Non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization

Abstract

A non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization includes the following steps: Step 1: extracting voltage and current signals before and after the load state changes, and organizing V-I trajectories according to time labels; Step 2: constructing a three-dimensional spatiotemporal V-I trajectory with a time label; Step 3: encoding the three-dimensional V-I trajectory into a color image with a time stamp; Step 4: constructing a recognition model based on transfer learning, and then inputting the three-dimensional color V-I trajectory image with time characteristics into the recognition model for training and testing; non-intrusive load monitoring is achieved through the above steps. The purpose of the invention is to propose a non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization to address the problem of low accuracy in identifying similar loads and multi-state loads with V-I trajectories.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, in particular to the technical field of load monitoring, and specifically to a non-invasive load monitoring method based on time-enhanced multi-dimensional feature visualization. Background Art

[0002] With the development of economy and society, energy demand continues to grow globally. According to the International Energy Agency, by 2040, global energy demand is expected to increase by 25%, of which electricity demand is expected to increase by 58% compared with 2018. Effective load monitoring is a necessary means to reduce energy consumption in households and industrial sectors.

[0003] Intrusive load identification (ILM) requires the installation of sensors on each of the user's electrical devices to obtain electrical appliance power consumption data, which has high implementation costs and the risk of leaking user privacy. Hart's non-intrusive load identification (NILM) does not require the installation of measurement devices for each appliance. It analyzes the total meter data of the home or factory to obtain the number, type and power consumption of electrical devices, saving costs and protecting user privacy. Non-intrusive load identification (NILM) has become a hot topic among scholars and researchers and is a key technology in smart grids and modern energy management systems.

[0004] Early researchers used a single steady state for load identification. These steady-state features include active power, reactive power, current waveform, steady-state current harmonics, power harmonics, and VI trajectory, among which VI trajectory is widely used due to its good identification effect. Compared with steady-state features, transient features generated during load switching are more closely related to equipment characteristics and can provide more information. Common transient features include power changes under transient conditions, starting current waveforms, voltage noise, edge size or kurtosis of current waveform peaks, etc.

[0005] However, the recognition accuracy of a single feature is limited. Therefore, based on the VI trajectory, the researchers used visualization technology to fuse multiple features to improve the recognition accuracy.

[0006] However, the VI trajectory is formed dynamically. Although some electrical appliances have similar VI trajectories, the formation process is different. The dynamic characteristics of the VI trajectory have been ignored in previous studies, which will result in insufficient resolution of the VI trajectory characteristics. Summary of the invention

[0007] The purpose of the present invention is to address the technical problem that the recognition accuracy of VI trajectory similar loads and multi-state loads is not high, and to propose a non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] The non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization includes the following steps:

[0010] Step 1: Extract the voltage and current signals before and after the load state changes, and organize the VI traces according to the time labels;

[0011] Step 2: Construct a three-dimensional spatiotemporal VI trajectory with time labels;

[0012] Step 3: Encode the 3D VI trajectory into a color image with a time stamp;

[0013] Step 4: Build a recognition model based on transfer learning, and then input the three-dimensional color VI trajectory image with time characteristics into the recognition model for training and testing;

[0014] The above steps can realize non-intrusive load monitoring.

[0015] In step 1, the following sub-steps are included:

[0016] Step 1.1: The aggregated power, current and voltage information can be obtained through the smart sensor to construct a VI curve with a time tag; assuming that only one load's working state changes at a certain moment, the current and voltage signals before and after the load state change are extracted to obtain the load's VI trajectory;

[0017] Step 1.2: Extract the steady-state current and voltage waveforms of T cycles before and after the event, denoted as v on 、v off 、i on and i off , taking the fundamental voltage phase angle as a reference, the periodic voltage waveform is fast Fourier transformed, the first intersection of the fundamental voltage and the horizontal axis is taken as the initial sampling point, and the load voltage and current with time tags are defined as:

[0018] v(t)=(v on +v off ) / 2

[0019] i(t)=(i off -i on ) / 2

[0020] A VI track with time labels can be formed by using the voltage and current in one cycle as the horizontal and vertical axes respectively.

