Non-intrusive household load identification method based on CUSUM event detection and convolutional neural network

Abstract

The invention discloses a non-intrusive household load identification method based on CUSUM event detection and a convolutional neural network, which is applied to total load data output by an intelligent electric meter installed at a user total incoming line, and comprises the following steps: data preprocessing; carrying out load event detection based on a sliding window bilateral CUSUM algorithm so as to determine an occurrence moment and a continuous interval of a load event; in a continuous interval corresponding to each load event, calculating an active power variation delta P, a reactive power variation delta Q, a current harmonic distortion variation delta Ihd and a current span Itc in steady-state data segments before and after the load event, and combining to form a multi-dimensional event feature vector; the method comprises the following steps: sequentially executing the steps on operation data of each electric device in a family in different operation modes to obtain a multi-dimensional event feature vector marked with a device type and an operation state label, and constructing a standardized device load feature database; and training a convolutional neural network classification model and realizing load identification.

Description

Technical Field

[0001] This invention relates to the field of smart grids and home energy management, and in particular to a non-intrusive home load identification method based on CUSUM event detection and convolutional neural networks. Background Technology

[0002] With socio-economic development, household load has gradually become a major component of electricity consumption, and residents' awareness of energy conservation is also increasing. Therefore, monitoring household load has become a crucial issue for achieving rational energy management. Currently, traditional load monitoring methods mostly employ invasive sensors to monitor individual appliances in real time. While this method can directly acquire the operating status of appliances, it requires the installation of independent measuring devices on each appliance, resulting in large hardware deployment, high installation costs, and complex system maintenance, making it unsuitable for large-scale deployment in ordinary households. In contrast, Non-Intrusive Load Monitoring (NILM) technology only needs to collect power signals at the main power line of the household. Through signal processing and pattern recognition algorithms, it can indirectly identify the operating status and energy consumption levels of different appliances, thus significantly reducing hardware investment and showing promising application prospects. However, existing NILM technologies still have significant shortcomings in practical applications: on the one hand, traditional CUSUM detection methods are easily affected by noise when handling appliance switching and mode changes, resulting in unclear event boundaries; on the other hand, most recognition algorithms rely on limited features and exhibit poor robustness when multiple appliances are detected in parallel with similar characteristics; furthermore, deep learning models have limitations, and existing CNN methods often lack optimized design for event features, leading to decreased recognition performance in complex scenarios. Therefore, improving the accuracy and stability of event detection and enhancing load recognition capabilities in complex appliance scenarios are key issues that current NILM technologies urgently need to address. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a non-intrusive household load identification method based on improved CUSUM event detection and convolutional neural networks. By introducing a multi-window mechanism (mean window, transient detection window, and variance calculation window) into the traditional CUSUM method, the detection accuracy of appliance on / off and mode switching events is significantly improved. Furthermore, by combining multi-dimensional event feature vectors containing active power, reactive power, harmonic distortion, and current span, and using a convolutional neural network (CNN) for identification, high-precision energy consumption decomposition in complex appliance scenarios is achieved.

[0004] The objective of this invention is achieved through the following technical solution: A non-intrusive household load identification method based on CUSUM event detection and convolutional neural networks is applied to the total load data output by a smart meter installed at the user's main power line, including: S1. Collect the voltage and current effective value sequence output by the smart meter, perform median filtering on the current effective value sequence to remove pulse noise, and perform downsampling and smoothing processing on the filtered data to reduce the high-frequency sampled data to a smooth data sequence with a preset sampling frequency; S2. For the smoothed data sequence, load event detection is performed based on the sliding window bilateral CUSUM algorithm. Three continuous sliding windows are set: a mean calculation window, a transient detection window, and a variance calculation window. The deviation between the mean of the transient detection window and the mean of the mean calculation window is calculated and accumulated. When the accumulated sum exceeds a preset variance threshold, it is determined that a load event may occur, and the variance calculation window is activated. When the variance of the data in the variance calculation window is less than the preset variance threshold, it is determined that the load has entered a new steady state, so as to determine the time and duration of the load event. S3. Within the duration interval corresponding to each load event obtained in step S2, calculate the change in active power ΔP, the change in reactive power ΔQ, the change in current harmonic distortion ΔIhd, and the current span Itc in the steady-state data segments before and after the occurrence of the load event, and combine them to form a multi-dimensional event feature vector. S4. By sequentially executing steps S1 to S3 on the operating data of each electrical device in the home under different operating modes, extract the multidimensional event feature vector of each device, and label the device type and operating status to construct a standardized device load feature database; this database is used to form the training sample set of the convolutional neural network, where each sample includes a multidimensional event feature vector and its corresponding device type and operating status label.

