CUSUM-Bi-LSTM-based edge-side online non-intrusive load identification method

Abstract

A kind of edge side non-invasive load identification method based on CUSUM-Bi-LSTM, the data acquisition is to obtain the active power and reactive power data at user bus;Event detection is to use double-sided CUSUM algorithm of composite dynamic sliding window on power data to detect events, extract the power change data of appliance switching-in and switching-out process period;Feature extraction is to process the detected events, extract the power change amount, cumulative and change amount features of active power and reactive power during event occurrence;Normalize the data feature set, divide the data into training set and test set;Load identification is to input training set data into the model with two one-dimensional convolution layers, Bi-LSTM network and fully connected layer for training, and use precision, recall and F1 score to evaluate the training results;Solidify the trained model into DSP chip for edge side load identification.The present application improves detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of non-intrusive load monitoring (NILM), and in particular to an edge-side online non-intrusive load identification method based on CUSUM-Bi-LSTM. Technical Background

[0002] With the rapid development of power electronics technology, electrical equipment is becoming increasingly diverse. Unsafe and poor electricity usage habits are also leading to an increasing number of electrical fires, resulting in varying degrees of property damage and energy waste. Non-intrusive load identification (NILM) can monitor users' electricity usage behavior and status, helping them understand appliance usage and energy consumption. Furthermore, NILM data helps power companies analyze energy consumption distribution within the power system, providing data support for energy conservation and emission reduction, and achieving optimized grid scheduling. Therefore, NILM technology is of great significance to the currently developing smart grids and smart cities.

[0003] Due to the current large and complex deployment of user devices, the wide variety of electricity consumption, and the wide fluctuation range of electricity consumption characteristics, load identification based on a single characteristic is difficult. Commonly used event detection algorithms struggle to address the problems of false detection caused by steady-state signal disturbances and prolonged transients. Furthermore, current researchers focus solely on the effectiveness of load identification algorithms on the desktop, neglecting the practical application and system design of non-intrusive load identification. Non-intrusive load identification can only achieve its maximum value and engineering significance when applied to the edge. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides an edge-side online non-intrusive load identification method based on CUSUM-Bi-LSTM to improve detection accuracy.

[0005] The technical solution adopted by the present invention to solve its technical problem is:

[0006] An edge-side online non-intrusive load identification method based on CUSUM-Bi-LSTM, the method comprising the following steps:

[0007] Step 1: Obtain active power P and reactive power Q data at the metering location and build an active power sequence and reactive power

[0008] Step 2: Use the double-sided CUSUM algorithm with a composite dynamic sliding window to perform event detection on the active power and reactive power sequence data, respectively, and extract the window time periods when the appliance to be detected is switched on and off from the active power and reactive power sequence data;

[0009] Step 3: Extract the average power value of the active power and reactive power sequence data mean window, the average power value of the detection window, the cumulative sum, the switch-in and switch-out status, and the time period of the event when the event occurs in the load to be detected, and further extract the feature to obtain the mean window mean minus the detection window mean, the cumulative sum value and The power difference within the event time period and the power mean within the event time period form the feature vector Y = [y1, y2, ..., y8];

[0010] Step 4: Normalize the eigenvector. The normalization formula is: y i is the original sample data value, and the minimum value in the sample data is marked as The maximum value in the sample data is marked as The result after the sample data is normalized is marked as The normalized eigenvector is The Bi-LSTM network model is used for recognition, with a DropOut coefficient of 0.5, Adam as the optimizer, and the softmax classifier outputs the recognition results of the load data.

[0011] Furthermore, in step 2, the process of the bilateral CUSUM algorithm is as follows:

[0012] 2.1: Bilateral CUSUM algorithm with dynamic sliding window;

[0013] 2.2: After the bilateral CUSUM algorithm of the dynamic sliding window, continuous jump detection is performed to eliminate events whose starting points are in the same time window.

