Low-frequency non-intrusive load monitoring method and system based on adaptive event detection

Abstract

The application discloses a low-frequency non-intrusive load monitoring method and system based on adaptive event detection, and the method comprises the following steps: event detection, feature extraction and load identification; in the event detection, the Bayesian information criterion is used as a detection window, and the window threshold is adaptively optimized through a power variable point weight model; the power time sequence of different electrical equipment is decomposed and grouped for load features by using a variational mode decomposition method; and according to a load identification model, the load identification and classification of the power curve waveform diagram of different electrical equipment are realized. The method disclosed by the application guarantees high precision of event detection and also shows high load identification accuracy under the condition of low-frequency sampling.

Description

Technical Field

[0001] This invention relates to a low-frequency non-intrusive load monitoring method and system based on adaptive event detection, belonging to the field of smart grids. Background Technology

[0002] With the ever-increasing energy demand and growing focus on energy efficiency in modern society, technological and industrial transformation is accelerating towards informatization, automation, and intelligence. The development of the power Internet of Things (IoT) places higher demands on power system sensing technologies. For intelligent management of user demand side, load monitoring is a crucial component. Traditional load monitoring technologies can no longer support the development of digital management on the user demand side, while NILM technology can accurately detect and identify abnormal load events without interrupting power operation or infringing on the privacy of residential users. This method also features high precision, low cost, and high efficiency, effectively reducing the maintenance costs of the power system, enabling online monitoring of residential user load data, and providing refined energy efficiency analysis for the interactive management between power grid companies and the residential demand side. This guides power companies to optimize the power supply and distribution structure, plan low-carbon and energy-saving electricity consumption patterns for users, and improve the energy efficiency of the power system.

[0003] In recent years, the construction of home energy management systems and research on load monitoring have attracted widespread attention from relevant scholars, and various intelligent residential load monitoring methods based on NILM technology have gradually emerged. At present, event detection-based NILM algorithms have made some progress in the field of residential load monitoring. In the field of high-frequency NILM, the literature "Lin Shunfu, Lin Yifeng, Li Yi, et al. Research on non-intrusive load identification technology based on high-frequency load feature imagery [J / OL]. Power System Technology: 1-11 [2024-04-03]" uses Gram angle field algorithm and Markov transfer field algorithm to add periodic steady-state current characteristics and periodic instantaneous active power characteristics to the image based on the grayscale image of UI characteristic curve, which improves the accuracy of load identification. However, this method is computationally complex and has high requirements for the measurement dimensions of residential load electrical quantity data. The literature “Wang Ying, Yang Wei, Xiao Xianyong, et al. Non-intrusive residential load monitoring method based on UI trajectory curve refined identification [J]. Power System Technology, 2021, 45(10): 4104-4113” uses goodness-of-fit test to capture appliance switching events, and then uses the k-means algorithm considering initial optimization and the convolutional neural network model to perform two-stage identification of the load, respectively, and proposes an event-based NILM method based on UI trajectory curve refined identification. The above methods improve the intelligence and refinement of event detection NILM methods to a certain extent. However, these methods exhibit significant computational complexity and impose very high sampling requirements (not only requiring a high sampling frequency, but also higher hardware performance requirements for NILM acquisition devices), thus hindering the low-cost and high-efficiency deep integration of data-driven high-precision event detection methods with NILM load identification terminals.

[0004] With the diversification of household appliances and their functions, residents' electricity environment is becoming increasingly complex. How to ensure high-precision NILM load identification while solving problems such as high equipment cost and complex calculation methods caused by high-frequency sampling has become a challenging research task in the field of load monitoring.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] This invention provides a low-frequency non-intrusive load monitoring method and system based on adaptive event detection, which can perform event detection under low-frequency sampling conditions to reduce sampling costs and computational complexity.

[0007] The technical solution of this invention is:

[0008] According to a first aspect of the present invention, a low-frequency non-intrusive load monitoring method based on adaptive event detection is provided, comprising: event detection, feature extraction, and load identification; in event detection, a Bayesian information criterion is used as the detection window, and the window threshold is adaptively optimized through a power variable point weight model; the power time series of different electrical devices is decomposed and grouped using a variational mode decomposition method; and the load identification and classification of the power curve waveforms of different electrical devices are realized based on the load identification model.