[0021] In step 2, a three-dimensional VI trajectory curve with time labels is constructed, including the following sub-steps:

[0022] Step 2.1: Construct the three-dimensional coordinates of voltage V, current I and time t, taking the time axis as the reference, so that each voltage and current data corresponds to a time point;

[0023] Step 2.2: Based on the time axis, connect any two consecutive points in the sequence to draw a three-dimensional trajectory graph.

[0024] In step 2.1, constructing a three-dimensional coordinate system with a time axis includes the following sub-steps:

[0025] Step 2.1.1: Define the voltage V, current I and time t in the three-dimensional coordinate system to correspond to the X-axis, Y-axis and Z-axis respectively, and normalize the voltage and current data so that they can be compared in a unified scale;

[0026] Step 2.1.2: Use the np.arange function to create a sequence of equally spaced timestamps starting from 0 to the total acquisition time T, specifically:

[0027] t = np.arange(0,T,Δt)

[0028] Among them, the time interval Δt is determined according to the sampling frequency, T is the total acquisition time length, and Δt is the sampling time interval, ensuring that each voltage and current data point has a corresponding time point;

[0029] Next, each voltage and current reading will be mapped to a point in three-dimensional space, where each point represents the voltage and current value at a specific point in time, specifically:

[0030] P i =[V norm (i),I norm (i),t(i)]

[0031] P i Represents a point in three-dimensional space, represented by the normalized voltage V norm (i), normalized current I norm (i) and the corresponding timestamp t(i).

[0032] In step 2.2, consider each point: P i =[V norm (i),I norm (i), t(i)] in three-dimensional space, construct a VI curve with time characteristics, for any continuous point in the sequence [V norm (i),I norm (i),t(i)] and

[0033] [V norm (i+1),I norm(i+1),t(i+1)], construct a line segment connecting these two points. For the line segment between two consecutive points, the parameterized equation is:

[0034]

[0035] λ is a parameter between 0 and 1, representing the i To P i+1 Interpolation of V norm (i+1) refers to the normalized voltage of the next point in the sequence, I norm (i+1) refers to the normalized current of the next point in the sequence, the subscript norm refers to the normalization process, t(i) refers to the timestamp of the current point, and t(i+1) refers to the timestamp of the next point;

[0036] The entire trajectory is formed by connecting the line segments in sequence. The mathematical representation of the trajectory can be regarded as a set of equations of these line segments, specifically:

[0037]

[0038] Where n is the total number of data points, Line i (λ) refers to the fact that the entire trajectory is a set of line segment equations connecting every two adjacent points in the sequence, and i refers to the index of each pair of adjacent points.

[0039] In step 3, in order to improve the feature expression of the three-dimensional VI trajectory with timestamp, the three-dimensional VI trajectory is encoded into a color image with timestamp using hue, saturation and brightness based on the temporal feature; the following sub-steps are included:

[0040] Step 3.1: Use the hue H to identify the direction of movement of the trajectory and record it as

[0041]

[0042] The function atan2(·) is a four-quadrant inverse tangent function, which calculates the phase angle between two consecutive points in the trajectory, and the angle range is from 0° to 360°, V j+1 Refers to the voltage value at point j+1, V j Refers to the voltage value at point j, I j+1 Refers to the current value at point j+1, I j refers to the current value at point j, T j+1 Refers to the timestamp of point j+1, T j refers to the timestamp of point j, v max Refers to the maximum voltage, i max Refers to the maximum current, t max Refers to the maximum value of time;

[0043] Define a new time-series tone matrix H(n x ,n y ), whose elements are obtained by grid (n x ,n y ) is calculated by the hue average value of the trajectory segment, expressed as

[0044]

[0045] Among them, A is a set that contains all the The index of the path point; A = {a|aThe path point passes through the grid |·| is the cardinality of a set;

[0046] It contains temporal features, that is, it is calculated based on the rate of change of the trajectory segment in the time series, and provides a time-varying hue value for each grid point in a 2N×2N matrix;