[0005] S5. Based on the training sample set, the convolutional neural network classification model is trained. The multidimensional event feature vector of the load event to be identified is input into the trained convolutional neural network model, and the matching probability of each equipment type is output. The equipment type with the highest probability and its operating status are used as the identification result, so as to realize non-intrusive identification and energy consumption decomposition of the operating status of each electrical equipment in the total load.

[0006] Furthermore, in step S1, The median filter is used to remove impulse noise and outliers caused by acquisition equipment failure and electromagnetic interference. The downsampling operation reduces the original high-frequency sampled data to one sampling point per second or a preset sampling interval, thereby reducing data dimensionality and computational complexity while preserving the load change trend.

[0007] Furthermore, in step S2, The sliding window bilateral CUSUM algorithm accumulates and calculates the sums in both directions of load increase and load decrease, and is used to simultaneously detect the power-on and power-off events of electrical equipment. The mean calculation window is used to characterize the average load level under the current steady state, the transient detection window is used to capture possible load abrupt changes, and the variance calculation window is used to determine whether the load has entered a new steady state, thereby accurately locking the start and end times of load events.

[0008] Furthermore, in step S3, the change in current harmonic distortion ΔIhd includes the changes in the 3rd, 5th, and 7th harmonic components and the total harmonic distortion rate before and after the load event, which are used to characterize the differences in harmonic characteristics of different types of electrical equipment.

[0009] Furthermore, the current span Itc is the difference between the maximum and minimum values ​​of the current waveform during the load event monitoring period, which is used to reflect the fluctuation amplitude of the current waveform during the load event process, so as to improve the ability to distinguish between resistive, motor and electronic electrical equipment.

[0010] Furthermore, in the standardized equipment load characteristic database, for each electrical device in the household, its operating data is collected when it is turned on, turned off, and under different power levels or working modes. By repeatedly executing steps S1 to S3, multi-dimensional event feature vectors of corresponding typical load events are extracted and labeled with equipment type and operating status tags.

[0011] Furthermore, the convolutional neural network model includes an input layer, at least one convolutional layer, a normalization layer, a pooling layer, a fully connected layer, and an output layer connected in sequence. The output layer outputs the matching probability distribution corresponding to each device type; By training the model using multidimensional event feature vectors and equipment type labels from the standardized equipment load feature database as supervision signals, the convolutional neural network model maintains high recognition accuracy and robustness even in scenarios involving parallel and similar electrical appliances.

[0012] Preferably, the present invention also provides a non-invasive household load identification system based on CUSUM event detection and convolutional neural networks, comprising: The data acquisition unit is used to acquire the effective value sequence of voltage and current output by the smart meter; The data preprocessing unit is used to perform median filtering on the current RMS value sequence to remove impulse noise, and to perform downsampling and smoothing on the filtered data to reduce the high-frequency sampled data to a smooth data sequence with a preset sampling frequency. The event detection unit is used to detect load events based on the sliding window bilateral CUSUM algorithm for the smoothed data sequence. It sets three continuous sliding windows: a mean calculation window, a transient detection window, and a variance calculation window. It calculates the deviation between the mean of the transient detection window and the mean of the mean calculation window and accumulates the sum. When the accumulated sum exceeds a preset variance threshold, it determines that a load event may have occurred and starts the variance calculation window. When the variance of the data in the variance calculation window is less than the preset variance threshold, it determines that the load has entered a new steady state, so as to determine the time and duration of the load event. The event vector feature extraction unit is used to calculate the change in active power ΔP, the change in reactive power ΔQ, the change in current harmonic distortion ΔIhd, and the current span Itc in the steady-state data segments before and after the occurrence of each load event within the duration interval corresponding to each load event, and combine them to form a multi-dimensional event feature vector. The event vector feature matching unit is used to construct a standardized equipment load feature database based on the multi-dimensional event feature vectors collected from each electrical device, and to match the multi-dimensional event feature vectors of the load event to be identified with the multi-dimensional event feature vectors in the standardized equipment load feature database to form a training sample set for the convolutional neural network. The load identification unit trains the convolutional neural network classification model based on the training sample set, inputs the multi-dimensional event feature vector of the load event to be identified into the trained convolutional neural network model, outputs the matching probability of each equipment type, and takes the equipment type with the highest probability and its operating status as the identification result, thereby realizing non-intrusive identification and energy consumption decomposition of the operating status of each electrical equipment in the total load.

[0013] Preferably, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the non-intrusive home load identification method.

[0014] Preferably, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the non-intrusive household load identification method.