[0009] During event detection, it is assumed that the time series of the measured power of electrical equipment follows a Poisson distribution. Based on this, the following two models are constructed, and the maximum likelihood function in the Bayesian information criterion is constructed according to these two models:

[0010] No events occurred in the power time series, so a static model M was constructed. s :

[0011] M s :p1,.....,p N ~P S (λ0);

[0012] A jump model M is constructed when load events occur in the power time series. J :

[0013] M J :p1,.....,p n-1 ~P S (λ1); p n ,.....,p N ~P S (λ2);

[0014] In the formula, λ0 represents the power time series P where no event occurs. S (λ0) follows a Poisson distribution in terms of expectation and variance; λ1 and λ2 represent the power time series P before the occurrence of the load event in the power time series. S (λ1) represents the expectation and variance of a Poisson distribution; λ2 represents the power time series P following a load event in the power time series. S (λ2) follows a Poisson distribution in terms of expectation and variance; p n This represents the marker point when a load event occurs, i.e., the nth sample in the power time series.

[0015] The adaptive optimization of the window threshold using a power variable point weight model includes:

[0016] In the power-change-point-weighted model, the jump model M J Transform the sequence into matrix form P s and P J :

[0017] static sequence P s Represented as:

[0018] P s =[p1,…,p n-1 ] T ;

[0019] Jump sequence P J Represented as:

[0020] P J =[p n ,…,p N ] T ;

[0021] In a static sequence, no events occur, so the power increment Δp0 corresponding to the sequence is zero; a transition sequence, however, is accompanied by a change in the state of the electrical equipment, so the power increment Δp0 corresponding to the sequence is not zero; δ is introduced. min This represents the minimum power increment required for switching the state of an electrical device; the power increment is described as follows:

[0022] Δp=Δp0-δ min ;

[0023] The power increment set is then:

[0024] Δp=[Δp1,…,Δp N ] T ;

[0025] Δp S =[Δp1,…,Δp n-1 ] T ;

[0026] Δp J =[Δp n ,…,Δp N ] T ;

[0027] In the formula, Δp N Δp represents the increment of the Nth sample in the power time series. S and Δp J These are the static power increment set and the transition power increment set, respectively; the power increment set Δp of the sequence to be detected is:

[0028] Δp=Δp S +Δp J =Δp J ;

[0029] Jump weights ω in the power variable point weight model i The assignment formula and the formula for calculating the jump information entropy F are as follows:

[0030] ω i =lg(Δp) i ); i∈(n,N);

[0031]

[0032] In the formula, N2 is the number of samples in the jump sequence;

[0033] By combining the Bayesian information criterion, the detection window threshold h is optimized.

[0034] The load recognition model is the CNN-SENet load recognition model, which introduces the squeeze-excitation attention mechanism into the convolutional neural network.

[0035] The CNN-SENet load identification model is specifically structured as follows: The model consists of one input layer, five convolutional layers, one pooling layer, one fully connected layer, and one output layer. The first four convolutional layers are standard 5x5 convolutions, while the fifth convolutional layer uses a 6x6 kernel with the SENet spatial attention module. The pooling layer performs max pooling and normalization on a 5x5 numerical matrix, operating in "SAME" mode. The output layer uses cross-entropy to construct the model's loss function, and Adam is selected as the gradient optimizer to dynamically update the CNN network parameters.

[0036] According to a second aspect of the present invention, a low-frequency non-intrusive load monitoring system based on adaptive event detection is provided, comprising: an event detection module, a feature extraction module, and a load identification module; the event detection module uses a Bayesian information criterion as a detection window and adaptively optimizes the window threshold using a power variable point weight model; the feature extraction module is used to decompose and group the load characteristics of the power time series of different electrical devices using a variational mode decomposition method; the load identification module is used to identify and classify the load of the power curve waveforms of different electrical devices according to the load identification model.

[0037] According to a third aspect of the present invention, a processor is provided that, when running, executes the low-frequency non-intrusive load monitoring method based on adaptive event detection as described above.

[0038] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium comprising a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the low-frequency non-intrusive load monitoring method based on adaptive event detection as described above.

[0039] The beneficial effects of this invention are as follows: Based on change-point detection, this invention establishes a power change-point weight model to achieve adaptive adjustment of BIC algorithm parameters, ensuring that the proposed algorithm has good generalization performance and excellent detection accuracy in different operating environments of electrical appliances. It introduces the VMD algorithm to perform multimodal decomposition and effective combination of features from the power curve waveform. Based on this, a CNN is used as an image classifier for the power curve waveforms of different household appliances, and the SENet attention mechanism is employed to enhance the CNN's feature aggregation capability. This not only improves the efficiency of the learning algorithm but also enhances the model's generalization ability and interpretability, enabling accurate identification of different household loads under different operating scenarios. Numerical examples show that the proposed method, under low-frequency sampling conditions, ensures high accuracy in event detection while also exhibiting high load identification accuracy. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0041] Figure 2 This is a block diagram of the event detection of the present invention;

[0042] Figure 3 This is a feature extraction flowchart of the present invention;

[0043] Figure 4 This is a block diagram of the load identification of the present invention;

[0044] Figure 5 This is a diagram of a transient event. Detailed Implementation

[0045] The invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited to the description.

