Non-intrusive high-power electric appliance access detection and risk early warning method, system and device and medium

Abstract

The invention discloses a non-intrusive high-power electric appliance access detection and risk early warning method, system and device and a medium, and the method comprises the steps: carrying out the differential threshold detection based on a general meter side instantaneous power sequence, recognizing an opening / closing event, and screening a high-power electric appliance candidate event through a first preset threshold; performing voltage zero-crossing phase alignment on the starting event, extracting steady-state current waveforms before and after the starting event, and calculating an event current component; inputting the event current component into a time sequence convolutional network for processing to obtain learnable image features, inputting the learnable image features into a two-dimensional convolutional neural network for classification, and outputting a load category to which the event belongs; obtaining a running time self-adaptive threshold value through a first clustering algorithm; and if the paired closing event is not detected in the operation time adaptive threshold value of the candidate event of the high-power electric appliance, triggering an early warning mechanism. According to the invention, the high-efficiency detection and risk prevention and control of the access of the high-power electric appliance are realized, and the practicability and the popularization value are relatively high.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring technology, and in particular to a non-intrusive method, system, equipment, and medium for detecting and warning of risks associated with high-power electrical appliances connected to the power grid. Background Technology

[0002] Most existing methods for monitoring high-power appliances rely on invasive monitoring techniques, which involve installing dedicated sensors at each appliance's terminals to collect energy consumption data. While these methods can obtain relatively accurate appliance operating information, they face challenges in practical application, including complex installation, high costs, and low user acceptance, hindering large-scale adoption in residential and office environments. In contrast, Non-Intrusive Load Monitoring (NILM) technology, by collecting voltage, current, or power signals from the main meter terminal, identifies and decomposes the operating status and energy consumption of various loads without requiring sensors on individual appliances, making it more practical and economical.

[0003] In traditional NILM methods, the mainstream load identification techniques mainly include signal processing methods based on event detection and feature extraction, statistical modeling methods based on hidden Markov models, and pattern recognition methods based on deep learning methods. Signal processing methods rely on manually designed features and can achieve good recognition results in specific scenarios, but they are not adaptable to scenarios with complex loads or multiple electrical appliances operating simultaneously. Statistical modeling methods infer the state of electrical appliances by establishing the implicit relationship between state transitions and observed power, which has the advantages of intuitive modeling and high inference efficiency, but they are often limited in handling multi-state electrical appliances and high-noise scenarios. Deep learning methods can automatically learn features and have high recognition accuracy, but they are heavily dependent on a large amount of labeled data, and the model's generalization ability is also easily limited. In addition, there are still obvious shortcomings in the current non-invasive detection and risk warning of high-power electrical appliances: (1) Load identification relies on traditional manual features or shallow models, and the recognition ability is limited; (2) There is a lack of running time threshold models that combine user historical behavior data, making it difficult to effectively monitor abnormal operating behavior; (3) Existing research ignores the safety risk characteristics of high-power electrical appliances and has not yet formed a systematic risk warning mechanism.

[0004] Therefore, there is an urgent need for a new method that can combine appliance recognition based on learnable image features, make full use of users' historical electricity consumption data to establish normal operating time thresholds for various appliances, and on this basis, build a detection and early warning mechanism for high-power appliances, thereby achieving accurate identification of high-power appliances connected to the system, effective monitoring of overtime operation, and timely early warning of potential risks. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a non-intrusive high-power appliance access detection and risk warning method, system, device and medium to solve the problem that current non-intrusive load monitoring technology lacks a specific risk warning mechanism for high-power appliances and is difficult to adaptively judge abnormal operating status by combining user historical behavior.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a non-invasive method for detecting and warning of risks associated with the connection of high-power electrical appliances, comprising: Based on differential threshold detection of the instantaneous power sequence on the main meter side, the system identifies on-start and off events, and filters candidate events for high-power appliances using a first preset threshold. The voltage zero-crossing phase is aligned for the turn-on event, the steady-state current waveforms before and after the turn-on event are extracted, and the event current component is calculated. The event current components are input into a temporal convolutional network for processing to obtain learnable image features. The learnable image features are input into a two-dimensional convolutional neural network for classification, and the load category to which the event belongs is output. Based on the historical identification results, a set of durations for each electrical appliance is constructed, and an adaptive threshold for running time is obtained through the first clustering algorithm; If no paired shutdown event is detected for the candidate high-power electrical appliance within the adaptive operating time threshold, an early warning mechanism is triggered.

