A sliding window bilateral CUSUM event detection method considering time threshold and its application

Abstract

This invention discloses a sliding window bilateral CUSUM event detection method and its application, considering time thresholds. Based on the concept of non-intrusive load monitoring, this invention analyzes the access and operation characteristics of residential two-stage single-phase photovoltaic grid-connected systems. By considering the flexible adjustment strategy of photovoltaic inverters, it incorporates the detection of abnormal photovoltaic access status into the monitoring scope, proposing a sliding window bilateral CUSUM event detection method considering time thresholds. This method can be used for appliance identification and solves the problem of difficulty in identifying residential photovoltaic systems based on non-intrusive schemes due to the uncertainty, continuous variability, and insignificant switching characteristics of photovoltaic equipment power generation. Furthermore, this invention improves the speed and accuracy of identification by combining event detection based on photovoltaic access characteristics with deep learning algorithms.

Description

Technical Field

[0001] This invention relates to a sliding window bilateral CUSUM event detection method and its application that considers time thresholds, belonging to the field of non-intrusive interactive energy monitoring. Background Technology

[0002] Distributed energy, represented by photovoltaics (PV), has been vigorously promoted in China. By the end of 2021, my country's grid-connected PV capacity reached 306 million kilowatts, ranking first in the world for seven consecutive years. However, large-scale distributed PV grid connection has had a significant impact on power flow, line losses, and power quality in the distribution network, posing a huge challenge to the safe and economical operation of the distribution network. With the increasing maturity of smart grid technology, existing research has mitigated the impact through various solutions such as active distribution networks, virtual power plants, and demand-side management. A necessary condition for these smart solutions is the observability of the distribution network. However, due to the unauthorized or unreported installation of some residential PV devices and the lack of separate metering for most residential devices in the distribution network, the visibility of the rapidly growing number of residential PV devices in the distribution network is actually very limited, including information such as the status, quantity, type, and location of each device. Furthermore, installing sensors on every household is impractical due to cost and privacy concerns. Therefore, developing a cost-effective method that does not require separate monitoring to estimate the regional-level household connectivity status is crucial. At this time, the non-intrusive load monitoring (NILM) method has a promising future for identifying photovoltaic devices. Implementing NILM-type algorithms on smart meters can provide real-time photovoltaic power generation / consumption information for the power grid or users, improve the observability of the distribution network, mitigate the negative impact of uncoordinated growth of photovoltaics in the distribution network, thereby improving the flexibility, reliability and efficiency of the system, and facilitating the effective transition from traditional power systems to dynamic smart grids. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a sliding window bilateral CUSUM event detection method and its application that considers time thresholds.

[0004] The technical solution adopted in this invention is as follows:

[0005] A sliding window bilateral CUSUM event detection method considering time thresholds includes:

[0006] The process involves acquiring a time series of electrical parameters to be detected, where the electrical parameters are active power, reactive power, or RMS current values; sampling the time series of electrical parameters using a sliding window and detecting whether an event occurs within each sliding window; the sliding window includes a mean calculation window W. m and transient detection window W dTwo consecutive sliding windows; for each sliding window, calculate the transient detection window W one by one. d The cumulative sum of input or output events corresponding to each time k within the time frame. And determine: If Then let the time delay factor d = 0; if At that time, let d = d + 1; until The timing of an input or cut-out event is determined by the inverse transformation of t = kd.

[0007] Where h is the cumulative sum threshold, representing the cumulative sum of the input and output events for each sliding window. The initial value is 0, which is represented as follows:

[0008]

[0009]

[0010] In the formula, β represents the noise introduced from the outside, and ψ(k) is the weight of k at each time step within the transient detection window calculated based on the Gaussian function; M d It is the average value of the electrical parameters in the current transient detection window; μ gauss It is the average value of the mean calculation window based on the Gaussian function;

[0011] After confirming that an input or cut-out event has occurred, continue to detect and calculate the cumulative sum of the corresponding input or cut-out events and determine the event end point: if the difference between the cumulative sum of input or cut-out events at the next time step and the current time step is less than the noise threshold, then the current time step is the event end point;

[0012] If the time difference between two consecutive events is less than a given time threshold, the two consecutive events are merged into one event.

