A three-level box coordinated non-intrusive load decomposition method and system

Abstract

The application discloses a kind of three-level box body coordinated non-intrusive load decomposition method and system, it is related to electric power load monitoring technical field;The method is collected voltage current waveform at the electric power total entrance of three-level box body of branch box, distribution box and metering box in low-voltage distribution network, branch box extracts power time-domain characteristics, uses weighted classification to determine each distribution box load type, encapsulation is under Dirichlet priori;Distribution box uses Bayesian online change point detection to detect equipment start-stop event, combined with priori, equipment type matching is generated, and start-stop label is issued;Metering box uses Seq2Point CNN model to carry out fine decomposition, and distribution box label is used as confidence enhancement signal to improve recognition reliability;Three-level model is trained through federal distillation cooperation, only soft label vector is transmitted, and original data does not come out metering box;The application significantly improves non-intrusive load decomposition precision, while protecting user electricity privacy.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring technology, specifically to a non-intrusive load decomposition method and system with three-level box coordination. Background Technology

[0002] Non-intrusive load monitoring (NILM) technology collects total voltage and current waveforms only at the power inlet. It leverages the unique electrical characteristics of each electrical device (steady-state power amplitude, transient starting current surge, harmonic distortion content, VI trajectory morphology, etc.) to decompose the operating status and power consumption of each device. This eliminates the need for sensors at each device, offering advantages such as low hardware cost, convenient installation, and high user acceptance. In recent years, with the development of deep learning technology, end-to-end NILM models, represented by Seq2Point CNN, have significantly improved device identification accuracy. Existing NILM technologies include the following three types:

[0003] The first approach is a single-household NILM (Non-Insulated Metering) solution based on deep learning. This solution installs a high-frequency sampling device (1kHz or higher) at the user's main inlet line to collect voltage and current waveform data. A CNN-LSTM hybrid deep learning model is then used to decompose the load of each electrical device within the user's premises, identifying the power consumption and operating time of each device. The advantage of this approach is that it eliminates the need for in-house sensor installation. However, its disadvantage is that the NILM model is trained only on single-household data; one metering box covering six households requires six independent sampling hardware sets and six independent models. It does not utilize the load characteristics and correlations of multiple users within the same box, nor does it leverage prior load composition information provided by the upper-level metering unit to assist in the decomposition.

[0004] The second approach is a multi-level load identification scheme for distribution networks in transformer substations. This scheme deploys load monitoring terminals with different sampling frequencies at three levels: transformer substations, branch boxes, and user meters. At the substation level, low-frequency sampling (15 minutes) is used for load classification; at the branch box level, medium-frequency sampling (1 minute) is used to detect equipment start-up and shutdown; and at the user level, high-frequency sampling (1 kHz) is used for fine-grained identification. The identification results from all three levels are uploaded to the cloud for data fusion. The advantage of this scheme is that it achieves multi-level deployment; the disadvantage is that the relationship between the three levels is only one of data aggregation, lacking a cross-level transfer mechanism for prior knowledge—the coarse classification results from the branch boxes are not used as Bayesian prior constraints for equipment matching in the distribution boxes, and the equipment start-up and shutdown tags detected by the distribution boxes are not used as enhancement signals for CNN inference in the metering boxes.

[0005] The third approach is a multi-user NILM collaborative training scheme based on federated learning. This scheme utilizes the FedAvg algorithm in federated learning, where multiple user terminals train their own local NILM models. The model parameters are then encrypted and uploaded to a central server for parameter averaging and aggregation. The advantage of this scheme is that it protects user privacy (raw electricity consumption data remains within the user's home). The disadvantage is that it adopts a flat federated architecture, with all clients participating equally in parameter averaging, failing to consider the inherent three-tiered hierarchy in the distribution network—users within the same distribution box have higher similarity in electricity consumption characteristics than users in different distribution boxes. Furthermore, the model structures at each level are heterogeneous (classification models for branch boxes, change point detection for distribution boxes, and CNNs for metering boxes), with inconsistent parameter dimensions, making it impossible to directly use FedAvg. Summary of the Invention

[0006] This invention provides a non-intrusive load decomposition method and system with three-level box coordination. It utilizes the three-level topology of the distribution network, namely branch boxes, distribution boxes, and metering boxes, to establish a progressive decomposition accuracy improvement path, which significantly improves the decomposition accuracy of non-intrusive loads.

[0007] This invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a non-invasive load decomposition method for three-level enclosure coordination, comprising the following steps:

[0009] Step 1: Collect three-phase voltage and current waveform data at the main incoming line of the branch box, calculate the total three-phase active power sequence, extract the time-domain features of the power sequence, and use weighted classification rules based on the time-domain features to determine the load type of each subordinate distribution box. Encapsulate the classification results into Dirichlet prior distribution parameters and send them to the corresponding distribution box gateway.

[0010] Step 2: The distribution box collects the total active power sequence at the main incoming line, receives the Dirichlet prior distribution parameters sent by the branch box, uses the Bayesian online change point detection algorithm to detect abrupt changes in the total power sequence, determines the equipment start event based on the power change before and after the abrupt change point, and performs equipment type matching for the start event by combining the Dirichlet prior distribution parameters and the equipment power fingerprint database, generates equipment start / stop tags and sends them to the subordinate metering box gateway.

[0011] Step 3: The metering box collects voltage and current waveform data at the main incoming line, calculates the total active power sequence, extracts the power sequence with a sliding window and inputs it into the Seq2Point CNN model, outputs the operating status probability of each device, receives the device start / stop tags issued by the distribution box as confidence enhancement signals to enhance the CNN output, and determines the device operating status and estimates the device power based on the enhanced confidence.