[0008] As a preferred embodiment of the non-intrusive high-power appliance access detection and risk warning method of the present invention, wherein: the step of screening candidate events for high-power appliances with a first preset threshold includes: Voltage and current sequences are collected from the user's main meter terminal to calculate the instantaneous power sequence; Perform sliding difference calculation on the instantaneous power sequence to detect the occurrence time of the load switching event; When the result of the sliding differential operation exceeds the set event detection threshold, a load switching event is determined to exist; If the steady-state power corresponding to the load switching event exceeds the first preset threshold, the load switching event is recorded as a candidate event for high-power appliances; otherwise, it is discarded. Based on the power step conservation principle, the turn-on event and the turn-off event are paired in the time neighborhood to obtain a complete event pair or an independent turn-on event marked as pending confirmation of turn-off.

[0009] As a preferred embodiment of the non-intrusive high-power electrical appliance access detection and risk warning method of the present invention, the calculation of the event current component includes: Select multiple power frequency cycles prior to the occurrence of the start-up event to calculate the steady-state baseline current; Select multiple power frequency cycles after the start-up event occurs, and calculate the steady-state current after the event; The event current component is calculated based on the steady-state baseline current and the post-event steady-state current. The event current component is then subjected to a first processing step to generate a standardized event current waveform.

[0010] As a preferred embodiment of the non-intrusive high-power electrical appliance access detection and risk warning method of the present invention, wherein: the step of inputting the event current component into a temporal convolutional network for processing to obtain learnable image features includes: The standardized event current waveform is input into a time-series convolutional network to extract the time-series feature matrix; The temporal feature matrix is ​​aligned by channel number through pointwise convolution to obtain the aligned feature matrix. Based on the aligned feature matrix, calculate the correlation matrix of different feature channels on the time axis; The correlation matrix is ​​normalized into a single-channel grayscale image, which is then output as a learnable image feature.

[0011] As a preferred embodiment of the non-intrusive high-power appliance access detection and risk warning method of the present invention, wherein: the load category to which the output event belongs includes: The learnable image features are input into a multi-layer two-dimensional convolutional neural network, and the learnable image features are processed layer by layer through batch normalization, non-linear activation and pooling operations. The extracted image features are input into a fully connected layer, processed by the Softmax function to obtain the probability distribution of the load categories, and the load category with the highest probability is selected as the recognition result, and the corresponding classification confidence is output.

[0012] The beneficial effects of this preferred technical solution are: it can automatically learn the differentiated characteristics of electrical appliance operation, significantly improving the accuracy and generalization ability of load identification.

[0013] As a preferred embodiment of the non-intrusive high-power appliance access detection and risk warning method of the present invention, wherein: obtaining the adaptive threshold for running time through the first clustering algorithm includes: Collect samples of the operating duration of each electrical appliance within a historical period to form a duration set. Perform robust preprocessing on the duration set to eliminate the influence of outlier data points. The first clustering algorithm is used to perform cluster analysis on the preprocessed duration set to determine the optimal number of clusters. Based on the posterior responsibility degree and prior business knowledge, the data clusters corresponding to normal working conditions are selected from the clustering results. Calculate the adaptive threshold for runtime and add a safety margin based on the distribution characteristics of the normal operating condition data cluster; The runtime adaptive threshold is updated and adjusted online to adapt to changes in user electricity consumption behavior.

[0014] The beneficial effects of this preferred technical solution are: by dynamically establishing an operating time threshold that conforms to the user's actual electricity consumption habits, it effectively distinguishes between normal use and abnormal overtime operation, and significantly reduces the false alarm rate.

[0015] As a preferred embodiment of the non-intrusive high-power electrical appliance access detection and risk warning method of the present invention, the triggering warning mechanism includes: When an event is detected and identified as a high-power appliance and the start time is recorded, runtime monitoring is initiated; if a paired shutdown event is detected within the runtime adaptive threshold, the event registration record is cleared; if the current time exceeds the runtime adaptive threshold and no shutdown event is detected, a timeout warning is triggered. The timeout warning level is determined by comprehensively considering classification confidence level, timeout duration, and continuous power. The actual runtime duration is written back to the historical dataset, and the runtime adaptive threshold is periodically updated.

[0016] Secondly, the present invention provides a non-invasive high-power electrical appliance access detection and risk warning system, comprising: The load event detection module is used for differential threshold detection based on the instantaneous power sequence of the main meter side, to identify on-off events and off-off events, and to filter candidate events of high-power appliances with a first preset threshold. The feature extraction module is used to perform voltage zero-crossing phase alignment on the turn-on event, extract the steady-state current waveform before and after the turn-on event, and calculate the event current component. The image feature extraction module is used to input the event current component into a temporal convolutional network for processing to obtain learnable image features. The load feature classification module is used to input the learnable image features into a two-dimensional convolutional neural network for classification and output the load category to which the event belongs. The threshold acquisition module is used to construct a set of durations for each electrical appliance based on historical recognition results, and to obtain an adaptive threshold for running time through the first clustering algorithm; The risk warning module is used to trigger a warning mechanism if no paired shutdown event is detected for the candidate high-power electrical appliance event within the adaptive operating time threshold.