[0013] Furthermore, the method for calculating the weight of k at each time step within the transient detection window based on the Gaussian function is as follows:

[0014]

[0015] Among them, u d N represents the median value of the maximum width of the transient detection window; α represents the backward offset of the detection window weights. m N d represents the length of the mean calculation window and the transient detection window, respectively; j is the position of the first sampling point of the current mean calculation window.

[0016] Furthermore, the method for calculating the average value of the mean calculation window based on the Gaussian function is as follows:

[0017] The weight representation at each time step within the mean calculation window is determined using a Gaussian function:

[0018]

[0019] In the formula u i u represents the position of each sampling point within the average value window. c This indicates the sampling point position in the middle of the mean calculation window; σ is a constant greater than 0.

[0020] Normalize the weight representation f(ui) to obtain the weight at each time step within the mean calculation window.

[0021]

[0022] In the formula, N m This indicates the length of the mean calculation window; j is the position of the first sampling point in the current mean calculation window;

[0023] The mean value of the Gaussian function calculation window is μ. gauss for:

[0024]

[0025] W m (u i ) indicates the mean calculation window within u i The electrical parameter values ​​corresponding to the sampling point locations.

[0026] Furthermore, the sliding distance of the sliding window is the length of the mean calculation window.

[0027] Furthermore, the cumulative sum and threshold h are expressed as:

[0028] h=(Δ min -β)T max ;

[0029] In the formula: T max =N max T s N max T is the maximum allowed number of delayed sampling points. s Δ is the sampling interval; min Indicates the noise threshold.

[0030] This invention is based on the concept of non-intrusive load monitoring. It analyzes the access and operation characteristics of residential two-stage single-phase photovoltaic grid connection and incorporates the detection of abnormal photovoltaic access status into the monitoring scope by considering the flexible adjustment strategy of photovoltaic inverters. It proposes a sliding window bilateral CUSUM event detection method that considers time thresholds. This method can be used for non-intrusive appliance identification and can solve the problem of difficulty in identifying residential photovoltaic based on non-intrusive schemes caused by the uncertainty, continuous variation and insignificant switching characteristics of photovoltaic equipment power generation.

[0031] A non-invasive electrical appliance identification method includes:

[0032] The system acquires the user's electricity meter information in real time and uses the sliding window bilateral CUSUM event detection method with time threshold consideration to detect whether a switching event has occurred. If a switching event occurs, the system constructs an input feature vector based on the electricity meter information before and after the event and inputs it into the trained recognition model to obtain the recognition result.

[0033] Furthermore, the user's electrical appliances include residential photovoltaic equipment.

[0034] Furthermore, the recognition model is structured as an Attention-CNN-BiLSTM neural network. This online recognition scheme, using the Attention-CNN-BiLSTM neural network model, improves both the speed and accuracy of recognition.

[0035] Furthermore, the input feature vector includes active power, reactive power, voltage amplitude, and current harmonics.

[0036] The beneficial effects of this invention are:

[0037] This invention analyzes the characteristics of photovoltaic (PV) access and operation in non-intrusive interactive energy monitoring, and fully considers the flexible behavior of PV inverters. It proposes a sliding window bilateral CUSUM event detection method that considers time thresholds, which can solve the problem of difficulty in identifying residential PV systems based on non-intrusive schemes due to the uncertainty, continuous variability, and insignificant switching characteristics of PV equipment power generation. Attached Figure Description

[0038] Figure 1 This is a flowchart of a sliding window bilateral CUSUM event detection method considering time thresholds according to the present invention;

[0039] Figure 2 A diagram illustrating the sliding window bilateral CUSUM algorithm process for a conventional electrical appliance, using power as an example.

[0040] Figure 3 The diagram illustrates the calculation principle of h;

[0041] Figure 4 A diagram illustrating the process of a sliding window bilateral CUSUM algorithm considering time thresholds for residential photovoltaic systems, with power as an example.

[0042] Figure 5 A block diagram of a neural network architecture;

[0043] Figure 6 This is a flowchart of a non-invasive electrical appliance identification method. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] To fully understand this invention, detailed structures and steps will be presented in the following description, and the embodiments of this application will be further described in detail in conjunction with the accompanying drawings and specific implementations, so as to illustrate the technical solutions proposed by this invention.