[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: 1. This invention adopts a non-intrusive overall architecture; a smart meter or data acquisition device is installed only at the main power inlet of the household to collect the effective value sequence of the total load voltage and current. Through load event detection and CNN recognition algorithms, the operating status and energy consumption of each electrical appliance are indirectly identified. By combining single-point data acquisition at the main power inlet with software algorithms to decompose the power consumption of each appliance, indirect monitoring is achieved, significantly reducing the number of hardware components and wiring complexity, lowering installation and maintenance costs, and improving the system's scalability and maintainability in residential settings. This solves the problems of traditional invasive monitoring methods, which require installing independent sensors on each appliance, resulting in large hardware deployments, complex installation and maintenance, high costs, and difficulty in widespread adoption in ordinary households.

[0016] 2. A median filtering algorithm is applied to the acquired current RMS value sequence to remove impulse noise and outliers introduced by equipment failures, electromagnetic interference, etc. Median filtering has a strong ability to suppress isolated spikes while relatively preserving the edge variations of the true signal, making the data entering the sliding window bilateral detection algorithm smoother and with fewer outliers. This reduces the interference of noise on the cumulative sum and statistics, improving the accuracy and stability of load event detection. This addresses the problem that directly using raw acquired data containing impulse noise, spikes, and outliers for load event detection can lead to abnormal CUSUM calculations, resulting in false positives, false negatives, and inaccurate load event boundaries.

[0017] 3. High-frequency sampled data is reduced to one sampling point per second or another preset low-frequency sampling rate, and the data is smoothed during this process to generate a smooth data sequence suitable for load event detection. Downsampling reduces the amount of data and the computational load of subsequent feature calculations and CUSUM accumulation operations, improving the real-time performance and deployability of the recognition system; appropriate smoothing can also preserve the load change trend, making the abrupt changes in load events still clear, thus ensuring the accuracy of load event detection while reducing computational load. This addresses the problems of large original high-frequency data volume, high computational load for direct processing, which can easily lead to excessive computational complexity in the back-end algorithm and poor system real-time performance; and the fact that high-frequency noise is not conducive to stable load event detection.

[0018] 4. Three continuous sliding windows are introduced in load event detection: a mean calculation window is used to estimate the current steady-state average current level; a transient detection window is used to monitor possible load abrupt changes; and a variance calculation window is used to determine whether a new steady state has been entered. A sliding window bilateral detection algorithm is used to accumulate and statistically analyze load increases and decreases. When the accumulated sum exceeds a threshold and meets the variance judgment, the start and end times of the load event are locked. This can more accurately lock the start and end times of events, significantly reduce false detections and missed detections caused by small fluctuations and noise, improve the accuracy and robustness of event detection, and provide high-quality load event fragments for subsequent feature extraction. This solves the problems of traditional methods that rely solely on a single window or only consider mean differences, which are easily affected by natural load fluctuations and noise. These methods also suffer from unclear event boundaries, false detections, and missed detections during appliance switching or mode switching, especially unstable identification of transient processes in low-power equipment and multiple appliances operating in parallel.

[0019] 5. In the steady-state data segment before and after a load event, the changes in active power ΔP, reactive power ΔQ, current harmonic distortion ΔIhd, and current span Itc are calculated and combined to form a multi-dimensional event feature vector. ΔP and ΔQ together characterize the changes in active / reactive power components, improving the differentiation of different load characteristics such as resistive, inductive, and capacitive loads. The change in current harmonic distortion ΔIhd reflects the differences in spectral characteristics between electronic and motor devices, providing additional distinguishing information for devices with similar characteristics but different topologies. The current span Itc reflects the amplitude of current waveform fluctuations during the load event, which is helpful in distinguishing motor devices with large starting currents from relatively stable resistive loads. The multi-dimensional event feature vector can more fully characterize the feature patterns of electrical appliance operating states, significantly improving the separability and recognition accuracy between devices. This solves the problem that existing technologies only utilize active power or a single current feature, which, when multiple appliances operate in parallel or have similar characteristics (such as multiple resistive loads or motorized appliances), lack sufficient distinguishability between features, leading to easy confusion in appliance type identification and low accuracy.

[0020] 6. For each household appliance operating under different power levels or modes during on / off states, preprocessing, event detection, and feature extraction are repeatedly performed to form a set of event feature vectors labeled with appliance type and operating status, constructing a standardized appliance load feature database. This database provides clear supervision signals and a unified feature representation for the CNN model, facilitating the model's learning of feature differences between different appliances and states. Furthermore, the database can be expanded and maintained based on the addition or removal of household appliances, improving the system's adaptability to new devices. This gives the load identification algorithm a trainable and transferable foundation, enhancing the model's generalization ability in different household scenarios. It also addresses the problem of overfitting or insufficient generalization ability that can easily occur in deep learning models lacking structured and fully labeled training samples.