[0017] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor, when executing the computer-executable instructions, implements the steps of a non-intrusive high-power electrical appliance access detection and risk warning method.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a non-intrusive high-power electrical appliance access detection and risk warning method.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: By integrating load recognition technology based on learnable image features with adaptive modeling based on time thresholds of user historical behavior, this invention achieves accurate detection and intelligent risk warning of high-power electrical appliance access. This invention can automatically learn the differentiated characteristics of appliance operation, significantly improving the accuracy and generalization ability of load recognition; by dynamically establishing operating time thresholds that conform to users' actual electricity consumption habits, it effectively distinguishes between normal use and abnormal overtime operation, greatly reducing the false alarm rate; simultaneously, by establishing an event-driven closed-loop early warning mechanism, it can promptly trigger alarms when appliances exceed their time limits, preventing safety hazards caused by prolonged appliance operation from the source, and significantly improving the initiative and reliability of user-side electricity safety management. This invention achieves efficient detection and risk control of high-power electrical appliance access, and has strong practicality and promotional value. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall process logic of a non-intrusive high-power electrical appliance access detection and risk warning method provided in one embodiment of the present invention.

[0022] Figure 2 This is a detailed flowchart illustrating a non-intrusive high-power electrical appliance access detection and risk warning method provided in one embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figures 1-2 As one embodiment of the present invention, a non-invasive method for detecting and warning of risks associated with the connection of high-power electrical appliances is provided, such as... Figure 1 The specific steps shown are as follows: S100: Based on the differential threshold detection of the instantaneous power sequence on the main meter side, identify the on and off events, and filter candidate events of high-power appliances with a first preset threshold; S200: Perform voltage zero-crossing phase alignment on the turn-on event, extract the steady-state current waveform before and after the turn-on event, and calculate the event current component; S300: Input the event current component into a temporal convolutional network for processing to obtain learnable image features; S400: Input learnable image features into a two-dimensional convolutional neural network for classification and output the load category to which the event belongs; S500: Construct a set of durations for each electrical appliance based on historical identification results, and obtain an adaptive threshold for running time through the first clustering algorithm; S600: If no paired shutdown event is detected for a candidate event of a high-power appliance within the adaptive threshold of running time, an early warning mechanism is triggered.

[0025] It should be noted that, to address the lack of a specific risk warning mechanism for high-power appliances in current non-intrusive load monitoring technologies, and the difficulty in adaptively judging abnormal operating states by incorporating user historical behavior, steps S100-S600 above achieve accurate detection and intelligent risk warning for high-power appliance access by integrating load recognition technology with learnable image features and adaptive modeling based on time thresholds of user historical behavior. This invention can automatically learn the differentiated characteristics of appliance operation, significantly improving the accuracy and generalization ability of load identification; by dynamically establishing operating time thresholds that conform to users' actual electricity consumption habits, it effectively distinguishes between normal use and abnormal overtime operation, greatly reducing the false alarm rate; simultaneously, by establishing an event-driven closed-loop warning mechanism, it can promptly trigger alarms when appliances exceed their time limits, preventing safety hazards caused by prolonged appliance operation from the source, and significantly improving the initiative and reliability of user-side electricity safety management. This invention achieves efficient detection and risk control for high-power appliance access, and has strong practicality and promotional value.

[0026] like Figure 2 The diagram shown is a detailed flowchart of the method provided in an embodiment of the present invention. The following is a summary of the process. Figure 2 Specific embodiments of the present invention will be described below.

[0027] In this embodiment of the invention, step S100, based on differential threshold detection of the instantaneous power sequence at the main meter side, identifies on-start and off events, and filters candidate events for high-power appliances using a first preset threshold, including: Voltage sequence collected from the user's main meter terminal V(t) With current sequence I(t) Calculate the instantaneous power sequence P(t) : Perform sliding difference calculations on the instantaneous power sequence to detect the timing of load switching events: When the result of the sliding difference operation exceeds the set event detection threshold When a load switching event occurs, it is determined that a load switching event has occurred; if the steady-state power corresponding to the load switching event exceeds a first preset threshold, i.e. If the load switching event is not recorded, it will be recorded as a candidate event for high-power appliances; otherwise, it will be discarded. Based on the power step conservation principle, on-events and off-events are paired within the time neighborhood to obtain complete event pairs or independent on-events marked as pending closure confirmation; each event contains ( , ) or only on at certain times (to be confirmed for closure), where, Indicates when the appliance is turned on. Indicates the time when the appliance is turned off.