[0047] Among them, n x Refers to the number of rows in the temporal tone matrix H, n y Refers to the number of columns of the temporal tone matrix H, refers to the row index in the grid, refers to the column index in the grid, H j It refers to the phase angle between the jth pair of consecutive points in the trajectory;

[0048] Step 3.2: Use saturation S to represent the power factor of the load, that is, the ratio of active power to apparent power, specifically:

[0049]

[0050] N is the total number of sampling points, P active is the active power, P apparent is the apparent power, V rms , I rms are the effective values ​​of load voltage and current, respectively, v n Refers to the voltage value of the nth sampling point, i n Refers to the current value at the nth sampling point;

[0051] Step 3.3: Brightness is represented by harmonic characteristics, and the steady-state current signal is decomposed using Fourier transform FFT:

[0052] I=A1sin(ωt+θ1)+A2sin(ωt+θ2)+...+A k sin(ωt+θ k )

[0053] A1,A2,Ak is the amplitude of each harmonic, θ1, θ2, θ k is the phase angle of each harmonic;

[0054] The harmonic amplitude and phase are numerically large, and fusing them with the 3D VI trace with time characteristics requires data processing of the voltage, current, and amplitude phase; to normalize the selected data, the following operations need to be performed:

[0055]

[0056] K represents the original value, K min Indicates the minimum value of the cycle, K max represents the maximum value of the cycle, and K' represents the normalized value;

[0057] Normalized data are rounded as follows:

[0058] K sure =Floor(K'×(N-1))

[0059] A' f =Floor(A f )÷N=x f ...y f

[0060] K sure and A' f represents the final eigenvalue, A f represents the basic amplitude, x f represents the quotient, y f Represents the remainder, Floor(A f ) refers to the basic amplitude A f Take the lower limit.

[0061] Build an image recognition model based on transfer learning, including but not limited to ECA-ResNet34, and then input the three-dimensional color VI trajectory image with time characteristics into the recognition model for training and testing.

[0062] Compared with the prior art, the present invention has the following technical effects:

[0063] 1) The method provided by the present invention is based on the recognition technology of time-enhanced multi-dimensional feature visualization, which improves the accuracy of load identification by integrating steady-state features, transient features and dynamic features. Using the tensorflow2.0 deep learning framework, experiments were conducted on the public data set PLAID, and a sample set was constructed in accordance with the method of the present invention, from which 80% of each device type was extracted as a training set, and the remaining 20% ​​was used as a test set. The overall load identification accuracy reached 98.4%; among them, the accuracy of compact fluorescent lamps, heaters, and laptops reached 100%; the VI trajectories of heaters and hair dryers were very similar, with accuracy rates of 100% and 98.7% respectively; air conditioners, as continuously variable devices, have variable power consumption characteristics and no fixed number of states, making them difficult to accurately identify. Their various operating conditions, such as cooling mode and heating mode, are easily confused with refrigerators and hair dryers, with an accuracy rate of 93.7%; the operating frequency of refrigerators usually changes with temperature, and the VI trajectory is uncertain, with an accuracy rate of 93.8%; multi-state loads such as microwave ovens and washing machines, with accuracy rates of 98.6% and 96.3% respectively. Precision (P), recall (R), the harmonic mean of precision and recall (F1), and total accuracy (A) were selected as evaluation indicators. Most of the P, R, F1, and A of various types of loads remained above 0.95, with average values ​​of 0.969, 0.977, 0.973, and 0.997, respectively.

