nodes monitor the collection accuracy and resource utilization in real time. When the C1 level event capture rate is <99.8% or the CPU utilization is >80%, a local alarm is immediately triggered and pushed to the operation and maintenance platform; a "Review Report on Optimization Effect" is generated weekly, and operation and maintenance personnel randomly check the collection data of 10% of terminals to confirm the rationality of the optimization parameters; a full terminal calibration is performed quarterly to ensure the long-term stability and effectiveness of the strategy.

[0265] In summary, this embodiment solves the inherent defects of traditional acquisition modes by adopting a dual-dimensional dynamic acquisition strategy of "load type-operation status". Together with the bionic acquisition terminal and edge decomposition algorithm mentioned above, it forms a complete technology chain of "acquisition-processing-application". The amount of data collected is reduced by more than 55%, and the key data capture rate reaches 99.6%. It provides efficient and accurate data support for refined management of building energy and early warning of equipment operation and maintenance, and has broad prospects for engineering applications.

[0266] Example 4

[0267] like Figure 4 As shown, an embodiment of the present invention provides a real-time correlation data collection and visualization method for load consumption and carbon footprint. Based on the multi-source data collected in Embodiment 1, the load decomposition results in Embodiment 2, and the dynamic data collection strategy in Embodiment 3, it achieves a closed-loop process of "carbon emission factor database construction - energy consumption - real-time carbon emission conversion - multi-dimensional visualization of carbon footprint - optimization of low-carbon decision-making model - effect verification". Specifically, it includes the following steps:

[0268] Step S41: Construction and Dynamic Update of Carbon Emission Factor Database

[0269] This step provides benchmark data support for carbon footprint conversion. The database is built based on national authoritative standards, combined with dynamic optimization of regional energy characteristics, and forms a correlation mapping with the load characteristic database of the aforementioned load type perception dynamic data acquisition strategy.

[0270] (1) Construction of a multi-dimensional factor database

[0271] The database adopts a hierarchical structure of "master database - sub-database". The master database stores the basic factors of energy type, and the sub-databases are subdivided according to dimensions such as region and energy purity, and are associated with the eight load types mentioned above. The core content is as follows:

[0272] The basic energy factor database covers six types of energy commonly used in buildings: electricity, natural gas, liquefied petroleum gas, urban heating, diesel, and solar energy. The basic factor values ​​are derived from the "Guidelines for the Compilation of Provincial Greenhouse Gas Inventories (2025 Edition)" and the National Energy Administration's "China Energy Statistical Yearbook".

[0273] Regional sub-database: Based on power grid factors, subdivided according to the seven major regional power grids in China (e.g., North China Power Grid). Northwest Power Grid ); Regarding gas factors, they are further subdivided according to the city's gas purity level (e.g., in Nanjing, the natural gas purity is 99.2%, and the factor is taken as...). The detailed data is obtained by calling the API interface of the local ecological and environmental bureau.

[0274] Load association mapping table: Establish the association relationship of "load type-energy type-factor ID", such as central air conditioning (electric drive) is associated with the power factor ID of East China Power Grid, and gas water heater is associated with the natural gas factor ID of Nanjing area to ensure accurate factor matching during conversion.

[0275] (2) Database dynamic update mechanism

[0276] The update adopts a dual-drive mode of "automatic synchronization + manual verification":

[0277] Automatic synchronization and update: Edge nodes automatically access the Ministry of Ecology and Environment's "Greenhouse Gas Emission Factor Database" and local energy departments' API interfaces at 2:00 AM on the 1st of each month (during off-peak hours) to synchronize the latest factor data; when the change of a certain type of energy factor is greater than 5%, the main database / sub-database data is automatically updated, and "historical data backtracking correction" is triggered - recalculating the carbon emission data of the corresponding load of that energy in the past 3 months to ensure data consistency.

[0278] Manual custom updates: Supports maintenance personnel to enter region-specific factors (such as self-owned power plant power supply factors and photovoltaic self-generated and self-consumed power factors). After entry, it must be verified by "two people + system verification" (it can only be effective if the deviation from the national benchmark factor is ≤10%), which meets the personalized accounting needs of industrial parks, green buildings and other places.