[0014] Furthermore, in 2.1, a dynamic sliding window is used to divide the data. The window is divided into a mean window MW and a detection window DW. The mean window MW and the detection window DW are set as dynamic windows, and their lengths are represented by L M With L D , the total length of the fixed window is L, when no event occurs, let L M Equal to the maximum value of MW L D Equal to the maximum value L of DW D =0.4L; when the detection window moves forward and detects a transient event, the length of MW is actively adjusted L M =0.4L, and shorten the length of the detection window L D =0.2L, in the DW window, a non-parametric bilateral CUSUM detection algorithm is used. and The expression is:

[0015]

[0016]

[0017] Where x0 is the mean window mean, p k Current power value, β external noise, k sampling time point,

[0018] when and When it is greater than the preset threshold η, it is determined that a transient event has occurred.

[0019] When an event is detected, the detection window in each bilateral sliding window will slide to the bottom of the window separately. The cumulative sum of the detection window will no longer change or will end when it slides to the bottom of the fixed window. A complete event detection is completed.

[0020] The beneficial effects of this invention are demonstrated by: A composite dynamic sliding window bilateral CUSUM event detection algorithm improves detection efficiency, refines the window, and enhances detection accuracy. Transient event feature extraction, combined with multi-feature fusion such as harmonics, reduces computational effort and enhances electrical signatures. The integration of an acquisition device and a DSP data processing chip enables edge-side load identification. This CUSUM-Bi-LSTM non-intrusive edge-side load identification algorithm offers low computational effort and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is an event detection algorithm diagram based on composite dynamic sliding window bilateral CUSUM.

[0022] Figure 2 It is the flow chart of event detection algorithm.

[0023] Figure 3 This is a flowchart of the edge-side non-intrusive load identification method based on CUSUM-Bi-LSTM. DETAILED DESCRIPTION

[0024] The implementation of the present invention is described in detail with reference to the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0025] Reference Figures 1 to 3 , a non-intrusive edge-side load identification method based on CUSUM-Bi-LSTM, comprising the following steps:

[0026] Step 1: Obtain active power and reactive power data at the user bus through the data acquisition device and build an active power sequence and reactive power

[0027] Step 2: Use the double-sided CUSUM algorithm with a composite dynamic sliding window to perform event detection on the active power and reactive power data, respectively, and extract the window time periods for different electrical appliances to be switched on and off from the total power data sequence;

[0028] Step 3: Extract the average power value of the average window, the average power value of the detection window, the cumulative sum, the switch-in and switch-out status, and the time period when the event occurs. Further feature extraction is obtained by subtracting the detection window mean and the cumulative sum from the average window mean. and The power difference and the power mean within the event time period form the feature vector Y = [y1, y2, …, y8];

[0029] Step 4: Normalize the feature vector, input it into the Bi-LSTM network for training, and output the recognition result of the load data through the softmax classifier;

[0030] Step 5: Solidify the trained model into the DSP chip for edge-side load identification.

[0031] In the step 1, the active power and reactive power operation data of each electrical appliance at the user bus are obtained through the data acquisition device to obtain the active power sequence and reactive power

[0032] In step 2, the process of the bilateral CUSUM algorithm is as follows:

[0033] 2.1: Bilateral CUSUM algorithm with dynamic sliding window;

[0034] 2.2: After the bilateral CUSUM algorithm of the dynamic sliding window, continuous jump detection is performed to eliminate events whose starting points are in the same time window.

[0035] Furthermore, in 2.1, a variable-length sliding window is used to divide the sequence P; the window can be divided into a mean window (MW) and a detection window (DW). The mean window (MW) and the detection window (DW) are set as dynamic windows, and their lengths are represented by L and L respectively. M With L D .like Figure 1 As shown in the figure, the total length of the fixed window is L. When no event occurs, let L M Equal to the maximum value of MW L D Equal to the maximum value L of DW D=0.4L; when the detection window moves forward and detects a transient event, the length of MW is actively adjusted L M =0.4L, and shorten the length of the detection window L D =0.2L, in the DW window Using the non-parametric bilateral CUSUM detection algorithm, and The expression is:

[0036]

[0037]

[0038] Where x0 is the MW window mean, p k Current power value, β is the noise introduced by the outside world, and k is the sampling time point.

[0039] When 0<S k When η<η, it means that the sampling time series data has changed, but the fluctuation has not reached the threshold η set in advance. At this time, it is necessary to record the delay time at this time, and record this delay time point as Δ, then let the delay time Δ=Δ+1. and When it is greater than the preset threshold η, it is determined that a transient event has occurred and the time of the event is estimated.