[0021] 7. A CNN model consisting of an input layer, convolutional layer, normalization layer, pooling layer, fully connected layer, and output layer is established. Supervised training is performed using event feature vectors and their labels from a standardized equipment load feature database. During the online recognition phase, the unknown load event feature vectors are input into the trained CNN model, which outputs the matching probability for each equipment category. The model with the highest probability is selected as the recognition result. The CNN model, through convolution and nonlinear mapping, can automatically learn deep combination relationships between features, enhancing its ability to fit complex patterns. Normalization and pooling operations improve the model's robustness to feature scale changes and small perturbations. Experimental verification shows that the average recognition accuracy of this invention is better than 95%, maintaining good recognition performance even in scenarios involving multiple appliances operating in parallel or with similar characteristics.

[0022] 8. This invention employs a holistic approach encompassing data preprocessing, a sliding window bilateral detection algorithm, multi-dimensional event feature vectors, and CNN training and prediction. Starting with source noise suppression and data dimensionality reduction, progressing to robust load event detection and multi-dimensional feature extraction, and finally to deep learning model training and recognition for a standardized equipment load feature database, each stage is designed with the unified goal of high-precision, non-intrusive load identification in complex home scenarios. The input and output of each stage are tightly coupled, gradually suppressing errors and amplifying discriminative power. The overall system demonstrates superior performance compared to traditional methods in terms of recognition accuracy, robustness, computational efficiency, and home deployability. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the principle of the method of the present invention.

[0024] Figure 2 This describes the load event detection process based on the sliding window bilateral CUSUM algorithm.

[0025] Figure 3 This is a load identification network architecture based on convolutional neural networks.

[0026] Figure 4 This describes the changes in accuracy and loss of the training set.

[0027] Figure 5 This is the result of load classification. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0029] Example 1 A non-invasive household load identification method based on improved CUSUM event detection and convolutional neural networks, see [link to relevant documentation]. Figure 1 After preprocessing the total household load data, the data undergoes event detection, event vector feature extraction and matching, and finally load identification. These five progressively advanced steps achieve household energy load identification based on non-intrusive load monitoring. The details are as follows: S1. Collect total load data from the smart meter installed at the user's main incoming line, including the effective value sequence of voltage and current. Due to equipment failure or electromagnetic interference, the raw total load data contains noise and outliers, which will lead to misjudgment if used directly for event detection. Preprocessing of the raw current effective value sequence is necessary. First, a median filtering algorithm (window size of 3 sampling points) is used to remove obvious impulse noise. Then, to reduce computational complexity and dimensionality reduction of the data, while preserving the data variation trend, the high-frequency sampling data is reduced to one sampling point per second, generating a smooth data sequence suitable for subsequent load event detection.

[0030] S2. A sliding window-based bilateral CUSUM algorithm is used to monitor load events in the preprocessed smooth data sequence to detect load abrupt changes caused by the switching on or off of electrical equipment. The load event detection part sets up three consecutive sliding windows: a mean calculation window, a transient detection window, and a variance calculation window.

[0031] The mean calculation window is used to calculate the average current, i.e., the average load level, during the current steady-state operation of the current load. The window length m is 50 sampling points (corresponding to 50 seconds, with a sampling frequency of 1Hz). The transient detection window is used to monitor possible sudden changes in the load. The window length n is 10 sampling points (10 seconds). The variance calculation window is used to determine whether the load has entered a new steady state. The window length v is 50 sampling points (50 seconds). The CUSUM cumulative sum threshold H is set to 2.5. This value was determined experimentally and satisfies H... min <H<H max The conditions are set to balance detection sensitivity and anti-interference capabilities. Variance threshold H v Set to 1×10 -3 This value is determined based on the maximum steady-state variance of each device (see Table 1, the steady-state variance range for each device is 2.1 × 10⁻⁶). -6 ~8.2×10 -4 ).

[0032] Table 1. Steady-state operating variance of each device

[0033] Specifically, the sliding window bilateral CUSUM algorithm calculates the deviation between the mean of the transient detection window and the mean of the mean calculation window, and then sums them up.

[0034] Assume a time series of current RMS values This defines the parameters for the mean calculation window and the transient detection window.

[0035] (1) Mean calculation window Mean calculation window The length is defined as The average effective value of the current in this window is denoted as The calculation formula is shown in Equation 1: (1); In the formula, This represents the starting time of the sliding window.

[0036] (2) Transient detection window Transient detection window The length is defined as The average effective value of the current in this window is denoted as The calculation formula is shown in Equation 2: (2); Based on this, to detect the input and output of events for the detection equipment, the events input and output are cumulatively summed respectively. , The definition is given by equations (3) to (4): (3); (4); In the formula, , for The accumulation of time, This describes the fluctuations during stable current operation.