[0046] Example 1: As Figure 1-5 As shown, according to a first aspect of the present invention, a low-frequency non-intrusive load monitoring method based on adaptive event detection is provided, comprising: event detection, feature extraction, and load identification; in event detection, a Bayesian information criterion is used as the detection window, and the window threshold is adaptively optimized through a power variable point weight model; secondly, the power time series of different electrical devices is decomposed and grouped using a variational mode decomposition method to remove redundant and irrelevant features of event loads; and based on the load identification model, load identification and classification of the power curve waveforms of different electrical devices are realized.

[0047] Furthermore, such as Figure 1The diagram illustrates the process of key components based on NILM technology. Event detection aims to automatically identify and label start-up and shutdown events of electrical appliances from electrical data sequences; feature extraction aims to mine differentiated characteristics of electrical quantities such as voltage, current, and power of different appliances to achieve good distinguishability; load identification aims to determine the type of each electrical device based on the load characteristics obtained from event detection and decomposition, providing technical support for users' energy consumption monitoring and energy optimization management. The event detection process of this invention is as follows: Figure 2 As shown, the principle of feature extraction is as follows: Figure 3 As shown, the principle of load identification is as follows: Figure 4 As shown.

[0048] In event-based NILM load identification, the Bayesian Information Criterion (BIC) is introduced as a probabilistic model for change point detection to monitor changes in statistical parameters of power sequences from different devices. Thus, the problem of load event detection is transformed into the problem of time series anomaly detection, i.e., continuously monitoring change points in the time series and marking moments when significant data changes occur. The mathematical model of the BIC algorithm is as follows:

[0049] Let the power sequence of a certain electrical device over time be P = {p t Let t = 1, 2, 3...N, and M be the probabilistic model for change point detection. A maximum likelihood function L(p,M) is proposed, and then a formula for calculating the BIC of model M is derived:

[0050] BIC(M)=-2log(L(p,M))+k*log(N);

[0051] In the formula, k represents the number of model parameters; N is the number of samples; k*log(N) is a penalty term introduced by BIC related to the number of model parameters. Its function is to avoid overfitting, that is, to ensure that the model selection problem finds the optimal balance between model complexity and the model's ability to describe the power sequence dataset (i.e., the likelihood function L). Generally, as model complexity increases (k increases), the likelihood function L also increases, thereby reducing the value of BIC and keeping it within a small range of variation. When model complexity continues to increase (k exceeds a certain value), the rate of increase in L will not effectively suppress the increase in the BIC value, at which point the model is prone to overfitting. The penalty term of BIC is not only directly related to k, but also considers the number of samples N. This is so that when the training data is large enough, the penalty term can effectively prevent excessive model complexity while improving model accuracy, reducing the possibility of overfitting.

[0052] During event detection, it is assumed that the time series of the measured power of electrical equipment follows a Poisson distribution. Based on this, the following two models are constructed, and the maximum likelihood function in the Bayesian information criterion is constructed according to these two models:

[0053] No events occur in the power time series, i.e., there is no static model M of the signal with a changing point. s :

[0054] M s :p1,.....,p N ~P S (λ0);

[0055] The power time series shows load events and a step signal jump model M. J :

[0056] M J :p1,.....,p n-1 ~P S (λ1); p n ,.....,p N ~P S (λ2);

[0057] In the formula, λ0 represents the power time series P where no event occurs. S (λ0) follows a Poisson distribution in terms of expectation and variance; λ1 and λ2 represent the power time series P before the occurrence of the load event in the power time series. S (λ1) represents the expectation and variance of a Poisson distribution; λ2 represents the power time series P following a load event in the power time series. S (λ2) follows a Poisson distribution in terms of expectation and variance; p n This indicates the marker point at which a load event occurs, further dividing the MJ sequence into two parts: p1, ..., p n-1 (Sample size denoted as N1) and p after the event occurs n ,.....,p N (The sample size is denoted as N2).