[0012] Secondly, the present invention provides a three-level box-coordinated non-intrusive load decomposition system, comprising:

[0013] The sensing hardware layer is deployed at the main power inlet of the three-level enclosure, including a high-frequency sampling module for branch boxes, a medium-frequency sampling module for distribution boxes, and a high-frequency sampling module for metering boxes, which are used to collect voltage and current waveform data of the main incoming line of each level enclosure;

[0014] The edge computing layer comprises edge computing gateways deployed in the three-level enclosure, including branch box edge computing gateways, distribution box edge computing gateways, and metering box edge computing gateways. The branch box edge computing gateway includes a three-phase feature extraction module, a coarse-grained load classification module, and a prior knowledge encapsulation module, used to perform the operation in step 1. The distribution box edge computing gateway includes a prior knowledge receiving module, a BOCPD change point detection module, and a device matching module, used to perform the operation in step 2. The metering box edge computing gateway includes a tag receiving module, a Seq2Point CNN inference module, and a device fingerprint matching module, used to perform the operation in step 3.

[0015] The cloud collaboration layer communicates with the edge computing layer and is used to generate load breakdown reports.

[0016] The present invention has the following beneficial effects:

[0017] 1. This invention establishes a three-level progressive NILM architecture: "coarse classification of branch boxes → detection in distribution boxes → fine decomposition of metering boxes." All sensing is completed at the main inlet of each level of box, eliminating the need to install sensors in subordinate units or equipment. The core of this architecture lies in the fact that the decomposition result of each level is not used as the endpoint, but rather as a Bayesian prior injection for the next level—the load type classification result of the branch box is encapsulated as a Dirichlet prior, constraining the candidate space for matching equipment in the distribution box; the start / stop label of the distribution box serves as a confidence enhancement signal during CNN inference in the metering box, improving the confidence of air conditioner identification at that moment. Simultaneously, the new equipment types identified by the lower-level VI trajectory matching are fed back to update the fingerprint database of the upper level, forming a closed loop of "prior injection → fine decomposition → fingerprint backhaul → prior update."

[0018] 2. This invention establishes an electrical feature fingerprint system covering four dimensions: steady-state features, transient features, harmonic features, and VI trajectory features. The metering box level extracts the above features only through the V / I waveform of the main incoming line, without the need to install sensors at the device end, thus achieving completely non-invasive simultaneous identification of multiple devices.

[0019] 3. This invention addresses the heterogeneous structure of the three-level box-type NILM model by designing a federated distillation method based on temperature-scaling soft labels. It replaces the tens of thousands of parameters of the CNN model or the original high-frequency sampled data transmitted in the network with 10-dimensional soft-label vectors, solving the problem that traditional FedAvg parameter averaging is unsuitable for heterogeneous models. The three-level distillation architecture, consisting of "local training in the metering box → local aggregation in the distribution box → global distillation in the cloud," fully leverages the natural hierarchical advantage of "similar user equipment usage habits within the same distribution box." Attached Figure Description

[0020] Figure 1 This is a flowchart of the non-invasive load decomposition method in this invention.

[0021] Figure 2 This is an architecture diagram of the non-intrusive load decomposition system in this invention.

[0022] Figure 3 This is a bar chart showing the global model accuracy at the branch box level, distribution box level, and metering box level in Example 3.

[0023] Figure 4 This is a bar chart showing the detection accuracy of high-power equipment at the distribution box level in Example 3, with and without prior constraints.

[0024] Figure 5 This is a bar chart showing the accuracy of device identification at the metering box level under single-level NILM and three-level combined NILM conditions in Example 3.

[0025] Figure 6 This is a convergence curve of the federal distillation co-training in Example 3.

[0026] Figure 7 This is a bar chart comparing the communication efficiency for privacy protection in Example 3.

[0027] Figure 8 This is a pie chart showing the breakdown of the electrical equipment used by user #1 for 24 hours in Example 3. Detailed Implementation

[0028] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0029] To facilitate understanding of this invention, the following terms are now explained.

[0030] NILM: Non-invasive load monitoring technology.

[0031] Seq2Point CNN: is a deep learning architecture that maps input sequences to a single scalar value. It is mainly used in scenarios such as non-intrusive load monitoring. In this invention, it is simply referred to as "CNN".

[0032] BOCPD: Bayesian online change point detection algorithm.

[0033] DTW: Dynamic Time Warping Algorithm.

[0034] FedAvg: Federated Averaging is the core aggregation algorithm in federated learning. It achieves collaborative modeling without data leaving the domain through multiple rounds of "local training + weighted averaging", aiming to significantly reduce communication overhead.

[0035] Example 1

[0036] Please see Figure 1 As shown, a non-intrusive load decomposition method for three-level box coordination includes:

[0037] Step 1: Branch Box Level Coarse-grained Load Type Classification. Three-phase voltage and current waveform data are collected at the main incoming line of the branch box. The total three-phase active power sequence is calculated, and the time-domain characteristics of the power sequence are extracted. Based on these time-domain characteristics, a weighted classification rule is used to determine the load type of each subordinate distribution box. The classification results are encapsulated as Dirichlet prior distribution parameters and sent to the corresponding distribution box gateway.

[0038] Specifically:

[0039] A high-frequency sampling unit (sampling frequency not less than 1kHz) is deployed at the main incoming line of the branch box to synchronously collect three-phase voltage and current waveform data. The total three-phase active power sequence is calculated with a period of 15 minutes.

[0040] Four temporal features of the power sequence were extracted: daytime power ratio. (Reflecting daytime electricity consumption intensity), peak-to-valley ratio (Reflecting the degree of load fluctuation), coefficient of variation (Reflecting load stability), evening power ratio (Reflecting evening electricity consumption intensity). Based on the above four characteristics, a weighted classification rule is used to determine the load type of each subordinate distribution box:

[0041] ;

[0042] in , For feature set, For type In features The typical range of values ​​is shown above. For example, the residential load exhibits a bimodal characteristic (morning and evening). Commercial load is high during the day and low at night. Public facilities remained stable throughout the day. ), industrial load working time is concentrated ( ); It refers to the weights of the "weighted classification rule", i.e., the features. The weight values, This indicates an indicator function, where the condition in the square brackets is true. When the conditions are not met .

[0043] The classification results are encapsulated as Dirichlet prior distribution parameters:

[0044] The concentration parameter for the dominant type (the type with the highest confidence level) is set to... ( (The confidence level is set as the dominant type), while other type concentration parameters are set as baseline values. 0. The prior distribution parameters are sent to the corresponding distribution box gateway via the MQTT protocol.