[0279] Step S42: Real-time correlation and conversion of energy consumption and carbon emissions

[0280] Based on the high-precision energy consumption data (electrical parameter accuracy ±0.01kW) collected in Example 1, carbon emissions are calculated in real time through a dynamic conversion engine with a conversion delay of ≤100ms, completed synchronously with energy consumption data collection. Differentiated conversion formulas are designed for single-energy loads and mixed-energy loads, and the formula parameters are linked to those in the previous examples, such as the operating status correction coefficient. Load condition identification results from Example 3, ambient temperature correction factor Environmental parameter data collected from Example 1.

[0281] (1) Carbon emission conversion formula for multiple energy types

[0282] The method of "calculation by energy type + weighted summarization" is adopted, and conversion formulas are designed for single-energy-driven loads and mixed-energy-driven loads respectively. The formula parameter definitions fully cover the calculation logic.

[0283] 1) Carbon emission conversion for single energy load

[0284] This formula applies to loads that use only one energy source (such as LED lighting and electrically driven elevators).

[0285]

[0286] In the formula, Carbon dioxide emissions from a single energy load, in units of Q represents the real-time energy consumption of the load. The unit for electricity load is kWh, and the unit for gas load is MJ. The acquisition accuracy is ±0.01 kWh / MJ. The specific carbon emission factor for the corresponding energy, in units of or The data is obtained from the factor database through a three-dimensional matching process based on "load type - energy type - region", with the precision retained to four decimal places. This is a correction factor for load operating conditions during steady-state operation. =1.0, during startup / impact state =1.02 (considering slightly lower energy efficiency during startup), during shutdown. =0.98 (considering energy consumption attenuation during shutdown); The ambient temperature correction factor ranges from 0.95 to 1.05, divided according to the outdoor temperature range (1.05 for below -10℃, 1.0 for 25℃±5℃, and 0.95 for above 35℃), and is collected by the terminal's built-in temperature and humidity sensor.

[0287] 2) Carbon emission conversion of mixed energy loads

[0288] This formula is applicable to loads that utilize multiple energy sources (such as wind-solar-storage hybrid air conditioning).

[0289]

[0290] In the formula, n is the number of energy types used by the load (n≥2); The real-time energy consumption of the i-th energy source is expressed in kWh or MJ and is collected by the built-in energy metering module of the load. The specific carbon emission factor for the i-th energy source is given by [value], in units of [value]. or ; This is the correction coefficient for the operating state corresponding to the i-th energy source, and its value is determined according to the same rules as for single energy load; the definitions of the other parameters are the same as those for the single energy load conversion formula.

[0291] (2) Data association and quality control

[0292] Construct a five-dimensional data chain encompassing energy consumption, carbon emissions, load, space, and time to ensure data traceability and verifiability.

[0293] Multi-dimensional data association: Carbon emission data is bound to attributes such as "load ID, load type, operating status, spatial location (floor-area number-equipment coordinates), collection timestamp (accurate to milliseconds), factor ID, and correction coefficient" collected in Example 1 to form structured data records. Each record contains 32 data fields (covering raw energy consumption data, factor information, conversion process data, and verification results), and supports filtering and querying by any dimension.

[0294] Conversion quality control: A triple verification mechanism is set up: ① Range verification: An alarm is triggered when the carbon emission data deviates from the historical data of the same type of load by more than 30%; ② Logic verification: When the carbon emission intensity (carbon emission per unit energy consumption) under different conditions of the same load deviates by more than 5%, the factor matching and correction coefficient values ​​are automatically checked; ③ Accuracy verification: An invasive carbon emission meter (accuracy ±0.1%) is used to sample and verify the C1 level core load, with a sampling ratio of ≥5%, to ensure that the overall conversion error rate is ≤3%.

[0295] Step S43: Multi-dimensional visualization of carbon footprint

[0296] Based on a cloud-based visualization platform, building space information and load attributes are integrated, and the load priority classification in Example 3 is combined to achieve multi-dimensional and interactive display of carbon emission data.