[0058] Next, we construct the maximum likelihood functions for the two models:

[0059] L(i)=Nlog(λ0)-N1log(λ1)-N2log(λ2);

[0060] Its corresponding BIC value is:

[0061]

[0062] In the formula, λ is λ0, λ1 or λ2;

[0063] The criteria for change point detection are derived as follows:

[0064] {max(BIC(i))}≥h;

[0065] In the formula, h represents the threshold parameter of the BIC detection window. When the BIC value of a certain detection window is greater than the set threshold h, it indicates that the power sequence of that window corresponds to the jump model M. J Using the maximum likelihood estimation method, the event times in this power sequence can be further determined as follows:

[0066]

[0067] The accuracy of the BIC algorithm largely depends on the value of the threshold parameter h. Therefore, this invention introduces the PCW model to evaluate the algorithm's performance when different thresholds are set, and maximizes the algorithm's event detection performance under different operating environments by dynamically optimizing the threshold h. The modeling approach of PCW is highly compatible with the mathematical model of BIC, both dividing the power curve of electrical equipment into two states: "static" and "jumping," and modeling them separately.

[0068] Furthermore, the adaptive optimization of the window threshold using the power variable point weight model includes:

[0069] In the power-change-point-weighted model, the jump model M J Transform the sequence into matrix form P s and P J :

[0070] static sequence P s Represented as:

[0071] P s =[p1,…,p n-1 ] T ;

[0072] Jump sequence P J Represented as:

[0073] P J =[p n ,…,p N ] T ;

[0074] In a static sequence, no events occur, so the corresponding power increment Δp0 is zero. However, a transition sequence involves a change in the state of the electrical equipment, so the corresponding power increment Δp0 is not zero. It is important to note that the power curve of the electrical equipment contains measurement noise. To avoid false event detection, δ is introduced. min This represents the minimum power increment required for the state transition of electrical equipment; therefore, the power increment can be further described as:

[0075] Δp=Δp0-δ min ;

[0076] The power increment set is then:

[0077] Δp=[Δp1,…,Δp N ] T ;

[0078] Δp S =[Δp1,…,Δp n-1 ] T ;

[0079] Δp J =[Δp n ,…,Δp N ] T ;

[0080] In the formula, Δp N Δp represents the increment of the Nth sample in the power time series. S and Δp J These are the static power increment set and the transition power increment set, respectively; clearly, the static sequence has no transition events, and the corresponding power increments are all less than δ. min Then we have Δp S =0. Therefore, the power increment set Δp of the sequence to be detected has:

[0081] Δp=Δp S +Δp J =Δp J ;

[0082] At this point, the PCW model initializes each transition event as a weightless and volumeless particle in the weight space. These particles contain different transition information of the electrical device, and the sum of the transition information contained in all particles is denoted as the transition information entropy F. Clearly, different particles contribute differently to F. This paper quantifies the particle's contribution to F as the transition weight ω, meaning that the larger the power transition containing transient information in the particle, the larger the particle's transition weight ω, and the larger its contribution to F. The transition weight ω in the power transition weight model... i The assignment formula and the formula for calculating the jump information entropy F are as follows:

[0083] ω i =lg(Δp) i ); i∈(n,N);

[0084]

[0085] Finally, by combining the Bayesian information criterion, the detection window threshold h is optimized. Under different thresholds, the magnitude of the jump information entropy is compared. The larger the jump information entropy F, the better the BIC algorithm performs in event detection. Thus, the dynamic optimization of the BIC window threshold parameter h can be achieved.

[0086] Furthermore, after event detection is completed, this invention introduces the VMD algorithm and combines it with the CNN-SENet model for feature extraction and load identification, such as... Figure 3 , 4 As shown, the extraction and identification of power curve waveform features belong to image target recognition and classification. The VMD algorithm is introduced to perform multimodal decomposition and effective combination of power curve waveform features. On this basis, CNN is used as an image classifier for power curve waveforms of different household appliances, and the SENet attention mechanism is adopted to enhance the feature aggregation capability of CNN. This not only improves the efficiency of the learning algorithm, but also enhances the generalization ability and interpretability of the model, enabling accurate identification of different household loads under different operating scenarios.