[0045] Step 2: High-power equipment start-up and shutdown event detection guided by Bayesian prior at the distribution box level. The distribution box collects the total active power sequence at the main incoming line, receives the Dirichlet prior distribution parameters from the branch boxes, and uses the Bayesian online change point detection algorithm to detect abrupt changes in the total power sequence. Based on the power change before and after the abrupt change, the distribution box determines the equipment start-up event. Combining the Dirichlet prior distribution parameters and the equipment power fingerprint database, the distribution box performs equipment type matching for the start-up event, generates equipment start-up and shutdown tags, and sends them to the subordinate metering box gateway.

[0046] Specifically:

[0047] The distribution box is equipped with an intermediate frequency sampling unit (sampling frequency not less than 1kHz) at the main incoming line to calculate the total active power sequence with a period of 1 minute. Receive Dirichlet prior distribution parameters from the branch box. .

[0048] The Bayesian Online Change Point Detection (BOCPD) algorithm is used to detect equipment start-up and shutdown events. The running length is defined. Let be the number of consecutive time steps since the last point of change. The recursive formula for its joint posterior probability is:

[0049] ;

[0050] in for The total active power observed at time t. The hazard function at the point of change uses a constant hazard rate. The runtime transition probability is:

[0051] ;

[0052] in This is the risk function for the point of change.

[0053] The predicted distribution of the data for a given runtime adopts a Gaussian distribution. ,parameter and By running length Estimation of the corresponding historical data segment.

[0054] When the posterior probability of the point of change Exceeding the threshold At that time, the judgment The time point is the point of change. Calculate the power change before and after the point of change. ( and These are the averages of 10 sampling points before and after the point of change, respectively. and When this occurs, it is determined to be a device startup event.

[0055] For each startup event, device type matching is performed using the branch box prior distribution parameters and the device power fingerprint database. The matching score for each candidate device consists of two parts:

[0056] ;

[0057] in For equipment The median of the rated power range (e.g., 2000W for an air conditioner). This indexes the load type (residential, commercial, utility, industrial) corresponding to the device in the Dirichlet prior. Weights are assigned... Select the candidate with the highest overall score that exceeds the threshold. The device type is used as the matching result. Device start / stop tags are generated. It is sent to the subordinate metering box gateway via the MQTT protocol.

[0058] Step 3: Meter Box-Level Seq2Point CNN Fine-Grade Equipment Decomposition. The meter box collects voltage and current waveform data at the main incoming line, calculates the total active power sequence, extracts the power sequence using a sliding window, inputs it into the Seq2Point CNN model, outputs the operating state probability of each device, receives the device start / stop tags issued by the distribution box as confidence enhancement signals to enhance the CNN output, and determines the device operating state and estimates the device power based on the enhanced confidence.

[0059] Specifically:

[0060] The metering box deploys a high-frequency sampling unit (sampling frequency not less than 1kHz) at the main incoming line. Total active power is calculated with a period of 1 second. At the current sampling time Centered on, extract length is The power sequence of a sliding window at each sampling point (with the center point of the window being the target prediction time):

[0061] ;

[0062] Will Input a Seq2Point CNN model. The specific structure of the CNN model is as follows:

[0063] Convolutional layer 1: Conv1D, 30 filters, kernel length 10, stride 1, activation function ReLU;

[0064] Convolutional layer 2: Conv1D, 30 filters, kernel length 8, stride 1, activation function ReLU;

[0065] Convolutional layer 3: Conv1D, 40 filters, kernel length 6, stride 1, activation function ReLU;

[0066] Convolutional layer 4: Conv1D, 50 filters, kernel length 5, stride 1, activation function ReLU;

[0067] Convolutional layer 5: Conv1D, 50 filters, kernel length 5, stride 1, activation function ReLU;

[0068] Global Average Pooling;

[0069] Fully connected layer: 128 neurons, ReLU activation function, Dropout rate 0.3;

[0070] Output layer: neurons ( (where is the number of device categories), the activation function is Sigmoid, and each neuron outputs the probability that the corresponding device is in a running state.

[0071] CNN output vector middle, Indicates the first The probability that a device of this type is in operation at the current moment.

[0072] The core idea of ​​the Seq2Point strategy is to use the center point of the sliding window as the prediction target, with one prediction result per window. Compared to Seq2Seq (window → sequence), Seq2Point avoids the accumulation of errors between adjacent time steps in the output sequence, resulting in more stable training.

[0073] The time when the equipment start / stop label is issued to the distribution box Enhance the output confidence of the CNN corresponding device by utilizing the labeled signal:

[0074] ;

[0075] in The label enhancement factor. For the indicator function (tag device type and current device) (1 for a match, 0 otherwise). For unlabeled moments, the raw CNN output is used directly.

[0076] when At that time, the judgment of the first The device is in an ON state. The power estimate for an ON device is the median of its steady-state power range multiplied by the enhanced confidence level.

[0077] ;

[0078] in, The power estimate for the operating equipment. and For the first The range of steady-state active power recorded in the fingerprint database for this type of device.

[0079] For the first occurrence of an unknown device launch event whose type cannot be determined by CNN ( Significant but all (All values ​​are below 0.5), triggering the non-intrusive VI trajectory recognition process. Normalized VI trajectory features are extracted from the voltage and current waveforms of 5 power frequency cycles before and after this moment (a total of 200 sampling points). The trajectory feature vector is obtained by resampling the VI scatter points at 64 equal intervals with voltage as the horizontal axis and current as the vertical axis. The Dynamic Time Warping (DTW) algorithm is used to match the VI trajectory templates of all devices in the pre-built fingerprint database one by one.

[0080] ;

[0081] in To align the paths, boundary conditions, monotonicity, and continuity constraints must be met. The device type with the smallest DTW distance is selected as the identification result. When the minimum DTW distance is lower than the matching threshold, the unknown device is registered in the local fingerprint database to enrich the local device type coverage.