[0040] In step 2, when an event is detected, the detection window in each bilateral sliding window will slide to the bottom of the window separately. The cumulative sum of the detection window no longer changes or ends after sliding to the bottom of the fixed window, and a complete event detection is completed.

[0041] In the second step, after performing the bilateral cumulative CUSUM event detection of the sliding window, continuous jump detection is performed to eliminate events whose starting points are in the same time window.

[0042] In step 3, feature extraction obtains the mean window mean minus the detection window mean, and the positive and negative cumulative sums and The sum of the difference between the maximum and minimum values ​​within the event window, the power mean value and the harmonic components within the event time period are taken as the characteristic vector matrix Y=[y1,y2,…,y8] of an event.

[0043] In step 4, the extracted feature vectors were normalized and input into a Bi-LSTM network. A DropOut factor of 0.5 was added to prevent overfitting, and the fully connected results were output using a Sigmoid function. Adam was used as the optimizer, with an initial learning rate of 0.01 and 100 iterations. The load data recognition results were output using a softmax classifier.

[0044] In step five, the trained CUSUM-Bi-LSTM algorithm model is solidified into the DSP chip, and the serial port of the DSP chip and the USART of the data acquisition device exchange data. The data acquisition device collects data in real time, and the DSP chip performs event detection, feature processing, and load identification on the collected data in real time, and returns the load identification data to the data acquisition device, which uploads it to the host computer.

[0045] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. An edge-side online non-intrusive load identification method based on CUSUM-Bi-LSTM, characterized by: The method comprises the following steps: Step 1: Obtain active power P and reactive power Q data at the metering location and build an active power sequence and reactive power Step 2: Use the double-sided CUSUM algorithm with a composite dynamic sliding window to perform event detection on the active power and reactive power sequence data, respectively, and extract the window time periods when the appliance to be detected is switched on and off from the active power and reactive power sequence data; Step 3: Extract the average power value of the active power and reactive power sequence data mean window, the average power value of the detection window, the cumulative sum, the switch-in and switch-out status, and the time period of the event when the event occurs in the load to be detected, and further extract the feature to obtain the mean window mean minus the detection window mean, the cumulative sum value and The power difference within the event time period and the power mean within the event time period form the feature vector Y = [y1, y2, ..., y8]; Step 4: Normalize the eigenvector. The normalization formula is: ,y i is the original sample data value, and the minimum value in the sample data is marked as The maximum value in the sample data is marked as The result after the sample data is normalized is marked as , the normalized eigenvector is ,Bi-LSTM network model is used for recognition, DropOut with a coefficient of 0.5 is added, Adam is used as the optimizer, and the recognition results of load data are output through the softmax classifier.

2. The CUSUM-Bi-LSTM-based edge-side online non-intrusive load identification method according to claim 1, characterized in that: In step 2, the process of the bilateral CUSUM algorithm is as follows: 2.1: Bilateral CUSUM algorithm with dynamic sliding window; 2.2: After the bilateral CUSUM algorithm of the dynamic sliding window, continuous jump detection is performed to eliminate events whose starting points are in the same time window.

3. The CUSUM-Bi-LSTM-based edge-side online non-intrusive load identification method according to claim 2, characterized in that: In 2.1, a dynamic sliding window is used to divide the data. The window is divided into a mean window MW and a detection window DW. The mean window MW and the detection window DW are set as dynamic windows, and their lengths are represented by L M With L D , the total length of the fixed window is L, when no event occurs, let L M Equal to the maximum value of MW L D Equal to the maximum value L of DW D =0.4L; when the detection window moves forward and detects a transient event, the length of MW is actively adjusted L M =0.4L, and shorten the length of the detection window L D =0.2L, in the DW window, a non-parametric bilateral CUSUM detection algorithm is used. and The expression is: Where x0 is the mean window mean, p k Current power value, β external noise, k sampling time point, when and When it is greater than the preset threshold η, it is determined that a transient event has occurred.

4. The CUSUM-Bi-LSTM-based edge-side online non-intrusive load identification method according to claim 3, characterized in that: When an event is detected, the detection window in each bilateral sliding window will slide to the bottom of the window separately. The cumulative sum of the detection window will no longer change or will end when it slides to the bottom of the fixed window. A complete event detection is completed.

Images