[0064] 2) The method provided by the present invention uses visualization technology and takes the time axis as the reference to construct a dynamic color VI trajectory based on time information, thereby improving the resolution of the feature. In order to verify that the addition of time information can significantly improve the resolution of the feature, under the same conditions as the present invention, the time information is removed and the experiment is conducted again. The overall load recognition accuracy is 90.5%, and it is 98.4% after adding the time characteristics, which is an increase of 7.9%; for multi-state loads such as air conditioners, refrigerators, and washing machines, the accuracy is 83.9%, 46.5%, and 72.7%, and it is 93.7%, 93.8%, and 96.3% after adding the time characteristics, which is an increase of 9.8%, 47.3%, and 23.6%; the accuracy of hair dryers and heaters is 87% and 80%, and due to the similarity of the trajectories, there is confusion between each other. The addition of time characteristics is 100% and 98.7%, which is an increase of 13% and 18.7%. The accuracy (P), recall (R), harmonic mean of precision and recall (F1) and total accuracy (A) were used as evaluation indicators. The average values ​​of P, R, F1 and A of various types of loads were 0.853, 0.856, 0.850 and 0.983, respectively, which were lower than 0.969, 0.977, 0.973 and 0.997 with the time characteristics added.

[0065] 3) The method provided by the present invention performs transfer learning based on the ECA-ResNet34 recognition model, which improves the applicability and generalization ability of the model. To verify the feasibility of transfer learning, the model trained and tested on the PLAID dataset is migrated to the WHITED dataset using the migration method described in step 4. The results show that all loads have good classification effects. F1-score is selected as the evaluation indicator. Most of the F1-scores are maintained above 99%, and F-macro is 98.8%. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0067] Figure 1 is a flow chart of the present invention;

[0068] Figure 2 Schematic diagram of the ResNet34 image recognition network in the embodiment. DETAILED DESCRIPTION

[0069] like Figure 1 As shown, a non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization includes the following steps:

[0070] Step 1: Extract the voltage and current signals before and after the load state changes, and organize the VI traces according to the time labels;

[0071] Step 2: Establish the three-dimensional coordinates of voltage (V), current (I) and time (t), taking the time axis as the reference, so that each voltage and current data has a corresponding time point to form a three-dimensional coordinate; construct a line segment connecting any two consecutive points in the sequence to form a three-dimensional trajectory diagram;

[0072] Step 3: Using HSV color coding technology, the hue (H) represents the movement direction of the trajectory, the saturation (S) represents the power factor of the load, and the brightness (V) represents the harmonic characteristics, forming a three-dimensional color VI trajectory image with a time label;

[0073] Step 4: Build an ECA-ResNet34 recognition model based on transfer learning, and input the three-dimensional color VI trajectory image with time characteristics into the network for training.

[0074] In step 1, data preprocessing is performed to organize VI traces according to time tags, including the following sub-steps:

[0075] Step 1.1: The smart sensor can obtain aggregated power, current and voltage information to construct a VI curve with a time tag. Assuming that only one load has a change in working state at a certain moment, the current and voltage signals before and after the load state change can be extracted to obtain the VI trajectory of the load.

[0076] Step 1.2: Extract the steady-state current and voltage waveforms of T cycles before and after the event, denoted as v on 、v off 、i on and i off Taking the fundamental voltage phase angle as a reference, the periodic voltage waveform is subjected to a fast Fourier transform, and the first intersection of the fundamental voltage and the horizontal axis is taken as the initial sampling point. The load voltage and current with time labels are defined as

[0077] v(t)=(v on +v off ) / 2

[0078] i(t)=(i off -i on ) / 2

[0079] A VI track with time labels can be formed by using the voltage and current in one cycle as the horizontal and vertical axes respectively.

[0080] In step 2, a three-dimensional VI curve with time labels is constructed, including the following sub-steps:

[0081] Step 2.1: Construct the three-dimensional coordinates of voltage (V), current (I) and time (t), taking the time axis as the reference, so that each voltage and current data corresponds to a time point;

[0082] Step 2.2: Based on the time axis, connect any two consecutive points in the sequence to draw a three-dimensional trajectory graph.

[0083] In step 2.1, how to construct a three-dimensional coordinate with a time axis includes the following sub-steps:

[0084] Step 2.1.1: Define the voltage (V), current (I), and time (t) in the three-dimensional coordinate system corresponding to the X-axis, Y-axis, and Z-axis, respectively. Normalize the voltage and current data so that they can be compared on a uniform scale.