[0082] Step 4: Three-level federated distillation collaborative training. The metering box trains a Seq2Point CNN model using local data, and the model output is temperature-scaled to generate soft tags, which are then uploaded to the corresponding distribution box. The distribution box collects the soft tags from all its subordinate metering boxes, performs a weighted average aggregation, and uploads them to the cloud. The cloud collects the aggregated soft tags from all distribution boxes, calculates global soft tags, and distributes them to the gateways at each level of the distribution box. In the next round of training, the local models at each level are optimized by adding a KL divergence distillation loss term to the loss function.

[0083] The goal of federated distillation collaborative training is to continuously optimize the decomposition accuracy of each level of the NILM model using a three-level hierarchical structure, without requiring the original high-frequency sampling data from the metrology chamber to leave the local machine. The training process consists of two phases: bottom-up soft-label uploading and top-down global model distribution.

[0084] Specifically:

[0085] Step 4.1, Local Training of the Measurement Box and Generation of Soft Labels:

[0086] Metering box A Seq2Point CNN model is trained using locally acquired high-frequency power sequences and corresponding device status labeling data (from occasional user feedback or labels issued by the distribution box). After each training batch, the model outputs a logits vector for a batch of window samples. Generating soft tags through temperature scaling:

[0087] ;

[0088] in Number of equipment categories; soft labels It's a normalized probability vector that retains richer information about inter-category relationships compared to hard labels (one-hot vectors). For example, in soft labels, "fixed-frequency air conditioner" and "electric heater" are both temperature-controlled devices, and their probability values ​​are quite similar, reflecting their similarity in electricity consumption characteristics. This kind of inter-category relationship knowledge cannot be conveyed by hard labels alone. The size of a soft label vector is only... byte( (Only 40 bytes at a time) are uploaded to the gateway of the corresponding distribution box.

[0089] Step 4.2, partial aggregation of the distribution box:

[0090] Distribution box Collect soft tags from all its subordinate metering boxes and their respective training sample sizes Perform weighted average aggregation:

[0091] ;

[0092] Aggregation weights are directly proportional to sample size—larger bins (with more diverse equipment and abundant labeled data) contribute more to the aggregation results. (The aggregated weights are then...) This reflects the statistical characteristics of equipment usage in the area covered by the distribution box (e.g., a single building). The distribution box gateway will... Uploaded to the cloud.

[0093] Step 4.3, Cloud-based global distillation and model deployment:

[0094] Cloud-based aggregation of soft tags from all distribution boxes and its total sample size Calculate global soft tags:

[0095] .

[0096] Global soft tags The data is distributed to gateways at all levels of the enclosure. In the next training round, each local model adds a KL divergence distillation loss term to the original loss function.

[0097] ;

[0098] in The cross-entropy loss is for binary classification (calculated independently for each device). Let KL divergence be the KL divergence. For distillation loss weighting, This is the temperature squared scaling factor (used to compensate for the gradient shrinkage caused by temperature scaling).

[0099] The core advantages of federated distillation are: (1) It only transmits 10-dimensional soft label vectors (40 bytes), instead of tens of thousands of parameters of CNN models, reducing communication by more than 60%; (2) The transmission of soft labels does not involve the original electricity consumption data, fundamentally protecting user privacy; (3) KL divergence distillation can achieve collaborative optimization of heterogeneous models (different structures of models at each level), solving the problem that the averaging of traditional FedAvg parameters is not applicable; (4) Users in the same distribution box have similar usage habits, and local aggregation makes full use of the prior information at this level.

[0100] Step 5: Generation of Multi-Granularity Load Decomposition Report

[0101] Based on the above three-level decomposition results, a non-intrusive load decomposition report covering four granularities is generated:

[0102] First granularity (branch box → distribution box): Output the load type composition, load percentage of each type, and load characteristic parameters (peak-valley ratio, day-to-day ratio, volatility level) of the area covered by each distribution box.

[0103] Second granularity (distribution box → metering box): Output the total load of each metering box, the start-stop event sequence statistics of high-power equipment (air conditioner, charging pile, electric water heater, etc.), and the abnormal load mode mark (the power consumption of a certain metering box deviates significantly from that of other metering boxes in the same distribution box).

[0104] The third granularity (meter box → user): outputs the total electricity consumption of each user, the running time statistics of each device, the estimated electricity consumption of each device, and the proportion of electricity consumption during peak and off-peak periods.

[0105] The fourth granularity (user → device): outputs the device-level power consumption details for each user (running time, estimated power, and power consumption percentage of each device), the ranking of high-energy-consuming devices (Top 3), the energy efficiency deviation compared with the average of similar users, and personalized energy-saving suggestions (such as "the running time of the fixed-frequency air conditioner is 1.5 times that of the average in the same area, it is recommended to replace it with an inverter air conditioner").

[0106] The core of the three-level joint non-intrusive load decomposition method proposed in this invention is "non-intrusive"—all measurements are completed at the main power inlet of each level of enclosure. The power consumption composition of each subordinate unit (distribution box / metering box / user) within the enclosure is inferred solely by analyzing the total voltage and current waveform signals, without requiring additional sensors to be installed in the subordinate units or on the equipment. The three-layer architecture each performs its specific function:

[0107] At the branch box level (first level), a coarse-grained load classification engine based on three-phase time-domain feature analysis is deployed. The branch box only collects the three-phase voltage and current waveforms of the main incoming line, extracting the time-domain features of the total three-phase active power (daytime ratio, peak-to-valley ratio, coefficient of variation, evening ratio). Combined with the power sequence proportions reported by each subordinate distribution box, a weighted classification rule is used to determine the load type (residential / commercial / public utilities / industrial) of the area covered by each distribution box. The classification results are encapsulated as Dirichlet prior distribution parameters and sent to the corresponding distribution box as a Bayesian prior constraint for distribution box-level equipment matching.