[0046] In power system load monitoring, if a monitored quantity, such as active power, reactive power, or the effective value of current, undergoes a sudden change due to load switching, the CUSUM change point detection algorithm can determine the timing of the load switching. Since the direction of change of the sudden signal can be either positive or negative, a bilateral CUSUM algorithm is generally used. This invention provides a sliding window bilateral CUSUM event detection method considering a time threshold, comprising:

[0047] Obtain the time series {x(k)} of the electrical parameters to be detected, k = 1, 2, ... , where k represents time, and x(k) is the electrical parameter value at time k. The electrical parameters are commonly monitored quantities in the process of power system load monitoring, generally active power, reactive power, or RMS current, etc.

[0048] A sliding window is used to sample the time series of electrical parameters to be tested and to detect whether an event occurs within each sliding window; the sliding window includes a mean calculation window W. m and transient detection window W d Two consecutive sliding windows, each with a length of N. m N d Process analysis, such as Figure 4 As shown, for each sliding window, the transient detection window W is calculated one by one. d The cumulative sum of input or output events corresponding to each time k within the time frame. Conventional electrical appliances typically enter a stable operating state quickly after being started, such as Figure 2 As shown, for residential photovoltaic (PV) systems, the PV equipment does not immediately enter a stable working state after startup, such as... Figure 4 As shown, it is necessary not only to determine the start-up time of the photovoltaic equipment, but also when it enters steady state to avoid misjudgment. Therefore, a bilateral cumulative sum (CUSUM) event detection algorithm with a composite window considering time thresholds is proposed to address this characteristic. Specifically:

[0049] First, the concept of a time difference threshold between two consecutive events is introduced, which can be used to mitigate the false alarm problem. When the time difference is within a certain time threshold, we treat them as a single event (i.e., the second event is discarded as a false alarm).

[0050] Δt<T th

[0051] Where Δt is the time difference between two consecutive events, T th This is a given time threshold. This threshold or time limit is derived from an investigation of the transient characteristics of typical smart devices.

[0052] Secondly, the average value window is processed using a Gaussian function. In event detection, sampled values ​​near the time when an event might occur are affected by the event's initiation and have little reference value. Sampled values ​​further away from the time when an event might occur will cause the average value to be biased towards the load level before the transient event, and the corresponding weights should be reduced. Therefore, using a Gaussian window to obtain the mean is an ideal method for calculating the average value. The weights at each time point within the mean calculation window determined by the Gaussian function are expressed as follows:

[0053]

[0054] In the formula u i u represents the position of each sampling point within the average value window. c The value σ represents the sampling point position in the middle of the mean calculation window; σ is a constant greater than 0 that determines the smoothness of the Gaussian window. The larger the value of σ, the better the smoothness of the Gaussian window and the more uniform its effect on each sampling point; the smaller the value of σ, the more prominent the filtering effect of the Gaussian window on the preceding and following data; when σ approaches infinity, the Gaussian window becomes the mean window.

[0055] Normalize the weight representation f(ui) to obtain the weight at each time step within the mean calculation window.

[0056]

[0057] In the formula, j is the position of the first sampling point of the current mean calculation window; then the average value μ of the mean calculation window calculated by the Gaussian function is... gauss for:

[0058]

[0059] W m (u i ) represents the mean calculation window W m Inside u i The electrical parameter values ​​corresponding to the sampling points.

[0060] The characteristics of a transient detection window generally need to balance noise immunity and sensitivity to transient events, which are almost impossible to achieve simultaneously. However, the load typically increases continuously during a transient event. Gradually amplifying the continuously increasing sample values ​​during the transient event can distinguish between transient events and noise, thereby improving detection accuracy. Within the detection window, since the accumulation effect of triggering the detection window may be due to excessive noise, the weight ratio of the accumulated values ​​in the initial segment of the detection window should be reduced. As the load level increases, the weight of the increasing electrical parameter values ​​is gradually increased, ultimately amplifying the effect of ramp-up events and enabling rapid identification of transient events. Therefore, an improved Sigmoid function is used to modify the transient detection window. The weight of k at each time point within the transient detection window is expressed as:

[0061]

[0062] Among them, u d α represents the median value of the maximum width of the transient detection window; α is a constant greater than 0, representing the backward offset of the detection window weights. The larger α is, the more sensitive the detection is to the climbing characteristics.