[0028] It should be noted that the event detection threshold The threshold setting is primarily based on the need to distinguish between the natural fluctuations of the total power sequence during steady-state operation and the significant power changes generated during load switching. This threshold needs to be set higher than the maximum power fluctuation caused by grid noise, measurement errors, and the start-up and shutdown of low-power equipment to avoid misjudging normal fluctuations as load events; simultaneously, this threshold needs to be lower than the minimum power step generated during the switching of the target appliance to be detected, to ensure that the target event can be effectively captured.

[0029] It should be noted that the first preset threshold is usually set at 3000W, which is determined based on the minimum operating power of common high-power appliances and residential electricity safety regulations. By using this threshold, typical high-power devices such as electric heaters, induction cookers, and instant water heaters can be included in the monitoring scope, while conventional low-power loads such as lighting, televisions, and computers are filtered out, thereby achieving focused monitoring resources and effective risk identification.

[0030] In this embodiment of the invention, step S200 involves aligning the voltage zero-crossing phase of the turn-on event, extracting the steady-state current waveforms before and after the turn-on event, and calculating the event current components, including: Calculate the steady-state baseline current by selecting multiple power frequency cycles before the start-up event occurs; Select multiple power frequency cycles after the start-up event occurs and calculate the steady-state current after the event; The event current component is calculated based on the steady-state baseline current and the post-event steady-state current. The event current component is then processed to generate a standardized event current waveform.

[0031] Specifically, for each enable event Alignment is made with the voltage zero-crossing point, and the length of the power frequency cycle is set to be... Sampling points. The average steady-state baseline current is calculated over M cycles prior to the start-up event, expressed by the formula: Specifically, multiple power frequency cycles after the start-up event are selected, and the steady-state current after the event is calculated. The formula is as follows: Specifically, based on the steady-state baseline current and the post-event steady-state current, the event current component is calculated, and the single-event appliance current waveform is approximated as follows: Specifically, the event current components are denoised, amplitude normalized, and length standardized to generate a standardized event current waveform.

[0032] In an optional embodiment, the first processing can also employ a wavelet threshold denoising combined with dynamic time warping. First, wavelet transform is used to remove high-frequency noise interference in the event current waveform. Then, the phase difference of different event waveforms is aligned through the dynamic time warping algorithm. Finally, amplitude normalization is performed, thereby improving the consistency and comparability of the data while preserving the key details of the waveform.

[0033] In another alternative embodiment, the first processing can also introduce an adaptive moving average filtering and piecewise linear representation process. First, the size of the filtering window is automatically adjusted according to the signal noise level to smooth random fluctuations. Then, the waveform is divided into characteristic segments and its main changing trend is fitted with straight line segments. Finally, it is uniformly scaled to a standard length, which effectively reduces noise while highlighting the morphological characteristics of the load switching process.

[0034] In this embodiment of the invention, step S300, which inputs the event current component into a temporal convolutional network for processing to obtain learnable image features, includes the following sub-steps C1 to C4: In C1: The standardized event current waveform is input into the temporal convolutional network to extract the temporal feature matrix; Specifically, the standardized event current waveform is input into a one-dimensional temporal convolutional network (TCN) to obtain the time... The channel feature matrix is ​​expressed by the formula: in, n The length of the time step (the effective length after TCN extraction); m This refers to the number of channels (the number of feature map channels in the final output of TCN).

[0035] Specifically, the matrix F is broken down row by row (by time): For the first t The feature vector of each event frame. At time frame t, the first i The activation values ​​of each channel feature subspace. The matrix is ​​arranged as follows (rows = time, columns = channels): in, Let m represent the eigenvalue of the m-th channel at time n.

[0036] In C2: The temporal feature matrix is ​​aligned by channel number through pointwise convolution, resulting in the aligned feature matrix, expressed by the formula: In C3: Based on the aligned feature matrix, calculate the correlation matrix of different feature channels on the time axis; Specifically, define channel vectors LGC's ( i, j The element measure of the correlation between different channels on a common time axis is expressed as: in, This represents the similarity between channel i and channel j over the entire time series. This represents the feature value of the i-th channel at time point t. This represents the feature value of the j-th channel at time point t.

[0037] In C4: Normalize the correlation matrix into a single-channel grayscale image. As a learnable image feature output.

[0038] In an optional embodiment, the step of extracting learnable image features can also employ a method combining short-time Fourier transform and attention enhancement. First, the event current component is converted into a time-frequency spectrum through STFT. Then, the spectral features are recalibrated using a channel attention mechanism. Finally, the enhanced time-frequency image features are generated by compression through a convolutional layer, thereby simultaneously preserving the time-domain and frequency-domain discrimination information of the signal.