[0108] At the distribution box level (second level), a medium-granularity event detection engine based on Bayesian change point detection and prior constraints is deployed. The distribution box only collects the total incoming active power sequence, without touching the independent measurements of each outgoing circuit. The Bayesian Online Change Point Detection (BOCPD) algorithm is used to calculate the posterior probability of the running length to detect abrupt changes (equipment start-up and shutdown events) in the total power sequence in real time. For detected start-up events, combined with the load type prior issued by the branch box—e.g., "the dominant type of this distribution box is residential load"—the candidate range of equipment matching is significantly narrowed (excluding impossible equipment such as industrial motors). The start-up and shutdown tags of the matched high-power equipment are issued to the subordinate metering boxes.

[0109] At the metering box level (Level 3), a fine-grained device decomposition engine based on Seq2Point CNN and DTW fingerprint matching is deployed. The metering box only collects the total incoming line voltage and current waveforms, and extracts the total active power sequence using a 599-point sliding window, which is then input into the CNN. The CNN directly predicts the operating status of each device on the total power window—different devices leave different identifiable traces in the total power signal due to differences in power amplitude, start-stop transient characteristics, and harmonic content. The CNN achieves decomposition by learning the mapping relationship between these traces and device status. The high-power device tags issued by the distribution box are used as confidence enhancement signals during CNN inference, improving the identification reliability of the corresponding devices. For unknown devices whose types cannot be determined by the CNN, their VI trajectories are extracted, and the device type is identified by curve morphology matching with a pre-built electrical fingerprint database using Dynamic Time Warping (DTW).

[0110] The three-level model is collaboratively trained through federated distillation: the local CNN training in the metering box outputs temperature-scaled soft tags (10-dimensional probability vectors) which are then uploaded to the distribution box; the distribution box performs a weighted average aggregation of the soft tags from its subordinate metering boxes and uploads them to the cloud; the cloud distills the global soft tags and distributes them to each level. Throughout the process, only the soft tags (approximately 40 bytes) are transmitted between the boxes, while the raw 1kHz sampling data (approximately 4000 bytes per second) does not leave the metering box, fundamentally protecting users' electricity privacy.

[0111] Example 2

[0112] Please see Figure 2 As shown, a three-level box-based collaborative non-intrusive load decomposition system is presented. The overall architecture of the system is divided into three layers: a sensing hardware layer, an edge computing layer, and a cloud collaboration layer.

[0113] The sensing hardware layer deploys high-frequency / medium-frequency sampling units at the main power inlet of the three-level enclosure. Branch boxes deploy three-phase voltage and current high-frequency sampling modules (sampling rate ≥ 1kHz, 3 Hall current sensors + 3 resistive voltage divider sensors), covering up to the 50th harmonic. Distribution boxes deploy medium-frequency sampling modules (sampling rate ≥ 1kHz, Rogowski coil current sensor + voltage sampling circuit). Metering boxes deploy high-frequency sampling modules (sampling rate ≥ 1kHz, miniature CT current sensor + voltage sampling circuit + RS485 multi-meter polling interface). Each sampling module has a built-in high-precision real-time clock (GPS / BeiDou timing, synchronization accuracy ±1ms) to ensure consistent time reference for the three levels of data. All sensors are installed inside the enclosure at the main power inlet, without entering the user's premises or installing sensors on individual equipment, achieving completely non-intrusive sensing.

[0114] Edge computing layer: Each of the three-level enclosures deploys an edge computing gateway, which autonomously runs the NILM algorithm engine corresponding to the layer.

[0115] The branch box edge computing gateway includes:

[0116] The three-phase feature extraction module is used to perform FFT transformation on the three-phase voltage and current waveforms and calculate the three-phase active / reactive power and power factor.

[0117] The coarse-grained load classification module is used to extract time-domain feature vectors and determine the load type of each distribution box using weighted classification rules.

[0118] The prior knowledge encapsulation module is used to encapsulate the classification results into Dirichlet prior distribution parameters and distribute them via the MQTT protocol.

[0119] The distribution box edge computing gateway includes:

[0120] The prior knowledge receiving module is used to receive Dirichlet prior parameters sent by the branch box.

[0121] The BOCPD change point detection module is used to calculate the posterior probability of the run length and detect abrupt changes in the total power sequence in real time.

[0122] The device matching module is used to match the device type of the startup event using prior constraints and the device power fingerprint database.

[0123] The local aggregation module is used to collect the soft tags of subordinate metering boxes, aggregate them by weighted average, and then upload them to the cloud.

[0124] The metering box edge computing gateway includes:

[0125] The tag receiving module is used to receive start / stop tags for high-power equipment issued by the distribution box.

[0126] The Seq2Point CNN inference module deploys a lightweight TFLite model (model size < 5MB), using a 599-point sliding window for point-by-point device status prediction.

[0127] The device fingerprint matching module is used to extract the VI trajectory of unknown devices, match it with the local fingerprint database using DTW, and support online registration of new device types.

[0128] The soft tag generation module is used to generate soft tags for uploading by scaling the temperature (T=4.0).

[0129] The user-level decomposition output module is used to cross-validate the estimated power of the equipment with the sub-meter power consumption data and output the power consumption decomposition results of each user's equipment.

[0130] Cloud-based collaboration layer: Deploys a global distillation and aggregation module (collects aggregated soft tags from distribution boxes → global distillation → distributes global models), a multi-granularity decomposition report generation module (aggregates three-level decomposition results → four-granularity load reports), and a device fingerprint database management module (maintains and updates the global device fingerprint database, and periodically distributes newly added device types to gateways at all levels).

[0131] The data flow is: Branch box gateway (prior parameter α) → Distribution box gateway (equipment start / stop tag) → Metering box gateway (equipment status, power estimation). The reverse flow is: Metering box gateway (soft tag). → Distribution box gateway (aggregated soft tag) → Cloud (Global Soft Tag) → Distribute to gateways at all levels for local distillation training.