[0063] Normalizing h(k) yields the weight of k at each time step within the transient detection window:

[0064]

[0065] The final calculation of the transient detection window W d The cumulative sum of input or output events corresponding to each time k within the time frame. The method is as follows:

[0066]

[0067]

[0068] In the formula, β represents the noise introduced from the outside, and M d It is the average value of the electrical parameters in the current transient detection window, expressed as:

[0069]

[0070] The cumulative events of each sliding window's entry and exit, and The initial value is 0.

[0071] Based on the acquired input, accumulated cut-off events, and judgment: If Then let the time delay factor d = 0; if At that time, let d = d + 1; until The timing of the input or output event is determined by the inverse transformation of t = kd; the calculation principle of h is as follows: Figure 3 As shown, the calculation formula is as follows:

[0072] h=(Δ min -β)T max

[0073] In the formula: T max =N max T s N max T is the maximum allowed number of delayed sampling points. s Δ represents the sampling interval; β is related to the noise in the measurement. min Indicates the noise threshold.

[0074] After confirming the occurrence of an input or cutoff event, continue to detect and calculate the cumulative sum of the corresponding input or cutoff events and determine the event end point: if the difference between the cumulative sum of input or cutoff events at the next time step and the current time step is less than the noise threshold, then the current time step is the event end point.

[0075]

[0076] This completes the detection of an input or output event.

[0077] As a preferred embodiment, the sliding distance of the sliding window is the length N of the mean calculation window. m This is to ensure that it can be detected at any time. Figure 1 The figure shows a complete flowchart of a sliding window bilateral CUSUM event detection method that considers time thresholds. As shown in the figure, the method specifically includes:

[0078] First, initialize and determine the constant values ​​in the method, including: Δ min , β, σ, α and calculate h;

[0079] The system iteratively reads data blocks acquired through the sliding window and assigns initial parameter values ​​to each data block: d = 0 and the initial value of the cumulative sum of input and output events is 0; it then determines whether the current k is within the current data block: NN d -N m If k < N, then perform an input or output event detection; otherwise, use a sliding window to read the next data block and update N = N + N. d .

[0080] For each sliding window, first calculate μ of the current sliding window. gauss Then jump to the transient detection window of the sliding window, k = k + N m Calculate input, cut-out event accumulation and First, determine if the cumulative sum of input and cut-out events is greater than 0. If not, it indicates that the current state is steady and proceeds to the next time step. If yes, further determine if the cumulative sum of input and cut-out events is greater than h. If yes, it indicates that an event has been detected, assign k = k + d + 1, and determine the event's end point. Otherwise, assign d = d + 1 and determine if the cumulative sum of input and cut-out events at time k + d is greater than h. If yes, it indicates that an event has been detected, assign k = k + d + 1, and determine the event's end point. Otherwise, further determine the magnitude of the cumulative sum of input and cut-out events at time k + d and time k + d + 1. If the cumulative sum of input and cut-out events at time k + d + 1 is greater than that at time k + d, assign d = d + 1 and determine if the cumulative sum of input and cut-out events at time k + d is greater than h. Otherwise, it indicates that the current state is steady, assign k = k + d, and proceed to the next time step.

[0081] The sliding window bilateral CUSUM event detection method of the present invention, which considers time thresholds, can be applied to appliance identification and can solve the problem of difficulty in identifying residential photovoltaic systems based on non-intrusive schemes due to the uncertainty, continuous variability, and insignificant switching characteristics of power generation from photovoltaic equipment. Specifically, a non-intrusive appliance identification method includes:

[0082] The system acquires users' electricity meter information in real time and uses the aforementioned sliding window bilateral CUSUM event detection method, which considers time thresholds, to detect whether a switching event has occurred. If a switching event occurs, an input feature vector is constructed based on the electricity meter information before and after the event, and then input into the trained recognition model to obtain the recognition result. As a preferred solution, the recognition model adopts a CNN-Attention-LSTM neural network model. CNN is a type of deep neural network structure, commonly used to extract feature information, and has significant advantages in data processing. Generally, CNN can be divided into: input layer, hidden layer, and output layer. The hidden layer can be further divided into convolutional layer, pooling layer, and fully connected layer. Neurons in each layer are interconnected, while neurons in the same layer are not connected. When the input data X is given, the feature map C of the convolutional layer can be represented as follows:

[0083]

[0084] In the formula, For convolution operation; W is the weight vector of the convolution kernel; b represents the offset; f(·) is the activation function, which can be ReLU, sigmoid, tanh, etc.