[0039] In another alternative embodiment, the step of extracting learnable image features can also introduce a multi-scale waveform slicing and Gram angle field conversion process. First, the current waveform is divided into overlapping segments according to different time scales. Then, each one-dimensional segment is converted into a two-dimensional polar coordinate image through Gram angle field. Finally, these images are stacked to form a multi-channel feature map to capture the spatiotemporal variation pattern of waveform morphology at multiple granularities.

[0040] In this embodiment of the invention, step S400 inputs learnable image features into a two-dimensional convolutional neural network for classification, and outputs the load category to which the event belongs, including: Learnable image features are input into a multi-layer two-dimensional convolutional neural network, and the learnable image features are processed layer by layer through batch normalization, non-linear activation and pooling operations. The extracted image features are input into a fully connected layer, processed by the Softmax function to obtain the probability distribution of the load categories, and the load category with the highest probability is selected as the recognition result. The corresponding classification confidence score is then output, as expressed by the formula: Specifically, the cross-entropy loss function is used during the training phase: in, The true value is one-hot. During the inference phase, the category with the highest probability is used as the identification result, and the confidence score is output for weighting the warning level.

[0041] In an optional embodiment, load category identification can also employ a dual-path feature fusion and graph neural network classification method. Learnable image features and time-domain statistical features extracted from the original waveform are respectively input into two parallel feature extraction networks. Then, the graph neural network is used to fuse and model the relationship between the two types of features and electrical appliances. Finally, load classification is completed based on the fused graph structure representation.

[0042] In another alternative embodiment, the identification of load categories can also employ a multi-scale feature pyramid and attention classifier approach. First, multi-scale information is extracted from image features through a feature pyramid network. Then, key features are screened using a dual spatial and channel attention mechanism. Finally, a prototype classifier is used to calculate the similarity between the sample to be identified and the prototypes of each category, and the load category is determined based on the similarity score.

[0043] In this embodiment of the invention, step S500, which constructs a set of durations for each electrical appliance based on historical identification results and obtains an adaptive threshold for running time using a first clustering algorithm, includes: Collect samples of the operating duration of each electrical appliance within the historical period to form a duration set, and perform robust preprocessing on the duration set to eliminate the impact of outlier data points; The first clustering algorithm is used to perform cluster analysis on the preprocessed duration set to determine the optimal number of clusters. Based on the posterior responsibility degree and prior business knowledge, the data clusters corresponding to normal working conditions are selected from the clustering results. Calculate the adaptive threshold for runtime and add a safety margin based on the distribution characteristics of the normal operating condition data cluster; The runtime adaptive threshold is updated and adjusted online to adapt to changes in user electricity consumption behavior.

[0044] In an optional embodiment, the first clustering algorithm may also employ peak detection and modal clustering based on kernel density estimation. By performing peak detection on the kernel density estimation curve of the duration sample, the sample region corresponding to each significant peak is divided into a cluster. Then, the main peak region is selected as the normal operating condition cluster according to the business logic. This method does not require a preset number of clusters and has no specific assumptions about the data distribution pattern.

[0045] In another optional embodiment, the first clustering algorithm can also introduce a method of combining spectral clustering with silhouette coefficient optimization. First, the distance matrix between samples is converted into a similarity matrix and a Laplacian matrix is ​​constructed for feature decomposition. Then, clustering is performed in the dimensionality reduction space. At the same time, the optimal number of clusters is automatically determined by the silhouette coefficient. Finally, the cluster with the highest internal consistency is selected as the representative of normal working conditions.

[0046] In this embodiment of the invention, steps S100~S400 are performed on the user's historical data to obtain a set of electrical appliances A, and for each electrical appliance... The formula for summarizing duration samples is as follows: Robust preprocessing can be performed. Given a sample duration... (From electrical appliances) Historical events), in each candidate cluster number Next, fit a one-dimensional Gaussian mixture model (GMM): Specifically, EM is used to obtain the maximum likelihood estimate. and maximum likelihood value The logarithmic form is: Specifically, the number of parameters (degrees of freedom): For one-dimensional data, the K-component GMM contains mixed weights with K-1 degrees of freedom, K means, and K variances. The Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) are then expressed as follows: Where n is the sample size (the number of historical events for this appliance).

[0047] Specifically, the selection principles include: Priority ; If the difference in BIC between the two candidates is very small (e.g.) This can be combined with AIC and business priors (number of device physical modes) to assist in decision-making. Intuitively, AIC penalizes model complexity less, favoring a larger K; BIC follows... Increasing the penalty makes it heavier, and it tends to favor a smaller K that is sufficient to explain the data.