[0132] This invention establishes for the first time a three-level progressive NILM architecture: "coarse classification of branch boxes → detection in distribution boxes → fine decomposition of metering boxes." All sensing is completed at the main inlet of each level of box, eliminating the need to install sensors in subordinate units or equipment. The core of this architecture lies in the fact that the decomposition result of each level is not used as the endpoint, but rather as a Bayesian prior injection for the next level. The load type classification result of the branch box (e.g., "the area covered by this distribution box is residential load") is encapsulated as a Dirichlet prior, constraining the candidate space for matching equipment in the distribution box (excluding impossible equipment such as industrial motors); the start / stop label of the distribution box (e.g., "an air conditioner was started at 9:27") serves as a confidence enhancement signal during CNN inference in the metering box, increasing the confidence of air conditioner identification at that moment from 0.62 to 0.74. At the same time, the new equipment types identified by the lower-level VI trajectory matching are fed back to update the fingerprint database of the upper level, forming a closed loop of "prior injection → fine decomposition → fingerprint backhaul → prior update."

[0133] This invention establishes an electrical fingerprint system encompassing four dimensions: steady-state characteristics (active power range, reactive power range, power factor—reflecting the electrical characteristics of the equipment during normal operation), transient characteristics (start-up power ratio, start-up duration—reflecting the impact characteristics of the equipment during start-up and shutdown), harmonic characteristics (3rd / 5th / 7th harmonic content—reflecting the nonlinear load characteristics of the equipment, such as the significant 3rd harmonic generated by an air conditioning compressor), and VI trajectory characteristics (normalized voltage-current Lissajous figure—reflecting the impedance characteristics of the equipment, such as a straight line for purely resistive equipment and a closed curve for nonlinear equipment). The metering box level extracts these characteristics solely from the V / I waveform of the main input line, eliminating the need for sensors at the equipment end and achieving completely non-invasive simultaneous identification of multiple devices.

[0134] This invention addresses the heterogeneous structure of the three-level enclosure NILM model (branch boxes use a weighted classification model to output 4 classes, distribution boxes use BOCPD to detect events, and metering boxes use a CNN to output 10 classes of device status), and designs a federated distillation method based on temperature-scaled soft labels. It replaces the tens of thousands of parameters of the CNN model or the raw high-frequency sampling data (approximately 4000 bytes per second) transmitted over the network with 10-dimensional soft label vectors (approximately 40 bytes), solving the problem that traditional FedAvg parameter averaging is not suitable for heterogeneous models. The three-level distillation architecture (metering box local training → distribution box local aggregation → cloud-based global distillation) fully utilizes the natural hierarchical advantage of "similar user equipment usage habits within the same distribution box."

[0135] Example 3

[0136] The following example of a residential community will be used to illustrate the workflow of this invention.

[0137] Application Scenario: A medium-sized residential community in a city, with a total of 288 households. The power distribution topology is as follows: 1 public transformer → 1 branch box (deployed in the community's power distribution room) → 3 distribution boxes (deployed in the ground floor distribution rooms of Buildings 1, 2, and 3 respectively) → 12 metering boxes (4 per building, each metering box covering 6 households). All sensing devices are installed at the main power inlet inside the boxes, without entering the households.

[0138] Hardware Deployment: Branch Box Main Inlet: Three-phase Hall current sensors ×3 (range 0-500A) + Resistor voltage divider sensors ×3 (range 0-400V), 1kHz synchronous sampling. Distribution Box Main Inlet (×3 units): Rogowski coil current sensor (range 0-200A) + voltage sampling circuit, 1kHz sampling. Metering Box Main Inlet (×12 units): Miniature CT current sensor (range 0-100A) + voltage sampling circuit + RS485 ×4 interfaces (polling 6 sub-meters in the box), 1kHz sampling. Edge Computing Gateways at All Levels: ARM Cortex-A55 quad-core processor, 4GB LPDDR4 memory, 32GB eMMC. Metering Box Gateway Deploys TFLiteSeq2Point CNN Inference Model (Model File <5MB).

[0139] Example execution: Using data from August 15, 2025 as an example.

[0140] At 9:00 AM, the branch box gateway performed coarse-grained classification of the total three-phase active power data from the past 24 hours. The characteristic values ​​and classification results for the three distribution boxes are as follows:

[0141] Distribution Box 1 (Building 1): , , , All four features fall within the typical range of "residential load," with a confidence level of 0.85. Dirichlet prior. .

[0142] Distribution Box 2 (Building 2): , , , Determined to be residential load, .

[0143] Distribution Box 3 (Building 3, ground floor includes shops): , , , The daytime electricity consumption ratio is relatively high (shops' daytime electricity consumption), indicating a mixed residential and commercial sector, with a residential-dominated confidence level of 0.65. .

[0144] At 9:15 AM, the gateway of distribution box 1 ran BOCPD in a 1-minute active power sequence. A change was detected at the 27th sampling point (corresponding to 9:27 AM), and the posterior probability of the change was... (Exceeding the threshold of 0.8), power surge Equipment matching calculation:

[0145] air conditioner: ;

[0146] Electric water heater: ;

[0147] Charging stations: ;

[0148] Industrial motors: ;

[0149] The air conditioner scored the highest, and the event was determined to be an air conditioner start-up event. The tag “(t=9:27, ΔP=2100W, type=fixed frequency air conditioner, confidence=0.73)” was generated and sent to the four metering boxes under distribution box 1.

[0150] At 9:30 AM, metering box 1-1 received the air conditioner start tag. A power window of 599 points centered at 9:27 AM was extracted and input into a Seq2Point CNN. The original CNN output showed an air conditioner ON probability of 0.62. Label augmentation was then performed. ): The system is set to ON. Air conditioner power estimation: At that moment, both the refrigerator (probability 0.85, power estimated 210W) and the lighting / standby base load (approximately 300W) were simultaneously identified, with the total power estimated... The measured power of the metering box's total incoming line is close to 2005W, with a deviation of only 0.35%, indicating good consistency.