[0085] Pooling layers perform downsampling on the convolutional output, preserving strong features and removing weak features, while reducing the number of parameters and preventing overfitting. This invention utilizes CNNs to extract features from the original data, uncovering the interrelationships between multidimensional data and removing noise and unstable components. The processed, relatively stable information is then fed as a whole into a Long Short-Term Memory (LSTM) artificial neural network. LSTM is a type of recurrent neural network (RNN) that solves the gradient vanishing problem during RNN training and exhibits good performance when processing long-term sequences. LSTM consists of multiple identical cell structures, each composed of a forget gate, an input gate, and an output gate. LSTM also includes BiLSTM.

[0086] The attention mechanism is a model that simulates human brain attention. It draws on the fact that the human brain concentrates its attention on specific areas at a given moment, while reducing or even ignoring attention to other parts. By assigning different weights to the model's input features, the attention mechanism highlights more critical influencing factors, helping the model make more accurate judgments without increasing the model's computational and storage overhead. Therefore, introducing the attention mechanism into LSTM models can effectively highlight factors affecting the load, thereby improving prediction performance.

[0087] This invention introduces the Attention mechanism into the CNN-BiLSTM structure, and the neural network block diagram is as follows. Figure 5 As shown.

[0088] Figure 6 This is a flowchart of a non-intrusive appliance identification method based on an Attention-CNN-BiLSTM neural network according to the present invention. It acquires matching external environmental data and total user electricity consumption data in real time, and uses the aforementioned sliding window bilateral CUSUM event detection method considering time thresholds to detect whether a switching event has occurred based on the acquired active power. If a switching event occurs, an input feature vector is constructed by fusing the steady-state features before and after the event and the transient features at the time of the event, and then input into the trained identification model to obtain the identification result. The Attention-CNN-BiLSTM neural network can be trained to obtain the final identification model. The training dataset can be obtained by constructing a database of photovoltaic power, voltage, and current features from an actual data acquisition platform. Each sample includes an input feature vector: a fused feature obtained by fusing the steady-state features before and after the event and the transient features at the time of the event, and the ground truth value of the appliance.

[0089] This invention analyzes the characteristics of photovoltaic access and operation in non-intrusive interactive energy monitoring, and fully considers the flexible behavior of photovoltaic inverters, improving the detection method. Finally, based on the parameters within the detection time points, input features are constructed and input into a trained Attention-CNN-BiLSTM neural network for identification. This can solve the identification difficulties caused by the uncertainty, continuous variability, and insignificant switching characteristics of photovoltaic equipment power generation, and improve the speed and accuracy of identification.

[0090] Utilizing existing laboratory resources, a physical data acquisition platform was built for high-frequency, low-frequency data collection. Firstly, the laboratory was equipped with a 20kW rooftop photovoltaic system and an experimental electric vehicle for random charging, along with appliances such as refrigerators, air conditioners, and lights. Data was collected using voltage and current sensors to obtain steady-state and transient signals of the intelligent device load. The collected data primarily included active and reactive power, voltage, current, and harmonic data, with a sampling frequency of 100Hz. The data was stored on a dedicated server. The data was divided into training and testing sets. The training set was used to train the identification model, while the testing set was used to evaluate the event detection algorithm, the identification model, and the overall performance of the identification method.

[0091] For multi-class classification problems, the performance can be evaluated using a confusion matrix. The confusion matrix contains the number of classes correctly classified and the number of classes classified as other than the correct one, providing a visual indication of the classification algorithm's performance.

[0092] Precision Rate (PR) represents the proportion of correctly classified events out of all detected events; Recall (RE) represents the proportion of events that were detected out of all detected events. The formulas for calculating these two metrics are as follows:

[0093]

[0094]

[0095] In the formula, TP represents the number of devices correctly classified into the corresponding on state; FP represents the number of devices classified as on but actually off; FN represents the number of devices classified as off but actually on; and TN represents the number of devices classified as off and actually off.