[0297] (1) Spatial Dimension Visualization: Carbon Footprint Heat Map

[0298] Integrating architectural models and geographic information systems, a four-level spatial heat map of "building-floor-area-equipment" is constructed. Carbon emission intensity (unit: [missing information]) is indicated by a four-color gradient: red (high carbon emissions), orange (medium carbon emissions), yellow (low carbon emissions), and green (zero carbon emissions). ):

[0299] Tiered display logic: The first-level display (building overview) presents the carbon emission and thermal distribution of the entire building, with red areas (carbon emission intensity > 100%). The first level identifies high-carbon emission areas; the second level (floor drill-up) displays the carbon emission distribution of each functional area (office area, meeting room, computer room) on a specific floor; the third level (equipment drill-up) displays the carbon emission data, carbon emission intensity, and ranking of a single piece of equipment in that area, helping to locate high-carbon emission equipment.

[0300] Real-time dynamic updates: Heat map data is refreshed every 5 seconds, with the refresh frequency synchronized with the dynamic acquisition frequency of Example 3; when the carbon emission intensity of a certain area exceeds the threshold (office area > Computer room > When this occurs, the corresponding area on the heat map flashes an alarm, and information is pushed to the mobile devices of maintenance personnel.

[0301] (2) Visualization of the time dimension: trend and comparative analysis

[0302] It provides a full-time dimension display function of "real-time-historical-predictive" and supports data comparison across multiple time periods:

[0303] Real-time trend curve: Displays the changes in total building carbon emissions, carbon emissions in various regions, and core load carbon emissions over the current 24 hours. The horizontal axis represents time (accurate to the minute), and the vertical axis represents the carbon emission rate. The inflection point of the curve indicates the change in the load operating status (such as the sudden increase in carbon emissions caused by the air conditioner starting at 9:00).

[0304] Historical comparative analysis: Supports selecting "daily / weekly / monthly / yearly" time granularity to compare carbon emission data at different time periods. For example, comparing the carbon emission curves of Monday this week and Monday last week, automatically calculating the emission reduction ΔE=E 历史 -E 当前 and emission reduction rate η=ΔE / E 历史 ×100%; Supports overlaying the implementation timeline of energy-saving measures (such as "replace LED lighting on April 1st"), intuitively demonstrating the carbon emission reduction effect after the implementation of the measures.

[0305] Short-term forecasting function: Based on historical data of the past 7 days, the ARIMA time series model is used to predict the carbon emission trend in the next 24 hours, and the peak carbon emission period (error ≤ 30 minutes) and the corresponding load type are marked, providing advance notice for peak electricity consumption and carbon quota scheduling.

[0306] (3) Visualization of management dimensions: three-level drill-down reports

[0307] Generate standardized reports that meet carbon quota management requirements, supporting three-level drill-down queries of "total carbon emissions - regional carbon emissions - equipment carbon emissions". The report data is associated with the C1-C3 priority categories in Example 3.

[0308] Level 1 Report (Total Carbon Emission Statistics): Statistics on total building carbon emissions and the carbon emission ratio of various energy sources (e.g., electricity accounts for 78% and gas accounts for 22%) are compiled daily / monthly / yearly. The report automatically compares the carbon emission quota indicators. When the cumulative carbon emissions reach 80% of the quota, an early warning is triggered. When the cumulative carbon emissions reach 90%, the operation of high-carbon emission non-core loads is restricted.

[0309] Secondary Report (Regional Carbon Emission Statistics): Statistics on carbon emission data by region, and calculation of carbon emission intensity for each region. The data and percentage of carbon emissions were used to identify the top three regions with the highest carbon emission percentages, providing a basis for regional energy-saving renovations.

[0310] Level 3 Report (Equipment Carbon Emission Statistics): Carbon emissions of individual equipment are categorized by load priority. C1-level core equipment is reported separately, with equipment carbon emission efficiency (carbon emissions per unit output) clearly indicated. / unit output value), for example, the carbon emission efficiency of a C2-level air conditioner is Higher than the average value of similar equipment ( It is recommended to carry out energy efficiency improvement renovations.

[0311] Step S44: Optimization of Low-Carbon Decision Model

[0312] The maximum entropy inverse reinforcement learning algorithm is introduced, and the load operation and comfort data of Example 1, the load decomposition data of Example 2, and the carbon emission data of this example are combined to quantify the balance relationship between "comfort-low carbon-cost-stability" and output the optimal operation strategy.