[0151] Simultaneously, a 720W power step was detected at the main inlet of metering box 1-1, but the output probability of the CNN for all 10 known device types was less than 0.3. The VI trajectory recognition process was triggered: normalized VI trajectories (64-point resampling) for 5 power frequency cycles before and after this moment were extracted, and DTW matching was performed against the fingerprint database one by one. Matching results: the shortest distance was for "microwave oven" (DTW distance = 0.08), and the second shortest was for "electric kettle" (0.35). It was determined that the microwave oven was started (VI trajectory is elliptical, power factor is close to 1.0, and start-up duration is extremely short <50ms). The system registered the microwave oven's VI trajectory in the local fingerprint database.

[0152] At 10:00 AM, the federated distillation cycle arrived. The local CNN in metering chamber 1-1 processed 3600 power windows over the past hour, generating 10-dimensional soft label vectors. (Values ​​correspond to 10 categories of equipment, with the highest values ​​for the first three being air conditioners (0.32), refrigerators (0.18), and other base loads (0.25). Distribution box 1 collects soft tags from its four subordinate metering boxes, and the weighted average is calculated based on the sample size. Upload to the cloud. The cloud aggregates tags from three distribution boxes and calculates global soft tags. Issued. Calculation for distribution box 1. Used for the next round of local distillation training ( ).

[0153] At 0:00 the following day, the system will generate a four-granularity load decomposition report:

[0154] Granularity 1 (Branch Box → Distribution Box): Distribution Box 1 covers 96 households in Building 1, with a total daily electricity consumption of 2,398 kWh. Load type: 96.5% residential, 3.5% public facilities. Peak-to-valley ratio is 1.82, which is typical residential load.

[0155] Granularity 2 (Distribution Box → Metering Box): Metering Box 1-1 (floors 1-6) has a daily electricity consumption of 628 kWh, with air conditioners starting 23 times / day (total for 6 households), electric water heaters starting 12 times / day, and charging stations charging 3 times / day. Metering Box 1-4 (floors 19-24) has a daily electricity consumption of 527 kWh (fewer residents on the top floor, high vacancy rate).

[0156] Granularity 3 (Metering Box → User): User 1-1-302 (3rd Floor, Unit 2) has a daily electricity consumption of 35.2 kWh. The air conditioner runs for 8.5 hours, accounting for 51.4% of the daily electricity consumption; the refrigerator runs for 24 hours, accounting for 14.2% of the daily electricity consumption; other equipment accounts for 34.4% of the total.

[0157] Granularity 4 (User → Equipment): User 1-1-302 Equipment Power Consumption Details—Air Conditioner 18.1 kWh (51.4%, running for 8.5 hours), Refrigerator 5.0 kWh (14.2%, running for 24 hours), Electric Water Heater 4.5 kWh (12.8%, running for 1.5 hours), Washing Machine 1.8 kWh (5.1%, running for 1.2 hours), Microwave Oven 0.5 kWh (1.4%, running for 0.3 hours), Lighting and Television 5.3 kWh (15.1%). Energy Saving Recommendations: The air conditioner's operating time is 37% higher than the regional average of 6.2 hours. It is recommended to adjust the temperature from 24℃ to 26℃ (estimated daily power saving of approximately 3.5 kWh); for the electric water heater, it is recommended to set the temperature to 50℃ and use the timer function to avoid evening peak hours; the results are as follows. Figure 8 As shown.

[0158] As can be seen from Example 3, the present invention has the following technical effects:

[0159] (1) Significantly improves the accuracy of non-intrusive load decomposition. Through a three-level progressive NILM architecture, the decomposition results of the upper level provide Bayesian prior constraints for the lower level, and the accurate labels of the lower level are sent back to correct the upper level. In a typical residential community scenario (1 branch box, 3 distribution boxes, 4 metering boxes under each distribution box, 6 households per metering box), the F1-score for the identification of 10 types of household appliances at the metering box level reaches 0.92, which is about 18% higher than that of single-level NILM (F1-score of about 0.78). Figure 5(As shown). The detection accuracy of high-power equipment at the distribution box level has increased from 65% without prior knowledge to 83% with prior knowledge (e.g.). Figure 4 (As shown). Prior injection of the four load categories at the branch box level reduces the candidate set for matching distribution box equipment by approximately 50%.

[0160] (2) Achieving continuous optimization of the global model under privacy protection. During the three-level federated distillation training process, only the 10-dimensional soft label vector (40 bytes) is transmitted between the boxes, and the original 1kHz high-frequency sampling data (containing sensitive user electricity privacy information, approximately 4000 bytes per second) does not leave the metering box edge gateway. The amount of communication data is reduced by approximately 60% compared to the FedAvg parameter transmission scheme (approximately 5MB of model parameters). Figure 7 As shown). In non-independent, identically distributed scenarios (each distribution box covers different user types), the global model accuracy is improved by 12% compared to training each metering box independently, and by 7% compared to the flat FedAvg architecture (e.g.). Figure 3 (As shown). Setting the distillation temperature T=4.0 allows the soft tag to fully retain similarity information between equipment classes. For example... Figure 6 As shown, after 50 rounds of training, the device recognition F1-score of this scheme reaches 0.888, while the device recognition F1-score of the flat FedAvg architecture is 0.834. In addition, the device recognition F1-score of each meter box after isolated training is only 0.786.

[0161] (3) Provides full-granularity load sensing capabilities from area to device. This invention realizes four-granularity non-intrusive load decomposition along the distribution topology—the power grid company can understand the load type composition and load characteristics of each distribution box downstream of the branch box, the property management company can grasp the usage statistics of high-power equipment and abnormal load patterns of each metering box, and users can obtain device-level electricity consumption details ("Air conditioning accounts for 45% of your total electricity consumption, and the running time is 1.5 times higher than the average of the same area") and personalized energy-saving suggestions. All sensing information is obtained in a non-intrusive manner, without the need for home renovation or the installation of sensors, resulting in low deployment costs and high user acceptance.