[0048] Specifically, based on EM's posterior responsibility Calculate the effective percentage of each cluster: Combination The cluster with the largest proportion and located in the main body of the data is selected as the normal operating condition cluster. If there is a clear business prior (e.g., water heater 20-40 minutes), priority should be given to those falling within the prior interval and... Larger clusters.

[0049] Specifically, the threshold can be given using parameterization or quantile methods: or in, To be assigned to a cluster The sample set. Incorporating a safety margin. (After 2-5 minutes, the following results were obtained:) Specifically, online updating and adaptive adjustment of the runtime adaptive threshold includes: The runtime adaptive threshold is updated on a rolling basis or exponentially smoothed, as expressed by the formula: when (e.g., 20) , temporarily do not fit GMM, and use inter-period quantile or inter-user prior thresholds as temporary values. .

[0050] In this embodiment of the invention, if no paired shutdown event is detected within the adaptive threshold of the running time for the candidate high-power electrical appliance event in step S600, the early warning mechanism is triggered as follows: When an event is detected and identified as a high-power appliance and its activation time is recorded, runtime monitoring is initiated; if If the event is not found, discard it; otherwise, register it in the event table. ); If a paired shutdown event is detected within the runtime adaptive threshold, the event registration record is cleared. If the current time exceeds the runtime adaptive threshold and no shutdown event is detected, a timeout warning is triggered. Based on classification confidence level and timeout duration The timeout warning level is determined by combining the continuous power output; The actual running duration Write back to the historical dataset and periodically update the runtime adaptive threshold to achieve online adaptive updating of the threshold.

[0051] In the above embodiments, this invention achieves accurate detection and intelligent risk warning for high-power electrical appliance access by integrating load recognition technology based on learnable image features with adaptive modeling of time thresholds based on user historical behavior. This invention can automatically learn the differentiated characteristics of appliance operation, significantly improving the accuracy and generalization ability of load recognition; by dynamically establishing operating time thresholds that conform to users' actual electricity consumption habits, it effectively distinguishes between normal use and abnormal overtime operation, greatly reducing the false alarm rate; simultaneously, by establishing an event-driven closed-loop warning mechanism, it can promptly trigger alarms when appliances exceed their time limits, preventing safety hazards caused by prolonged appliance operation from the source, and significantly improving the initiative and reliability of user-side electricity safety management. This invention achieves efficient detection and risk control for high-power electrical appliance access, and has strong practicality and promotional value.

[0052] Example 2, based on the previous example, provides an application example of a non-intrusive high-power electrical appliance access detection and risk warning method to verify and illustrate the technical effects of the method.

[0053] This embodiment simulates two residential units (each with one 220V phase) and covers four types of high-power appliances: electric water heater (EWH, 3.3 kW), electric oven (OVN, 3.5 kW), drum dryer (DRY, 3.6 kW), and central air conditioning outdoor unit (CAC, 4–5 kW). Raw voltage and current waveforms are collected, and active power is calculated. Data was collected continuously for 21 days, and the on / off times were manually checked. The event-level labels were {EWH,OVN,DRY,CAC}. The amount of data after deduplication and anomaly removal is shown in Table 1. The processed data was divided according to the rule of training / validation / test = 60%:20%:20%.

[0054] Table 1: Data volume after deduplication and anomaly removal.

[0055] Specifically, event detection: Only when It enters the high-power channel at that time. Zero-crossing voltage alignment is performed; [the following is taken before startup] , Use a mean template based on the power frequency cycle. Resampling to .

[0056] Specifically, the steps of LGC image feature and classifier processing include: TCN: 3 dilated convolution stacks (dilation factor 1 / 2 / 4), 64 channels, residual connections, and output. (n=25, m=64). 1×1 convolution alignment: LGC: ,element LGC normalization is used as a single-channel grayscale image. 2D-CNN classification head: 5 convolutional layers (3×3, stride 1, BN+ReLU+2×2 pooling), fully connected + Softmax. Training: Adam, lr=1e-3, batch=64, early stopping (patience=10), cross-entropy; class weights are balanced according to training set frequency.

[0057] Specifically, for each appliance Collect historical duration sets .exist Fitting a one-dimensional GMM: parameter ; , ,Pick have to Select the normal operating condition cluster. threshold or and add a safety margin have to If enabled If no shutdown event occurs, an alert is triggered; if the shutdown is normal, the registration is cleared.

[0058] Specifically, evaluation metrics are used to assess recognition performance, including: Accuracy, average F1 score (F1 macro), average precision (P macro), average recall (R macro), Precision, Recall, and F1 score. The formula for issuing alerts is: Specifically, the comparison method used in this embodiment includes: B1: Manual features (25 dimensions including step amplitude / power factor / harmonic energy) + SVM.