[0313] (1) Definition of input and output of decision model

[0314] Input feature parameters include load operation parameters (air conditioning temperature setting, lighting brightness, equipment running time, frequency converter frequency), carbon emission parameters (real-time carbon emission rate, cumulative carbon emission, carbon emission intensity), comfort parameters (indoor temperature and humidity, illuminance, PM2.5 concentration, collected by the terminal's built-in sensor with an accuracy of ±0.1℃, ±5lux, ±1μg / m³), environmental parameters (outdoor temperature and humidity, wind speed, electricity price during the day), and carbon quota parameters (remaining quota, quota unit price), totaling 22 input features. The data sampling frequency is consistent with the acquisition frequency in Example 3.

[0315] Output optimization strategy: Specific control suggestions are output for different load types. Output parameters include target operating parameters (such as air conditioning temperature setting of 26℃ and lighting brightness of 80%), expected carbon emissions (such as a 12% reduction from the benchmark), comfort score, and changes in operating costs (such as an 8% reduction from the benchmark). The strategy can be sent to the load control terminal with one click.

[0316] (2) Implementation of maximum entropy inverse reinforcement learning algorithm

[0317] The algorithm is implemented through "behavioral data to deduce the reward function - iterative policy optimization", and the specific process is as follows:

[0318] Reward Function Construction: Based on historical operational data (12 months of load operation, carbon emissions, and comfort data, with a sample size of ≥100,000), the maximum entropy inverse reinforcement learning algorithm is used to infer users' implicit preferences, and the comprehensive reward function is constructed as follows:

[0319]

[0320] In the formula, R is the overall reward value, which ranges from 0 to 100. The larger the value, the better the strategy. For comfort reward, the formula is: ,in These are actual environmental parameters. The values ​​are national standard values, and m represents environmental parameter dimensions (temperature, humidity, illuminance, etc.).

[0321] The formula for low-carbon rewards is as follows: ,in For actual carbon emissions, Baseline carbon emissions (historical average for the same period);

[0322] As a penalty for operating costs, the formula is: ,in For actual operating costs, Base cost;

[0323] The reward for operational stability is calculated based on the fluctuation range of load operating parameters. 10 points are awarded when the fluctuation range is ≤5%, and 2 points are deducted for every additional 5%, with a minimum of 0 points.

[0324] α, β, γ, and δ are weighting coefficients that satisfy α+β+γ+δ=1. The default values ​​are α=0.4 (comfort), β=0.5 (low carbon), γ=0.05 (cost), and δ=0.05 (stability). Users can customize and adjust these values ​​according to their management needs.

[0325] Strategy Iteration Optimization: The strategy is updated and optimized using the policy gradient descent method with a learning rate of 0.001. The model parameters are iterated every 7 days using newly added running data. The convergence condition is that the fluctuation of the comprehensive reward value in 3 consecutive iterations is ≤1%.

[0326] (3) Decision support for carbon quota management and transformation

[0327] Based on model output and visualization analysis results, it provides accurate data support for carbon quota management and low-carbon transformation:

[0328] Dynamic carbon allowance allocation: Combining historical carbon emission data (average over the past 12 months) and forecast data (for the next 12 months) for each region and load, the entropy weight method is used to decompose the total building carbon allowance to the regional and equipment levels, generating a secondary carbon allowance list of "region-equipment", such as the monthly allowance for data centers. Office area monthly quota When the quota is exceeded, energy-saving suggestions will be automatically pushed (such as adjusting the air conditioning temperature and reducing the running time of non-core loads).

[0329] Prioritization of retrofits: based on equipment carbon emission efficiency ( The system prioritizes high-carbon emission equipment for retrofitting based on four indicators: unit output, years of operation, retrofit investment amount, and post-retrofit emission reduction potential. It outputs a "Low-Carbon Retrofit Recommendation Report" which includes parameters such as retrofit projects, expected emission reductions, investment amount, and investment payback period. For example, it recommends prioritizing the retrofit of central air conditioning systems that exceed carbon emission efficiency standards by 20%.