[0087] Compared to wavelet transform or empirical mode decomposition (EMD), VMD offers better feature resolution. Furthermore, unlike EMD, which lacks mathematical background and is sensitive to noise, VMD is built on a solid mathematical foundation, including Wiener filters for signal denoising, Hilbert transforms for generating one-sided analytic signals, and frequency shifts for identifying baseband through mode mixing. VMD is an adaptive, fully non-recursive modal variational and signal processing method. Its adaptability is demonstrated by determining the number of modal decompositions for the power time series based on the actual electrical quantities of different devices, and automatically determining the optimal frequency band for each modal signal in subsequent signal decomposition steps. This decomposes the original signal into a discrete number of multi-order intrinsic mode functions (IMFs) with specific sparse properties, achieving effective decomposition of event device characteristics in the power series.

[0088] Each IMF obtained from eigenvalue decomposition is an amplitude and frequency modulated signal concentrated around the center frequency. Feature combination is needed to filter out the effective decomposition components of the power sequence, reconstruct the event device features, and thus obtain the optimal solution to the variational problem. First, signal analysis is performed using Hilbert transform to convert the one-dimensional power sequence signal into a two-dimensional signal with amplitude and phase on the complex plane. Second, Gaussian smoothing is used to demodulate the signal and estimate the bandwidth of each mode signal. Then, in each iteration, Lagrange multipliers and quadratic penalty functions are used to strengthen the reconstruction constraints based on the signal's spectral characteristics to improve the convergence of the results. Finally, L2 norm regularization is used to address the overfitting problem. For each different device power signal, different decomposition modes and center frequencies are updated to complete the effective extraction of event device features.

[0089] Furthermore, the load identification model considers the enhancement of spatial feature aggregation capabilities by the Squeeze-and-Excitation Network (SENet) attention mechanism, and introduces it into the CNN-SENet load identification model of a Convolutional Neural Network (CNN) to achieve load identification and classification of power curve waveforms of different electrical devices. Experimental results show that the proposed method has high identification accuracy and good generalization performance under different operating scenarios.

[0090] To complete the feature aggregation of convolution operations and further achieve refined classification output in the output layer, the CNN-SENet load identification model constructed in this invention realizes load identification and classification of power curve waveforms of different electrical devices, such as... Figure 1 As shown, specifically:

[0091] Standardization of power curve waveforms: The power curve waveforms are preprocessed and the window size is normalized; for power curve waveforms of different household appliance loads, they are uniformly adjusted to (256, 256, 3) JPG format, that is, the image pixel size is 256×256, and 3 feature map channels are used.

[0092] The CNN-SENet load recognition model is constructed as follows: The model consists of one input layer, five convolutional layers, one pooling layer, one fully connected layer, and one output layer. The first four convolutional layers are standard 5x5 convolutions; specifically, the fifth convolutional layer uses a 6x6 convolutional kernel with SENet spatial attention. The pooling layer performs max pooling and normalization on a 5x5 digit matrix using the "SAME" operation mode. The output layer uses cross-entropy to construct the model's loss function, and Adam is selected as the gradient optimizer to dynamically update the CNN network parameters.

[0093] Training and testing the classification model: In the proposed method, the training and testing sets are configured in an 8:2 ratio to train and test the classification model. This step is a semi-supervised machine learning training process, optimizing the parameters of the CNN-SENet model with manually assisted labeled data to achieve the goal of identifying the load of different household appliances.

[0094] Based on the above, the NILM load identification method based on adaptive event detection of the present invention is presented in the following steps:

[0095] 1) Data Acquisition: Analyze the differentiated electrical characteristics of different electrical appliances in residential buildings, and measure and collect power data of electrical appliances in different operating states;

[0096] 2) Event detection: Establish a power time series-based curve waveform database and use the PCW-optimized BIC algorithm to detect events and mark event times;

[0097] 3) Feature extraction: The VMD algorithm is used to perform multimodal decomposition and effective reconstruction of the power curve waveform of the event load to extract typical features of different electrical devices;

[0098] 4) Load identification: Construct a load identification model based on CNN-SENet, learn the load characteristic values ​​of different electrical devices through iterative training, and complete the identification and classification of loads on the test set data;

[0099] 5) Evaluation and optimization: After identification by the CNN-SENet model, the load labels are output, and the test results are compared with the true values ​​to evaluate and optimize the load recognition performance of the proposed method.