[0037] Based on the original CUSUM cumulative sum algorithm, a variance calculation window is added, and the calculation method is improved to increase the accuracy of detecting events entering steady state. Variance Calculation Window Its window length is The mean effective value of the current is denoted as The variance is In addition, the cumulative sum of events invested and cut off. , The event cumulative sum is updated. The updated cumulative sum calculation method can effectively improve the event detection effect of devices with small current. The calculation formula is shown in (5)~(6): (5); (6); This allows us to obtain the updated event accumulation and sum. , The calculation formula is shown in equations (7) to (8): (7); (8); In the formula, Additional coefficient This is the variance threshold. Additional coefficient. The selection of the value involved sequentially turning on multiple known devices, testing the recognition performance under different coefficients, and ultimately choosing the value that resulted in the highest recognition rate; variance threshold. The selection of is achieved by observing and calculating the variance of the effective current value of each device during steady-state operation, and letting . It is greater than the maximum value among them.

[0038] When the accumulated sum exceeds the preset threshold When a load event is detected, the variance calculation window is activated. The variance calculation window calculates the mean current. and variance , when the variance Less than the preset variance threshold When the load has entered a new steady state, the event's end time is accurately determined. The detection principle and determination method for load event termination are the same as for load event termination. For the load event detection process based on the sliding window bilateral CUSUM algorithm, please refer to [link to relevant documentation]. Figure 2 .

[0039] This embodiment effectively overcomes false detections caused by load fluctuations by introducing a variance window, significantly improving the accuracy of event detection, especially for low-power equipment events, at their end time. The sliding window bilateral CUSUM algorithm can detect whether transient load events have occurred and record the relevant load events for subsequent load identification algorithms. Compared to traditional non-intrusive load monitoring algorithms, this algorithm has higher event detection accuracy and can effectively record the entry and exit points of events, thereby improving the precision of event detection. This provides a solid data foundation for load identification.

[0040] S3. For each load event identified in step S2, extract a multi-dimensional event feature vector to characterize the equipment type from the steady-state data segment (duration interval) before and after the load event. This includes: (1) Change in active power (ΔP): The difference between the steady-state active power after the end of the load event and the steady-state active power before the start of the event; (2) Reactive power change (ΔQ): The difference between the steady-state reactive power after the load event ends and the steady-state reactive power before the event begins; (3) Change in current harmonic distortion (ΔIhd): Extract the changes in the 3rd, 5th, and 7th harmonics and total harmonic distortion before and after the load event; (4) Current span (Itc): The difference between the maximum and minimum values ​​of the current waveform within the duration of a load event.

[0041] These feature parameters together constitute a multi-dimensional event feature vector that can distinguish electrical equipment with different characteristics, such as resistive, motor, and electronic equipment.

[0042] To illustrate the effectiveness of multidimensional feature vectors, typical feature value ranges for different types of household appliances are listed (see Table 2): Rice cooker (resistive type): Features are as follows Stablize, and Extremely low. Air conditioner (motor type): Characterized by... and Significantly higher. Televisions (electronics): Characterized by... Lower, but Relatively high.

[0043] Table 2. Value range of characteristic parameters for each device

[0044] By combining these multi-dimensional features, it is possible to effectively distinguish devices with similar power but different types. For example, a 1kW resistance heater (whose features are similar to a rice cooker) ) with a 1kW air conditioning compressor (characteristics similar to an air conditioner, ), can be accessed and To accurately distinguish significant differences.

[0045] S4. Establish a standardized equipment load characteristic database for known electrical appliances to provide a comparison benchmark for load identification. By individually collecting operating data for each electrical appliance in a household under different operating modes (such as on, off, and high / low power switching), repeat steps S1-S3 to extract typical multi-dimensional event feature vectors for various load events of each electrical appliance. Label these vectors with equipment type and status tags to construct a standardized equipment load characteristic database. This database serves as the training sample set for the convolutional neural network model.

[0046] S5. A classification model based on convolutional neural networks (CNNs) is used to classify and identify the multidimensional event feature vectors of unknown load events. The load identification network architecture based on CNNs is as follows: Figure 3As shown, the main architecture of this model consists of 12 layers, including an input layer, two convolutional layers, two normalization layers, two activation layers, two max pooling layers, one fully connected layer, and one output layer. The specific composition and data flow are as follows: The input layer receives a 5-dimensional feature vector (corresponding to ΔP, ΔQ, ΔI). hd (3rd, 5th, and 7th harmonics and total distortion), I tc The input size is 5×1×1; the first convolutional layer uses 16 2×1 convolutional kernels with a stride of 1, followed by a batch normalization layer and a ReLU activation layer, outputting a feature map of size 4×1×16 through convolution operations; the first pooling layer uses max pooling with a window size of 2×1 and a stride of 1, outputting a feature map of size 3×1×16; the second convolutional layer uses 32 2×1 convolutional kernels with a stride of 1, followed by a batch normalization layer and a ReLU activation layer, outputting a feature map of size 2×1×32; the second pooling layer uses max pooling with a window size of 2×1 and a stride of 1, outputting a feature map of size 1×1×32; the fully connected layer flattens the 1×1×32 feature map into a 32-dimensional vector; the output layer uses the Softmax activation function to output the probability distribution of 12 device types.