[0059] B2: TCN + Fully Connected (no LGC, directly using the temporal average pooling vector of F).

[0060] B3: STFT spectrogram + 2D-CNN (window 256, 50% overlap).

[0061] Abl-1: LGC removes 1×1 alignment (approximates the splicing of m≠n into a non-square matrix and then scales it).

[0062] Abl-2: No voltage zero-crossing alignment.

[0063] The recognition performance on the test set is shown in Table 2, and the event detection results are shown in Table 3. Table 2 shows that LGC utilizes channel-to-channel correlation. The image representation is significantly improved compared to the pure time-series vector (B2) and the spectrum image (B3); 1×1 alignment and voltage zero-crossing alignment both make stable contributions to performance.

[0064] Table 2: Recognition performance on the test set.

[0065] Table 3: Event detection (on / off).

[0066] The results of the adaptive threshold method proposed in this invention compared with the baseline method that uniformly sets a fixed threshold of 45 minutes are shown in Table 4. The adaptive strategy effectively reduces false alarms by modeling differences in user habits, while improving the warning response speed by about 30%.

[0067] Table 4: Comparison Results.

[0068] In this embodiment, the "user" is used as the outer domain in the cross-user generalization test, and a 2-fold LOUO is performed (family A training → family B testing, and vice versa). The F1 macro of LGC is 0.936, which is more robust than B2 (0.902) and B3 (0.915), indicating that the image representation of channel correlation has better robustness to cross-domain differences.

[0069] Furthermore, the water heater learned a threshold of 42 minutes based on historical data. On one occasion, when it ran for 45.1 minutes, the system promptly issued a warning 3.1 minutes later, after which the user confirmed it was an abnormality due to the water heater not being turned off. The dryer's adaptive threshold was 52.3 minutes; a normal 49-minute operation did not trigger a false alarm, whereas a fixed threshold strategy would produce false alarms. These examples demonstrate that the present invention effectively avoids unnecessary false alarm interference while ensuring the accuracy of warnings.

[0070] As can be seen from the above, the present invention achieves accurate detection and intelligent risk warning of high-power electrical appliances by integrating load recognition technology with learnable image features and adaptive modeling based on time thresholds of user historical behavior.

[0071] Example 3: This example provides a non-intrusive high-power electrical appliance access detection and risk warning system, including: The load event detection module is used for differential threshold detection based on the instantaneous power sequence of the main meter side, to identify on-off events and off-off events, and to filter candidate events of high-power appliances with a first preset threshold. The feature extraction module is used to perform voltage zero-crossing phase alignment on the turn-on event, extract the steady-state current waveform before and after the turn-on event, and calculate the event current component. The image feature extraction module is used to input the event current component into a temporal convolutional network for processing to obtain learnable image features; The load feature classification module is used to input learnable image features into a two-dimensional convolutional neural network for classification and output the load category to which the event belongs. The threshold acquisition module is used to construct a set of durations for each electrical appliance based on historical recognition results, and to obtain an adaptive threshold for running time through the first clustering algorithm; The risk warning module is used to trigger a warning mechanism if no paired shutdown event is detected for a candidate event of a high-power electrical appliance within the adaptive threshold of the running time.

[0072] It should be noted that the technical solution of the non-intrusive high-power electrical appliance access detection and risk warning system is based on the same concept as the technical solution of the non-intrusive high-power electrical appliance access detection and risk warning method described above. For details not described in detail in the technical solution of the non-intrusive high-power electrical appliance access detection and risk warning system in this embodiment, please refer to the description of the technical solution of the non-intrusive high-power electrical appliance access detection and risk warning method described above.

[0073] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0074] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a non-intrusive method for detecting and warning of high-power electrical appliance access. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0075] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0076] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0077] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A non-intrusive method for detecting and warning risks associated with high-power electrical appliances connected to the system, characterized in that, include: Based on differential threshold detection of the instantaneous power sequence on the main meter side, the system identifies on-start and off events, and filters candidate events for high-power appliances using a first preset threshold. The voltage zero-crossing phase is aligned for the turn-on event, the steady-state current waveforms before and after the turn-on event are extracted, and the event current component is calculated. The event current components are input into a temporal convolutional network for processing to obtain learnable image features. The learnable image features are input into a two-dimensional convolutional neural network for classification, and the load category to which the event belongs is output. Based on the historical identification results, a set of durations for each electrical appliance is constructed, and an adaptive threshold for running time is obtained through the first clustering algorithm; If no paired shutdown event is detected for the candidate high-power electrical appliance within the adaptive operating time threshold, an early warning mechanism is triggered.

2. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 1, characterized in that, The step of filtering candidate events for high-power electrical appliances using a first preset threshold includes: Voltage and current sequences are collected from the user's main meter terminal to calculate the instantaneous power sequence; Perform sliding difference calculation on the instantaneous power sequence to detect the occurrence time of the load switching event; When the result of the sliding differential operation exceeds the set event detection threshold, a load switching event is determined to exist; If the steady-state power corresponding to the load switching event exceeds the first preset threshold, the load switching event is recorded as a candidate event for high-power appliances; otherwise, it is discarded. Based on the power step conservation principle, the turn-on event and the turn-off event are paired in the time neighborhood to obtain a complete event pair or an independent turn-on event marked as pending confirmation of turn-off.

3. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 2, characterized in that, The calculation of the event current component includes: Select multiple power frequency cycles prior to the occurrence of the start-up event to calculate the steady-state baseline current; Select multiple power frequency cycles after the start-up event occurs, and calculate the steady-state current after the event; The event current component is calculated based on the steady-state baseline current and the post-event steady-state current. The event current component is then subjected to a first processing step to generate a standardized event current waveform.

4. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 3, characterized in that, The step of inputting the event current component into a temporal convolutional network for processing to obtain learnable image features includes: The standardized event current waveform is input into a time-series convolutional network to extract the time-series feature matrix; The temporal feature matrix is ​​aligned by channel number through pointwise convolution to obtain the aligned feature matrix. Based on the aligned feature matrix, calculate the correlation matrix of different feature channels on the time axis; The correlation matrix is ​​normalized into a single-channel grayscale image, which is then output as a learnable image feature.

5. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 4, characterized in that, The load categories to which the output events belong include: The learnable image features are input into a multi-layer two-dimensional convolutional neural network, and the learnable image features are processed layer by layer through batch normalization, non-linear activation and pooling operations. The extracted image features are input into a fully connected layer, processed by the Softmax function to obtain the probability distribution of the load categories, and the load category with the highest probability is selected as the recognition result, and the corresponding classification confidence is output.

6. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 5, characterized in that, The step of obtaining the adaptive threshold for runtime through the first clustering algorithm includes: Collect samples of the operating duration of each electrical appliance within a historical period to form a duration set. Perform robust preprocessing on the duration set to eliminate the influence of outlier data points. The first clustering algorithm is used to perform cluster analysis on the preprocessed duration set to determine the optimal number of clusters. Based on the posterior responsibility degree and prior business knowledge, the data clusters corresponding to normal working conditions are selected from the clustering results. Calculate the adaptive threshold for runtime and add a safety margin based on the distribution characteristics of the normal operating condition data cluster; The runtime adaptive threshold is updated and adjusted online to adapt to changes in user electricity consumption behavior.

7. The non-intrusive high-power electrical appliance access detection and risk warning method as described in claim 6, characterized in that, The triggering early warning mechanism includes: When an event is detected and identified as a high-power appliance and the start time is recorded, runtime monitoring is initiated; if a paired shutdown event is detected within the runtime adaptive threshold, the event registration record is cleared; if the current time exceeds the runtime adaptive threshold and no shutdown event is detected, a timeout warning is triggered. The timeout warning level is determined by comprehensively considering classification confidence level, timeout duration, and continuous power. The actual runtime duration is written back to the historical dataset, and the runtime adaptive threshold is periodically updated.

8. A non-intrusive high-power electrical appliance access detection and risk warning system, employing the non-intrusive high-power electrical appliance access detection and risk warning method as described in any one of claims 1 to 7, characterized in that, include: The load event detection module is used for differential threshold detection based on the instantaneous power sequence of the main meter side, to identify on-off events and off-off events, and to filter candidate events of high-power appliances with a first preset threshold. The feature extraction module is used to perform voltage zero-crossing phase alignment on the turn-on event, extract the steady-state current waveform before and after the turn-on event, and calculate the event current component. The image feature extraction module is used to input the event current component into a temporal convolutional network for processing to obtain learnable image features. The load feature classification module is used to input the learnable image features into a two-dimensional convolutional neural network for classification and output the load category to which the event belongs. The threshold acquisition module is used to construct a set of durations for each electrical appliance based on historical recognition results, and to obtain an adaptive threshold for running time through the first clustering algorithm; The risk warning module is used to trigger a warning mechanism if no paired shutdown event is detected for the candidate event of the high-power electrical appliance within the adaptive threshold of the running time.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the non-intrusive high-power electrical appliance access detection and risk warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the non-intrusive high-power electrical appliance access detection and risk warning method according to any one of claims 1 to 7.

Images