[0162] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A non-intrusive load decomposition method for three-level box-type coordination, characterized in that, Includes the following steps: Step 1: Collect three-phase voltage and current waveform data at the main incoming line of the branch box, calculate the total three-phase active power sequence, extract the time-domain features of the power sequence, and use weighted classification rules based on the time-domain features to determine the load type of each subordinate distribution box. Encapsulate the classification results into Dirichlet prior distribution parameters and send them to the corresponding distribution box gateway. Step 2: The distribution box collects the total active power sequence at the main incoming line, receives the Dirichlet prior distribution parameters sent by the branch box, uses the Bayesian online change point detection algorithm to detect abrupt changes in the total power sequence, determines the equipment start event based on the power change before and after the abrupt change point, and performs equipment type matching for the start event by combining the Dirichlet prior distribution parameters and the equipment power fingerprint database, generates equipment start / stop tags and sends them to the subordinate metering box gateway. Step 3: The metering box collects voltage and current waveform data at the main incoming line, calculates the total active power sequence, extracts the power sequence with a sliding window and inputs it into the Seq2Point CNN model, outputs the operating status probability of each device, receives the device start / stop tags issued by the distribution box as confidence enhancement signals to enhance the CNN output, and determines the device operating status and estimates the device power based on the enhanced confidence.

2. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 1, characterized in that, Also includes: Step 4: The metering box uses local data to train a Seq2Point CNN model, performs temperature scaling on the model output to generate soft tags, and uploads them to the corresponding distribution box. The distribution box collects the soft tags of all its subordinate metering boxes, performs a weighted average aggregation, and then uploads the data to the cloud. The cloud collects aggregated soft tags from all distribution boxes, calculates global soft tags, and distributes them to the gateways of each level of the box. During the next round of training, the local models at each level add a KL divergence distillation loss term to the loss function for optimization. Step 5: Generate a load decomposition report with four granularities based on the three-level decomposition results of branch boxes, distribution boxes, and metering boxes: The first granularity is the load type composition and the proportion of each type of load from the branch box to the distribution box; The second granularity includes the total load from the distribution box to the metering box, the timing statistics of start-up and shutdown events of high-power equipment, and the marking of abnormal load modes; the third granularity includes the total electricity consumption from the metering box to the user, the running time statistics of each device, and the estimated electricity consumption of each device; the fourth granularity includes the device-level electricity consumption details from the user to the equipment, the ranking of high-energy-consuming equipment, and personalized energy-saving suggestions.

3. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 1, characterized in that, The time-domain features extracted from the power sequence in step 1 include: daytime power ratio, peak-to-valley ratio, coefficient of variation, and evening power ratio; the weighted classification rule is expressed as: ； in , For feature set, For type In features Typical range of values ​​for; For indicator functions; Features The weight values.

4. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 1, characterized in that, In step 2, the running length of the Bayesian online change point detection algorithm The joint posterior probability recursive formula is: ； ； ； in for The observed total active power at any given time; The transition probability at the runtime is given by a constant risk rate in the risk function at the point of change. , The predicted distribution of the data under a given runtime is a Gaussian distribution. ,parameter and By running length Estimation of the corresponding historical data segment; When the posterior probability of the point of change Exceeding the threshold At that time, the judgment The moment is the point of change.

5. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 1, characterized in that, In step 2, the comprehensive score for device type matching It consists of two items: ； in For equipment The median of the rated power range, This is the load type index corresponding to the device in the Dirichlet prior; As weight; This represents the change in power before and after the point of change. For the Dirichlet prior distribution parameters, select the one with the highest comprehensive score that exceeds the threshold. The device type is used as the matching result.

6. The non-intrusive load decomposition method and system for three-level enclosure coordination according to claim 1, characterized in that, In step 3, the confidence enhancement method for the CNN output of the equipment start / stop labels issued by the distribution box is as follows: ； in, Indicates the first The probability that a device of this type is currently in operation. The label enhancement factor. For indicator functions; For equipment start / stop labels.

7. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 1, characterized in that, In step 3 When the output probability of the CNN for all known device categories is lower than the preset threshold, the non-intrusive VI trajectory recognition process is triggered. Normalized VI trajectory features are extracted from the voltage and current waveforms of five power frequency cycles before and after this moment, and the trajectory feature vector is obtained by 64-point equidistant resampling of the VI scatter points. ; The dynamic time warping algorithm is used to match the VI trajectory templates of all devices in the pre-built fingerprint database one by one. The matching model is as follows: ,in Align the path; The device type with the smallest DTW distance is selected as the identification result. When the smallest DTW distance is lower than the matching threshold, the unknown device is registered in the local fingerprint database.

8. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 2, characterized in that, In step 4 The soft label generated by the metering box is: ;in This is the logits vector output by the Seq2PointCNN model for a batch of window samples; Number of equipment categories; The soft label for the distribution box after weighted average aggregation is as follows: ;in The training sample size for each of the metrology boxes; The global soft tag calculated in the cloud is: ;in This represents the total sample size of all distribution boxes and their subordinate units.

9. The non-intrusive load decomposition method for three-level enclosure coordination according to claim 8, characterized in that, In step 4, the training loss function for each level of the local model is: ； ； in The cross-entropy loss is used for binary classification. Let KL divergence be the KL divergence. For distillation loss weighting, This is the temperature squared scaling factor.

10. A three-stage, box-type, non-invasive load decomposition system, applicable to the method of any one of claims 1-9, characterized in that, include: The sensing hardware layer is deployed at the main power inlet of the three-level enclosure, including a high-frequency sampling module for branch boxes, a medium-frequency sampling module for distribution boxes, and a high-frequency sampling module for metering boxes, which are used to collect voltage and current waveform data of the main incoming line of each level enclosure; The edge computing layer comprises edge computing gateways deployed in the three-level enclosure, including branch box edge computing gateways, distribution box edge computing gateways, and metering box edge computing gateways. The branch box edge computing gateway includes a three-phase feature extraction module, a coarse-grained load classification module, and a prior knowledge encapsulation module, used to perform the operation in step 1. The distribution box edge computing gateway includes a prior knowledge receiving module, a BOCPD change point detection module, and a device matching module, used to perform the operation in step 2. The metering box edge computing gateway includes a tag receiving module, a Seq2Point CNN inference module, and a device fingerprint matching module, used to perform the operation in step 3. The cloud collaboration layer communicates with the edge computing layer and is used to generate load decomposition reports.

Images