[0047] First, the CNN model was trained in a supervised manner using a standardized equipment load feature database. The training hyperparameters were set as follows: initial learning rate of 0.001, adaptive descent strategy, and a learning rate reduction factor of 0.5 after 450 iterations; batch size of 32; maximum number of iterations of 500; cross-entropy loss function; and the Adam algorithm as the optimizer, with parameters set to β1=0.9, β2=0.999, and ε=10. -8 The training set contains 1200 samples (100 samples from each of the 12 devices), which are divided into training / validation / test sets in an 80 / 10 / 10 ratio.

[0048] After training, the multi-dimensional event feature vector of the load event to be identified is input into the trained CNN model. The model outputs the matching probability of the multi-dimensional event feature vector with each equipment type. The equipment type with the highest probability is output as the identification result, thereby completing the accurate identification of the operating status and energy consumption decomposition of each electrical device in the total load.

[0049] Example 2 To verify the superiority of the aforementioned non-invasive household load identification method over existing technologies, a systematic comparative experiment was conducted based on 29 days of energy consumption data collected from actual buildings. All comparative results are based on the processing and analysis of the same dataset. The experimental results are as follows: In event detection, the proposed sliding window bilateral CUSUM algorithm (mean window, transient window, variance window) was compared with the traditional single-window CUSUM method on the same dataset. Performance metrics included false positive rate, false negative rate, and absolute error of event start and end times. The effectiveness of the event detection proposed in this invention is directly evidenced by the algorithm's processing output of the total load data. As shown in Tables 3 and 4, this invention can stably and accurately extract a large number (a total of 3296) of events with clear start and end boundaries and feature vectors from the total load curve. Statistical analysis of these output events confirms the significant improvement in event detection accuracy and stability compared to traditional methods: the false positive rate is reduced from approximately 15% to <5%; the false negative rate is reduced from approximately 10% to <3%; and the error in event start and end times is reduced from over ±3 seconds to within ±1 second.

[0050] Table 3. Changes in RMS Current During Partial Load Events

[0051] Table 4. Changes in the Eigenvector of Partial Load Events

[0052] In terms of load identification, the multi-dimensional feature vector (ΔP, ΔQ, ΔI) proposed in this invention will be used. hd , I tc The load identification accuracy of the scheme combining multi-dimensional features with a CNN model was compared with that of the traditional scheme using only a single active power feature (ΔP) combined with a decision tree classifier. Experimental results fully demonstrate the synergistic advantages of multi-dimensional features and deep learning models: the average identification accuracy was significantly improved from 85% of the traditional method to 98.8%. For changes in training set accuracy and loss, please refer to [link to relevant documentation]. Figure 4 The figure shows that the model training accuracy converged rapidly to over 99%; meanwhile, the load classification results are shown in [reference needed]. Figure 5 The high-precision classification performance achieved by the model on the test set is intuitively demonstrated in the form of a confusion matrix. In challenging scenarios involving multiple appliances operating in parallel and appliances with similar characteristics, the method of this invention still maintains a recognition accuracy of over 95%; while the accuracy of traditional methods drops significantly to around 70% in the same scenario. Key evidence for this robustness advantage lies in the detailed data provided in Table 5. This table clearly shows that even for highly similar and easily confused device pairs (such as rice cookers and hair dryers), the method of this invention, thanks to its multi-dimensional features, still achieves recall rates of 98.9% and 92.1% respectively, fully demonstrating its effectiveness in complex scenarios.

[0053] Table 5 Evaluation of Device Recognition Performance

[0054] Taking rice cookers and hair dryers, two resistive devices with similar characteristics, as typical examples, traditional methods, relying solely on active power characteristics, are insufficient for differentiation. This invention utilizes both their reactive power change (ΔQ) and current span (I)... tc The differences between the devices (see Table 2 for specific feature ranges) were successfully identified, and this case is a typical example of how the present invention ensures a high recognition rate even among devices with high similarity.

[0055] In summary, through comparative experiments on the two stages of event detection and load identification, and through mutual verification with the original data table, model training graph, classification result graph, and performance evaluation table, it is fully demonstrated that the present invention, by combining the sliding window bilateral CUSUM algorithm with the multi-dimensional feature CNN model, has achieved significant progress in both event detection accuracy and load identification accuracy, and ultimately realizes high-precision non-intrusive identification of household loads.