[0330] Step S45: Verification of Module Implementation Results

[0331] This module selects multiple typical building scenarios to conduct full-cycle performance verification. The verification scenarios cover medium and large office buildings of 10-20 stories and park-type complex buildings. The number of deployed terminals matches the building load scale (covering 8 core load types, with ≥100 data collection terminals deployed). The continuous operation cycle is no less than 12 months. Combining the dynamic data collection strategy evaluation method of Example 3, a standardized verification system is established from three aspects: technical accuracy, economic benefits, and low-carbon benefits. The specific verification content and indicators are as follows:

[0332] Technical accuracy verification: A dual-benchmark verification system of "invasive carbon emission meter + manual calculation" is adopted, with verification standards set according to load priority: C1 level core load sampling ratio ≥10%, carbon emission error rate ≤2.5%; C2 level important load sampling ratio ≥8%, carbon emission error rate ≤3.0%; C3 level general load sampling ratio ≥5%, carbon emission error rate ≤5.0%; overall monthly average total carbon emission error rate ≤3.0%. Simultaneously, data transmission and display performance are verified, with carbon emission data update latency ≤100ms and heat map data refresh synchronization ≤5 seconds, both meeting the preset technical accuracy requirements. Furthermore, the verification results comply with the relevant provisions of GB / T32150-2015 regarding the accuracy of carbon emission data calculation.

[0333] Economic and Low-Carbon Benefit Verification: Using a traditional energy management model without this module as a benchmark, the full-cycle benefits are calculated through standardized statistical methods: Economically, relying on the optimization strategy of the low-carbon decision-making model, a building's overall energy saving rate of ≥10%, a reduction in operation and maintenance costs of ≥15%, and a reduction in the frequency of ineffective operation and maintenance of core high-energy-consuming equipment of ≥40% can be achieved; Low-carbonly, through precise carbon emission control and equipment upgrade optimization, the annual carbon emission reduction per unit building area is ≥ It can reliably exceed the preset annual carbon quota target, and the carbon emission data can be directly connected to the carbon trading platform reporting system, providing compliant data support for buildings to participate in carbon market trading and obtain carbon asset benefits.

[0334] This embodiment, through deep collaboration with Embodiments 1, 2, and 3, achieves real-time correlation and precise management of load energy consumption and carbon emission data, solving the problems of "data lag, rough accounting, and blind decision-making" in traditional building carbon emission management. It is the final application outlet of the entire technology system, providing efficient data support for building carbon quota management, low-carbon transformation, and carbon trading.

[0335] Beneficial effects: The methods and modules for building load data acquisition, decomposition, strategy optimization and carbon emission control provided by this invention construct a complete technical system of "data acquisition - load analysis - strategy control - carbon emission accounting", which solves the core problems of low data accuracy, ambiguous load identification, poor acquisition efficiency and disconnection of carbon emission control in traditional building energy management, and provides comprehensive technical support for efficient building energy utilization and low-carbon operation.

[0336] In terms of multi-source load data acquisition methods, a biomimetic structural coupling design is used to integrate multi-dimensional acquisition units by simulating a biological sensing and collaborative mechanism. This achieves high-precision capture of load parameters such as electricity and gas. The accuracy of electricity parameter acquisition reaches ±0.01kW, reducing the error by more than 60% compared to traditional single acquisition methods. In addition, the anti-interference characteristics of the biomimetic structure enable the acquisition unit to adapt to complex environments such as voltage fluctuations and electromagnetic interference within buildings. It maintains stable operation even under extreme conditions such as high temperature and humidity, and the data transmission success rate is increased to 99.8%, solving the problem of easy failure of traditional acquisition equipment in complex scenarios. At the same time, load operating parameters are integrated with environmental temperature, humidity, illuminance and other related data to form a multi-dimensional data acquisition matrix, breaking through the limitation of traditional acquisition that only focuses on energy consumption data, and providing data support for accurate judgment of load status and comfort optimization.

[0337] In terms of load identification efficiency and deployment, relying on the edge computing architecture to achieve localized data processing, the decomposition response latency is controlled within 500ms, which is 80% more efficient than centralized cloud processing. Through feature extraction and pattern matching optimization, the identification accuracy of 8 typical loads such as air conditioning and lighting reaches over 95%, resolving the contradiction between the high cost of traditional intrusive decomposition and the low accuracy of non-intrusive decomposition. The non-intrusive design eliminates the need to modify existing load loops, reducing installation costs by 70%. At the same time, the distributed deployment of edge nodes reduces the data transmission and storage pressure on the central server, reducing maintenance workload by 50% and significantly improving the feasibility of technology implementation. In addition, the real-time data processing capability of edge nodes enables the load decomposition results to be fed back to the control system in an instant, providing millisecond-level response support for the dynamic adjustment of load operation status, breaking down the time barrier between data processing and control execution.