[0100] Further explanation is provided below, based on a calculation example:

[0101] To verify the feasibility and superiority of the proposed non-intrusive load monitoring algorithm, two computational examples were designed. Example 1 utilizes the proposed BIC-PCW adaptive event detection algorithm to perform event detection on residential load monitoring data, demonstrating that the BIC-PCW algorithm can meet the event detection requirements of different residential loads through threshold adaptation, and comparing it with existing event detection algorithms. Example 2, based on Example 1, further extracts the power waveforms before and after the event, performs mode decomposition using VMD, and uses the decomposition results as training samples input into the SENet-CNN network to complete residential load identification. The hardware environment for the examples is a ThinkPad T14, CPU 12th i7-1260P, 2.10GHz, and the software simulation environment is PyCharm (Professional Edition), Python 3.9. The power time series obtained by low-frequency sampling is selected as the electrical quantity input of electrical equipment. Specifically, in this invention, the data comes from a simulation experimental environment built by a certain electricity meter company. The USB A / D data acquisition card is used to collect current and voltage data from six types of equipment, namely electric oven, electric kettle, fixed frequency air conditioner, variable frequency air conditioner, microwave oven and induction cooker, at a sampling frequency of 1Hz. Then, the active power of the equipment is obtained by calculating the effective value of the cycle according to the following formula.

[0102]

[0103] In the formula: u i For voltage sample values; i i is the current sample value; m is the periodic sampling point.

[0104] Both event detection and load identification are essentially classification problems. Event detection, by monitoring load power fluctuations, transforms the problem of whether a transient event has occurred into a binary classification problem, that is, determining whether the current monitored state is transient or steady-state. Load identification, on the other hand, transforms the identification task into a multi-class classification problem based on equipment characteristics, that is, determining which type of equipment characteristic it belongs to.

[0105] Therefore, this invention utilizes the evaluation method of classifiers in machine learning as the evaluation standard for simulation experiments.

[0106]

[0107]

[0108] In the formula, precision and recall are precision and recall, respectively. TP (True Position) represents the number of correctly classified positive samples; FP (False Position) represents the number of misclassified positive samples; and FN (False Negative) represents the number of misclassified negative samples. It is easy to see that the difference between precision and recall lies in FP and FN, meaning that they influence each other. Generally, improving precision often reduces recall. Therefore, a single evaluation metric cannot comprehensively reflect the classifier's performance. The following is a new classifier evaluation metric based on the above two indicators: the f1 score. The specific calculation method is shown in the following formula:

[0109]

[0110] In the formula, the F1 score comprehensively considers the impact of precision and recall on the classifier's performance evaluation, achieving a relative balance between the two and providing a more comprehensive reflection of the classifier's performance. A higher F1 score indicates better classifier performance.

[0111] Event detection, based on statistical theory, is the process of determining whether the operating state of equipment has changed by monitoring load power fluctuations. The entire transient event should include continuous jumps in equipment power; that is, the transient event begins with a sudden power change after the equipment has reached steady-state operation and ends with a new steady state after the change. For example... Figure 5 As shown, this invention defines the continuous change process between a power surge and the device entering another steady state as a complete transient event.

[0112] Then, the labeled samples are input into the BIC-PCW event detection algorithm, and the performance of BIC-PCW event detection is measured using the evaluation metrics mentioned above. Finally, the method of this invention is compared with existing literature to verify the superiority of the proposed method.

[0113] The BIC-PCW algorithm was used to detect events in laboratory data collected over a specific period. A total of 73 transient events occurred during this period, each consisting of 2-3 transient points, totaling 153 transient points across six types of equipment. The simulation experiment used transient points as the basic unit; the algorithm was considered to have successfully detected the event if it detected any one of the transient points, and this was used as the standard to measure the algorithm's performance. BIC-PCW uses transient information entropy as the objective function and iteratively adjusts the BIC threshold to adapt to different types of equipment. Table 1 shows the event detection results of BIC for the six types of equipment. The results indicate that the BIC-PCW algorithm performs excellently, with F1 scores for all six types of equipment exceeding 90%. In particular, resistive loads, which convert electrical energy into heat and consume a large amount of electricity, exhibit more pronounced power jumps in their transient events, making them easier to detect. Therefore, resistive loads such as electric kettles, electric ovens, and induction cookers all achieved F1 scores above 95%.

[0114] Table 1 BIC-PCW Event Detection Results

[0115] Load Category number of events precision recall <![CDATA[f1 score]]> electric oven 18 97.2％ 97.2％ 97.2％ electric kettle 2 100％ 100％ 100％ Fixed frequency air conditioner 22 97.6％ 85.1％ 90.9％ Inverter air conditioner 17 93.9％ 88.6％ 91.2％ Micro-wave oven 3 100％ 85.7％ 92.3％ induction cooker 11 95.7％ 95.7％ 95.7％ total 73 96.5％ 90.8％ 93.6％

[0116] To verify the superiority of the algorithm, the proposed algorithm was compared with CUSUM, GOF, and GRLT using power waveform data collected in the laboratory, and the event detection performance of different algorithms on six types of devices was evaluated by the f1 score. Table 2 shows the comparison results of the four algorithms. The experimental results show that the proposed algorithm outperforms the other three event detection algorithms in terms of event detection f1 score, both in single devices and overall across all devices.