[0096] Precision and recall are often conflicting parameters that are difficult to evaluate due to practical reasons. The F1 score is the harmonic mean of these two metrics, taking both into account. The formula is as follows:

[0097]

[0098] The test results are shown in the table below:

[0099]

[0100] As can be seen, the sliding window bilateral CUSUM event detection algorithm of this invention, which considers the time threshold, is used to detect the switching event. It obtains the ideal average value through the Gauss window and amplifies the climbing event through the Sigmoid function. Compared with other ordinary algorithms, it can accurately obtain the time point when the transient overshoot occurs and continues to the steady state.

[0101] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A sliding window bilateral CUSUM event detection method considering time thresholds, characterized in that, include: Obtain the time series of electrical parameters to be detected, wherein the electrical parameters are active power, reactive power, or effective current values; A sliding window is used to sample the time series of electrical parameters to be tested and to detect whether an event occurs within each sliding window; the sliding window includes a mean calculation window W. m and transient detection window W d Two consecutive sliding windows; for each sliding window, calculate the transient detection window W one by one. d The cumulative sum of input or output events corresponding to each time k within the time frame. And determine: If Then let the time delay factor d = 0; if At that time, let d = d + 1; until The timing of an input or cut-out event is determined by the inverse transformation of t = kd. Where h is the cumulative sum threshold, representing the cumulative sum of the input and output events for each sliding window. The initial value is 0, which is represented as follows: In the formula, β represents the noise introduced from the outside, and Ψ(k) is the weight of k at each time step within the transient detection window calculated based on the Gaussian function; M d It is the average value of the electrical parameters in the current transient detection window; μ gauss It is the average value of the mean calculation window based on the Gaussian function; After confirming that an input or cut-out event has occurred, continue to detect and calculate the cumulative sum of the corresponding input or cut-out events and determine the event end point: if the difference between the cumulative sum of input or cut-out events at the next time step and the current time step is less than the noise threshold, then the current time step is the event end point; If the time difference between two consecutive events is less than a given time threshold, the two consecutive events are merged into one event.

2. The method according to claim 1, characterized in that, The method for calculating the weight of k at each time step within the transient detection window based on Gaussian function calculation is as follows: Among them, u d N represents the median value of the maximum width of the transient detection window; α represents the backward offset of the detection window weights. m N d represents the length of the mean calculation window and the transient detection window, respectively; j is the position of the first sampling point of the current mean calculation window.

3. The method according to claim 1, characterized in that, The method for calculating the average value of the mean calculation window based on the Gaussian function is as follows: The weight representation at each time step within the mean calculation window is determined using a Gaussian function: In the formula u i u represents the position of each sampling point within the average value window. c This indicates the sampling point position in the middle of the mean calculation window; σ is a constant greater than 0. Normalize the weight representation f(ui) to obtain the weight at each time step within the mean calculation window. In the formula, N m This indicates the length of the mean calculation window; j is the position of the first sampling point in the current mean calculation window; The mean value of the Gaussian function calculation window is μ. gauss for: W m (u i ) indicates the mean calculation window within u i The electrical parameter values ​​corresponding to the sampling point locations.

4. The method according to claim 1, characterized in that, The sliding distance of the sliding window is the length of the mean calculation window.

5. The method according to claim 1, characterized in that, The cumulative sum and threshold h are expressed as: h=(Δ min -b)T max ; In the formula: T max =N max T s N max T is the maximum allowed number of delayed sampling points. s Δ is the sampling interval; min Indicates the noise threshold.

6. A non-invasive electrical appliance identification method, characterized in that, include: The user's electricity meter information is acquired in real time, and the sliding window bilateral CUSUM event detection method considering time thresholds as described in any one of claims 1-5 is used to detect whether a switching event has occurred based on the acquired electricity meter information. If a switching event occurs, an input feature vector is constructed based on the meter information before and after the event, and then input into the trained recognition model to obtain the recognition result.

7. The method according to claim 6, characterized in that, The user's electrical appliances include residential photovoltaic equipment.

8. The method according to claim 6, characterized in that, The structure of the recognition model is an Attention-CNN-BiLSTM neural network.

9. The method according to claim 6, characterized in that, The input feature vector includes active power, reactive power, voltage amplitude, and current harmonics.

Images