[0338] Regarding the dynamic load type acquisition strategy, the acquisition frequency and accuracy are dynamically adjusted based on the load type perception results. High-frequency, high-precision acquisition is used for core loads such as central air conditioning, while low-frequency acquisition is used for standby loads. While ensuring the quality of key data, the data transmission volume is reduced by 40%, thus reducing network bandwidth usage and energy consumption. A correlation model between load type and acquisition parameters is constructed. When a new load type is added, the system can automatically match the optimal acquisition strategy without manual reconfiguration, adapting to the dynamic changes in building load types and improving the system's scalability and intelligence. The acquisition strategy is linked with load control requirements, prioritizing the processing of data collected from high-energy-consuming loads to provide targeted data support for the generation of energy-saving control instructions. This improves the targeting of control measures by 50% and promotes the optimization of energy utilization efficiency.

[0339] Regarding the real-time correlation between load consumption and carbon footprint, it achieves real-time correlation between energy consumption data and carbon emission accounting, with a conversion delay of ≤100ms and an annual carbon emission data error rate of ≤3%, meeting the GB / T32150-2015 carbon trading data reporting standard. This solves the problems of lagging traditional carbon emission accounting and low data reliability, and constructs a closed-loop management system of "energy consumption collection - real-time carbon calculation - data traceability". It integrates BIM and GIS technologies to construct a four-level carbon emission heat map, combined with time-dimensional trend analysis and three-level drill-down reports, making carbon emission data intuitive and traceable. Based on the optimization strategy output by the maximum entropy inverse reinforcement learning model, it improves the comprehensive score of "comfort - carbon emission" of air conditioning systems, providing a scientific basis for low-carbon decision-making. Through dynamic allocation of carbon quotas and priority ranking of high-carbon emission equipment transformation, it takes into account both environmental and economic benefits, providing strong support for the implementation of energy conservation and carbon reduction goals in the construction industry.

[0340] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A system for collecting energy consumption data within a building, characterized in that, The system includes a plant-inspired vascular structure acquisition unit, a terminal distributed deployment and network networking module, a multi-source data collaborative acquisition and real-time preprocessing module, an edge node data processing and cloud platform linkage application module, and an anomaly early warning and full-cycle data traceability module. The acquisition unit adopts a four-layer biomimetic structure consisting of a protective epidermis, a biomimetic conduction layer, a core sensing layer, and an energy storage root layer. The energy storage root layer integrates a thermoelectric power generation module, a supercapacitor, and a backup power supply, utilizing the temperature difference between the equipment and the environment to achieve self-powering. The core sensing layer integrates a triple acquisition unit for electrical parameters, environmental parameters, and equipment status parameters, enabling synchronous acquisition of multi-source parameters.

2. The system according to claim 1, characterized in that, The multi-source data collaborative acquisition and real-time preprocessing module adopts a dual-mode acquisition strategy of power-level acquisition + event triggering. Differentiated normal acquisition cycles are set according to the load power level. When equipment vibration or temperature exceeds the preset threshold, the high-frequency acquisition mode is automatically triggered. After the acquired data is preprocessed by Kalman filtering and moving average method, abnormal data is marked and a priority transmission mechanism is adopted, while normal data is transmitted in batches with encryption.

3. A non-intrusive load decomposition method based on edge computing, applied to the building load energy consumption acquisition system of claim 1, characterized in that, Includes the following steps: Data preprocessing includes data decryption and format alignment, outlier correction and noise suppression, and data normalization; The format alignment formula is: , The sampling interval is used; outlier detection is performed using... Criteria, formula is Outliers are corrected using linear interpolation, as shown in the formula: ; Multi-dimensional feature extraction and dimensionality reduction: calculate time-domain, frequency-domain and coupling features, form a 22-dimensional feature vector, and then reduce the dimensionality to 8-dimensional through PCA; The reduced feature vectors are input into the lightweight ATCN network for load decomposition, and the power prediction values ​​of each independent load are output. The decomposition results are optimized by anomaly correction and power balance correction to obtain the final decomposition value. The anomaly correction formula is as follows: ,when When, the corrected formula is The power balance correction formula is: , ,in As power factor weights, ensure that after correction ; Performance verification and output.