[0117] Table 2 Comparison of Event Detection Results of Different Algorithms

[0118] BIC-PCW CUSUM GOF GRLT electric oven 97.20％ 88.20％ 84.80％ 69.80％ electric kettle 100％ 88.90％ 80.00％ 72.70％ Fixed frequency air conditioner 90.90％ 83.50％ 76.70％ 69.00％ Inverter air conditioner 91.20％ 81.20％ 75.00％ 68.80％ Micro-wave oven 92.30％ 80.00％ 71.40％ 66.70％ induction cooker 95.70％ 79.10％ 73.20％ 69.80％ total 93.60％ 83.40％ 77.60％ 69.30％

[0119] After the residential load undergoes event detection, it is necessary to further extract the steady-state power waveform characteristics before and after the transient event. Mode decomposition (VMD) is then used to perform mode decomposition on the raw power data. The first two intrinsic mode signals of the load, the final trend waveform, and the raw power waveform are used as inputs to the CNN-SENet model. Current and voltage waveform data for six types of residential equipment are collected over 12 hours, based on... The current and voltage are converted into active power, and the equipment data is manually labeled as samples to train the CNN-SENet load identification model. It is known that a total of 1856 transient events occurred during the data sample collection process, that is, the effective number of dataset samples is 1856 groups, which are randomly divided into training set and test set in a ratio of 8:2. In order to verify the superiority and effectiveness of the proposed method, the load identification method of the present invention is further compared with existing literature, and the performance of different load identification algorithms in 6 types of equipment is analyzed. Reference [1] uses the generalized likelihood ratio algorithm to detect equipment transient events, and then uses CEEMD to extract two-dimensional images of power sequence to train the CNN model. Reference [2] proposes a two-stage load identification method based on CNN according to the transient current amplitude characteristics. Reference [3] proposes an event detection method based on low-frequency power difference features, and uses it to train a Bi-LSTM load identification network (Reference [1]: Feng Changsen, Liu Pan, Wang Jiaying, et al. Non-intrusive load monitoring algorithm for residential users using limited low-frequency information [J]. Electric Power Automation Equipment, 2023, 43(11): 181-187. Reference [2]: Huang Youjin, Xiong Wei, Yuan Xufeng, et al. Low-frequency sampling non-intrusive load identification algorithm that integrates deep learning and amplitude features [J]. Electric Power Science and Engineering, 2020, 36(04): 10-16. Reference [3]: Zhou Buxiang, Zhao Wenwen, Zang Tianlei, et al. Non-intrusive load monitoring method based on low-frequency power difference features and dual long short-term memory network [J]. Electric Power Automation Equipment, 2023, 43(08): 167-173+209.). Table 3 shows the load identification results of different algorithms. It is easy to see that the method of the present invention outperforms the other algorithms in the identification performance of all six types of loads, with an F1 score of 93.4% for all monitored devices. Among them, the fixed-frequency air conditioner has the highest load identification rate, with an F1 score of 96.7%, while the electric oven has the lowest load identification rate, with an F1 score of only 89.3%.

[0120] Table 3 Load identification results of different algorithms

[0121]

[0122] As can be seen from the above technical solutions, traditional non-intrusive load monitoring algorithms rely on high-frequency data for accuracy, leading to high sampling equipment costs and complex calculation methods. This invention proposes a low-frequency non-intrusive load monitoring algorithm based on adaptive event detection. This method is based on the NILM algorithm framework for event detection and proposes a two-stage load detection approach. First, an adaptive event detection model is established based on PCW and BIC. This model does not require manual threshold adjustment and can automatically iterate and optimize according to equipment type to achieve higher accuracy in event detection. Second, the VMD algorithm is used to decompose and reorganize load features, filtering out redundant and irrelevant features. Finally, a load identification model is constructed using SENet-CNN, and the proposed method is compared with existing NILM algorithms for low-frequency data. Experimental results show that the proposed method performs well in different operating environments. The simulation examples involving six types of equipment all achieved load identification f1 scores above 90%, demonstrating that the proposed method has excellent event detection performance and load identification rate under low-frequency sampling conditions. This also indicates that this invention is more universally applicable in high-frequency environments containing richer load event information.