[0056] Example 3 Based on the same inventive concept, this application also provides a non-invasive household load identification system based on CUSUM event detection and convolutional neural networks, which can be used to implement the methods described in the above embodiments, as shown in the following embodiments. Since the principle of solving the problem in the non-invasive household load identification system based on CUSUM event detection and convolutional neural networks is similar to that of the non-invasive household load identification method based on CUSUM event detection and convolutional neural networks, the implementation of this device can refer to the aforementioned method implementation, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0057] This embodiment provides a specific implementation of a non-intrusive household load identification system capable of implementing a non-intrusive household load identification method, specifically including the following: The data acquisition unit is used to acquire the effective value sequence of voltage and current output by the smart meter; The data preprocessing unit is used to perform median filtering on the current RMS value sequence to remove impulse noise, and to perform downsampling and smoothing on the filtered data to reduce the high-frequency sampled data to a smooth data sequence with a preset sampling frequency. The event detection unit is used to detect load events based on the sliding window bilateral CUSUM algorithm for the smoothed data sequence. It sets three continuous sliding windows: a mean calculation window, a transient detection window, and a variance calculation window. It calculates the deviation between the mean of the transient detection window and the mean of the mean calculation window and accumulates the sum. When the accumulated sum exceeds a preset variance threshold, it determines that a load event may have occurred and starts the variance calculation window. When the variance of the data in the variance calculation window is less than the preset variance threshold, it determines that the load has entered a new steady state, so as to determine the time and duration of the load event. The event vector feature extraction unit is used to calculate the change in active power ΔP, the change in reactive power ΔQ, the change in current harmonic distortion ΔIhd, and the current span Itc in the steady-state data segments before and after the event, respectively, within the duration interval corresponding to each load event, and combine them to form a multi-dimensional event feature vector. The event vector feature matching unit is used to construct a standardized equipment load feature database based on the multi-dimensional event feature vectors collected from each electrical device, and to match the multi-dimensional event feature vectors of the load event to be identified with the multi-dimensional event feature vectors in the standardized equipment load feature database to form a training sample set for the convolutional neural network. The load identification unit trains the convolutional neural network classification model based on the training sample set, inputs the multi-dimensional event feature vector of the load event to be identified into the trained convolutional neural network model, outputs the matching probability of each equipment type, and takes the equipment type with the highest probability and its operating status as the identification result, thereby realizing non-intrusive identification and energy consumption decomposition of the operating status of each electrical equipment in the total load.

[0058] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps of the non-invasive household load identification method based on CUSUM event detection and convolutional neural networks in the above embodiments. The electronic device specifically includes the following: Processor, memory, communications interface, and bus; The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0059] The processor is used to call the computer program in memory. When the processor executes the computer program, it implements all the steps in the non-invasive household load identification method based on CUSUM event detection and convolutional neural network in the above embodiments.

[0060] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the non-invasive household load identification method based on CUSUM event detection and convolutional neural networks in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the non-invasive household load identification method based on CUSUM event detection and convolutional neural networks in the above embodiments.

[0061] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0062] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A non-invasive household load identification method based on CUSUM event detection and convolutional neural networks, characterized in that, The total load data output by the smart meter installed at the user's main power line includes: S1. Collect the voltage and current effective value sequence output by the smart meter, perform median filtering on the current effective value sequence to remove pulse noise, and perform downsampling and smoothing processing on the filtered data to reduce the high-frequency sampled data to a smooth data sequence with a preset sampling frequency; S2. For the smoothed data sequence, load event detection is performed based on the sliding window bilateral CUSUM algorithm. Three continuous sliding windows are set: a mean calculation window, a transient detection window, and a variance calculation window. The deviation between the mean of the transient detection window and the mean of the mean calculation window is calculated and accumulated. When the accumulated sum exceeds a preset variance threshold, it is determined that a load event may occur, and the variance calculation window is activated. When the variance of the data in the variance calculation window is less than the preset variance threshold, it is determined that the load has entered a new steady state, so as to determine the time and duration of the load event. S3. Within the duration interval corresponding to each load event obtained in step S2, calculate the change in active power ΔP, the change in reactive power ΔQ, the change in current harmonic distortion ΔIhd, and the current span Itc in the steady-state data segments before and after the occurrence of the load event, and combine them to form a multi-dimensional event feature vector. S4. By sequentially executing steps S1 to S3 on the operating data of each electrical device in the home under different operating modes, extract the multi-dimensional event feature vector of each device, and label the device type and operating status to construct a standardized device load feature database; this database is used to form the training sample set of the convolutional neural network, where each sample includes a multi-dimensional event feature vector and its corresponding device type and operating status label; S5. Based on the training sample set, the convolutional neural network classification model is trained. The multidimensional event feature vector of the load event to be identified is input into the trained convolutional neural network model, and the matching probability of each equipment type is output. The equipment type with the highest probability and its operating status are used as the identification result, so as to realize non-intrusive identification and energy consumption decomposition of the operating status of each electrical equipment in the total load.