4. The method according to claim 3, characterized in that, The lightweight ATCN network includes an input layer, a deep convolutional layer, an attention enhancement layer, and an output layer. The deep convolutional layer uses dilated causal convolution with a kernel size k=3, dilation coefficient d=[1,2,4], number of kernels K=16, and GELU activation function. The attention enhancement layer uses a channel attention mechanism, and the weight calculation formula is as follows: ,in L=60 is the length of the input sequence.

5. A dynamic data acquisition strategy method for load type sensing, characterized in that, Includes the following steps: Construct a multi-dimensional feature database covering all types of building loads, storing core feature parameters such as rated load power, starting power threshold, and steady-state fluctuation range. The database supports dynamic updates. A three-tiered mechanism—feature extraction, similarity matching, and status determination—is used to identify load type and operating status in real time. Similarity matching employs a cosine similarity algorithm, with the formula: ; Based on the identification results of load type and operating status, dynamically match differentiated data acquisition modes; Optimize the scheduling of acquisition resources and data transmission links based on load priority; Data feedback enables dynamic iterative optimization and effectiveness evaluation of the data collection strategy.

6. The method according to claim 5, characterized in that, The differentiated acquisition modes include event triggering + high-frequency supplementary acquisition + steady-state frequency reduction mode for intermittent high-fluctuation loads, fluctuation coefficient adaptive mode for continuous stable loads, and historical prediction + pre-triggering + full-process high-frequency mode for impulsive loads. The load priority is divided into three levels: C1, C2, and C3, which can be dynamically adjusted according to the device operating status. The priority is permanently bound to the physical ID of the acquisition terminal, and the core load occupies dedicated acquisition and transmission resources.

7. A method for real-time correlation and visualization of load consumption and carbon footprint, characterized in that, Includes the following steps: A carbon emission factor database with a master-sub-database hierarchical structure is constructed. The master database stores the basic carbon emission factors of commonly used building energy sources, while the sub-databases are subdivided by region and energy purity. The database is associated with building load type and supports dual-drive updates of automatic synchronization and manual verification. Based on high-precision energy consumption data of building load, combined with operation status correction coefficient and ambient temperature correction coefficient, real-time correlation conversion of energy consumption and carbon emissions is realized. By integrating building space information and load attributes, a multi-dimensional visualization of carbon footprint can be achieved; A low-carbon decision-making model is constructed by introducing the maximum entropy inverse reinforcement learning algorithm, and the output load operation optimization strategy takes into account multiple objectives. The effectiveness of the module implementation was verified from multiple dimensions, including technical precision, economic benefits, and low-carbon benefits.

8. The method according to claim 7, characterized in that, The real-time energy consumption-carbon emission correlation conversion includes conversion between single energy load and mixed energy load; the formula for single energy load conversion is as follows: Where Q is the real-time energy consumption. As a carbon emission factor, The operating condition correction factor is 1.0 for steady state, 1.02 for startup / shock, and 0.98 for shutdown. The ambient temperature correction factor is 1.05 for temperatures below -10℃, 1.0 for temperatures between 25℃ and 5℃, and 0.95 for temperatures above 35℃; the formula for converting mixed energy load is as follows. .

9. The method according to claim 7, characterized in that, The low-carbon decision-making model takes load operation parameters, carbon emission parameters, comfort parameters, environmental parameters, and carbon quota parameters as inputs, and uses the maximum entropy inverse reinforcement learning algorithm to construct a comprehensive reward function. The formula for the comprehensive reward function is as follows: α, β, γ, and δ are weighting coefficients that satisfy α+β+γ+δ=1. The default values ​​are α=0.4 (comfort), β=0.5 (low carbon), γ=0.05 (cost), and δ=0.05 (stability). Users can customize and adjust these values ​​according to their management needs. For comfort reward, the formula is: , The formula for low-carbon rewards is as follows: , As a penalty for operating costs, the formula is: .

10. The method according to claim 7, characterized in that, The multi-dimensional visualization of carbon footprint includes a four-level spatial carbon footprint heat map of building, floor, area, and equipment; real-time, historical, and predictive full-time trend and comparative analysis; and a three-level drilling management report of total carbon emissions, regional carbon emissions, and equipment carbon emissions. The heat map is color-coded according to carbon emission intensity and refreshed in real time. The trend analysis uses the ARIMA model to achieve short-term carbon emission prediction. The management report is linked to load priority and connected to carbon quota management requirements.

Images