[0123] According to a second aspect of the present invention, a low-frequency non-intrusive load monitoring system based on adaptive event detection is provided, comprising: an event detection module, a feature extraction module, and a load identification module; the event detection module uses a Bayesian information criterion as a detection window and adaptively optimizes the window threshold using a power variable point weight model; the feature extraction module is used to decompose and group the load features of the power time series of different electrical devices using a variational mode decomposition method; the load identification module is used to identify and classify the load of the power curve waveforms of different electrical devices according to the load identification model.

[0124] According to a third aspect of the present invention, a processor is provided that, when running, executes the low-frequency non-intrusive load monitoring method based on adaptive event detection as described above.

[0125] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the low-frequency non-intrusive load monitoring method based on adaptive event detection as described above.

[0126] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A low-frequency, non-intrusive load monitoring method based on adaptive event detection, characterized in that, include: Event detection, feature extraction, and load identification; In event detection, the Bayesian information criterion is used as the detection window, and the window threshold is adaptively optimized through the power variable point weight model. The variational mode decomposition method is used to decompose and group the load characteristics of the power time series of different electrical devices. Based on the load identification model, the load identification and classification of the power curve waveforms of different electrical devices are realized. During event detection, it is assumed that the time series of the measured power of electrical equipment follows a Poisson distribution. Based on this, the following two models are constructed, and the maximum likelihood function in the Bayesian information criterion is constructed according to these two models: No events occurred in the power time series, so a static model M was constructed. s : ； A jump model M is constructed when load events occur in the power time series. J : ； In the formula, Power time series representing no events occurring The expected value and variance follow a Poisson distribution; , This represents the power time series preceding the occurrence of a load event in a power time series. The expected value and variance follow a Poisson distribution; Represents the power time series following the occurrence of a load event in a power time series. The expected value and variance follow a Poisson distribution; This represents the marker point when a load event occurs, i.e., the nth sample in the power time series; The adaptive optimization of the window threshold using a power variable point weight model includes: In the power-change-point weighted model, the jump model M J Transform the sequence into matrix form and : static sequence Represented as: ； Jump sequence Represented as: ； No events occur in a static sequence, therefore the power increment corresponding to the sequence is... The value is zero; a transition sequence is accompanied by a change in the state of the electrical equipment, so the power increment corresponding to the sequence is zero. Not zero; introduce This represents the minimum power increment required for switching the state of an electrical device; the power increment is described as follows: ； The power increment set is then: ； ； ； In the formula, This represents the increment of the Nth sample in the power time series; and These are the static power increment set and the transition power increment set, respectively; the power increment set of the sequence to be detected. have: ； Jump weights in power variable point weight model Assignment formula and jump information entropy The calculation formula is as follows: ； ； In the formula, N2 is the number of samples in the jump sequence; By combining the Bayesian information criterion, the detection window threshold h is optimized.

2. The low-frequency non-intrusive load monitoring method based on adaptive event detection according to claim 1, characterized in that, The load recognition model is the CNN-SENet load recognition model, which introduces the squeeze-excitation attention mechanism into the convolutional neural network.

3. The low-frequency non-intrusive load monitoring method based on adaptive event detection according to claim 2, characterized in that, The CNN-SENet load identification model is specifically structured as follows: The model consists of one input layer, five convolutional layers, one pooling layer, one fully connected layer, and one output layer. The first four convolutional layers are standard 5x5 convolutions, while the fifth convolutional layer uses a 6x6 kernel with the SENet spatial attention module. The pooling layer performs max pooling and normalization on a 5x5 numerical matrix, operating in "SAME" mode. The output layer uses cross-entropy to construct the model's loss function, and Adam is selected as the gradient optimizer to dynamically update the CNN network parameters.

4. A low-frequency non-intrusive load monitoring system based on adaptive event detection, implementing the method of claim 1, characterized in that, include: Event detection module, feature extraction module, and load identification module; The event detection module uses the Bayesian information criterion as the detection window and adaptively optimizes the window threshold through a power variable point weight model. The feature extraction module uses variational mode decomposition to decompose and group the power time series of different electrical devices into load features. The load identification module uses a load identification model to identify and classify the load of different electrical devices based on their power curve waveforms.

5. A processor, characterized in that, The processor executes the low-frequency non-intrusive load monitoring method based on adaptive event detection as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the low-frequency non-intrusive load monitoring method based on adaptive event detection as described in any one of claims 1-3.

Images