2. The non-invasive household load identification method according to claim 1, characterized in that, In step S1, The median filter is used to remove impulse noise and outliers caused by acquisition equipment failure and electromagnetic interference. The downsampling operation reduces the original high-frequency sampled data to one sampling point per second or a preset sampling interval, thereby reducing data dimensionality and computational complexity while preserving the load change trend.

3. The non-invasive household load identification method according to claim 1, characterized in that, In step S2, The sliding window bilateral CUSUM algorithm accumulates and calculates the sums in both directions of load increase and load decrease, and is used to simultaneously detect the power-on and power-off events of electrical equipment. The mean calculation window is used to characterize the average load level under the current steady state, the transient detection window is used to capture possible load abrupt changes, and the variance calculation window is used to determine whether the load has entered a new steady state, thereby accurately locking the start and end times of load events.

4. The non-invasive household load identification method according to claim 1, characterized in that, In step S3, the change in current harmonic distortion ΔIhd includes the changes in the 3rd, 5th, and 7th harmonic components and the total harmonic distortion rate before and after the load event, which are used to characterize the differences in harmonic characteristics of different types of electrical equipment.

5. The non-invasive household load identification method according to claim 1, characterized in that, The current span Itc is the difference between the maximum and minimum values ​​of the current waveform during the load event monitoring period. It is used to reflect the fluctuation amplitude of the current waveform during the load event process, so as to improve the ability to distinguish between resistive, motor and electronic electrical equipment.

6. The non-invasive household load identification method according to claim 1, characterized in that, In the standardized equipment load characteristic database, for each electrical device in the household, its operating data is collected when it is turned on, turned off, and under different power levels or working modes. By repeatedly executing steps S1 to S3, multi-dimensional event feature vectors of corresponding typical load events are extracted and labeled with equipment type and operating status tags.

7. The non-invasive household load identification method according to claim 1, characterized in that, The convolutional neural network model includes an input layer, at least one convolutional layer, a normalization layer, a pooling layer, a fully connected layer, and an output layer connected in sequence. The output layer outputs the matching probability distribution corresponding to each device type; By training the model using multidimensional event feature vectors and equipment type labels from the standardized equipment load feature database as supervision signals, the convolutional neural network model maintains high recognition accuracy and robustness even in scenarios involving parallel and similar electrical appliances.

8. A non-invasive household load identification system based on CUSUM event detection and convolutional neural networks, characterized in that, include: The data acquisition unit is used to acquire the effective value sequence of voltage and current output by the smart meter; The data preprocessing unit is used to perform median filtering on the current RMS value sequence to remove impulse noise, and to perform downsampling and smoothing on the filtered data to reduce the high-frequency sampled data to a smooth data sequence with a preset sampling frequency. The event detection unit is used to detect load events based on the sliding window bilateral CUSUM algorithm for the smoothed data sequence. It sets three continuous sliding windows: a mean calculation window, a transient detection window, and a variance calculation window. It calculates the deviation between the mean of the transient detection window and the mean of the mean calculation window and accumulates the sum. When the accumulated sum exceeds a preset variance threshold, it determines that a load event may have occurred and starts the variance calculation window. When the variance of the data in the variance calculation window is less than the preset variance threshold, it determines that the load has entered a new steady state, so as to determine the time and duration of the load event. The event vector feature extraction unit is used to calculate the change in active power ΔP, the change in reactive power ΔQ, the change in current harmonic distortion ΔIhd, and the current span Itc in the steady-state data segments before and after the occurrence of each load event within the duration interval corresponding to each load event, and combine them to form a multi-dimensional event feature vector. The event vector feature matching unit is used to construct a standardized equipment load feature database based on the multi-dimensional event feature vectors collected from each electrical device, and to match the multi-dimensional event feature vectors of the load event to be identified with the multi-dimensional event feature vectors in the standardized equipment load feature database to form a training sample set for the convolutional neural network. The load identification unit trains the convolutional neural network classification model based on the training sample set, inputs the multi-dimensional event feature vector of the load event to be identified into the trained convolutional neural network model, outputs the matching probability of each equipment type, and takes the equipment type with the highest probability and its operating status as the identification result, thereby realizing non-intrusive identification and energy consumption decomposition of the operating status of each electrical equipment in the total load.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the non-intrusive household load identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the non-intrusive household load identification method according to any one of claims 1 to 7.

Images