[0085] Step 2.1.2: Use the np.arange function to create a sequence of equally spaced timestamps starting from 0 to the total acquisition time T. The time interval Δt is determined based on the sampling frequency t = np.arange(0,T,Δt)

[0086] T is the total acquisition time length, and Δt is the sampling time interval. Make sure that each voltage and current data point has a corresponding time point.

[0087] Each voltage and current reading is then mapped to a point in three-dimensional space. Each point represents the voltage and current value at a specific point in time.

[0088] P i =[V norm (i),I norm (i),t(i)]

[0089] P i Represents a point in three-dimensional space, represented by the normalized voltage V norm (i), normalized current I norm (i) and the corresponding timestamp t(i).

[0090] In step 2.2, considering each point P i =[V norm (i),I norm (i), t(i)] in three-dimensional space, a VI curve with time characteristics can be constructed. For any continuous point in the sequence [V norm (i),I norm (i),t(i)] and

[0091] [V norm (i+1),I norm (i+1),t(i+1)], construct a line segment connecting these two points. For a line segment between two consecutive points, the parameterized equation is

[0092]

[0093] λ is a parameter between 0 and 1, representing the i To P i+1 Interpolation of V norm (i+1) refers to the normalized voltage of the next point in the sequence, I norm (i+1) refers to the normalized current of the next point in the sequence, the subscript norm refers to the normalization process, t(i) refers to the timestamp of the current point, and t(i+1) refers to the timestamp of the next point.

[0094] The entire trajectory is formed by connecting these line segments in sequence. The mathematical representation of the trajectory can be regarded as a collection of these line segment equations

[0095]

[0096] Where n is the total number of data points, Line i(λ) refers to the fact that the entire trajectory is a set of line segment equations connecting every two adjacent points in the sequence, and i refers to the index of each pair of adjacent points.

[0097] In step 3, in order to improve the feature expression of the three-dimensional VI trajectory with timestamp, the three-dimensional VI trajectory is encoded into a color image with timestamp using hue, saturation and brightness based on the temporal feature. The following sub-steps are included:

[0098] Step 3.1: Use hue (H) to identify the direction of movement of the trajectory and record it as

[0099]

[0100] The function atan2(·) is a four-quadrant inverse tangent function, which calculates the phase angle between two consecutive points in the trajectory, and the angle range is from 0° to 360°, V j+1 Refers to the voltage value of point jv1, V j Refers to the voltage value at point j, I j+1 Refers to the current value at point j+1, I j refers to the current value at point j, T j+1 Refers to the timestamp of point j+1, T j refers to the timestamp of point j, v max Refers to the maximum voltage, i max Refers to the maximum current, t max Refers to the maximum time.

[0101] Define a new time-series tone matrix H(n x ,n y ), whose elements are obtained by grid (n x ,n y ) is calculated by the hue average value of the trajectory segment. It is expressed as

[0102]

[0103] Among them, A is a set that contains all the The index of the path point; A = {a|aThe path point passes through the grid |·| is the cardinality of a set.

[0104] Contains temporal features, i.e. it is calculated based on the rate of change of the trajectory segment in the time series. A time-dependent hue value will be provided for each grid point in a 2N×2N matrix.

[0105] Among them, n x Refers to the number of rows in the temporal tone matrix H, n y Refers to the number of columns of the temporal tone matrix H, refers to the row index in the grid, refers to the column index in the grid, H j Refers to the phase angle between the jth pair of consecutive points in the trajectory.

[0106] Step 3.2: Use saturation (S) to express the power factor of the load, which is the ratio of active power to apparent power.

[0107]

[0108] N is the total number of sampling points, P active is the active power, P apparent is the apparent power, V rms , I rms are the effective values ​​of load voltage and current, respectively, v n Refers to the voltage value of the nth sampling point, i n Refers to the current value at the nth sampling point.

[0109] Step 3.3: Brightness (V) is represented by harmonic characteristics. Decompose the steady-state current signal using Fourier transform (FFT)

[0110] I=A1sin(ωt+θ1)+A2sin(ωt+θ2)+...+A k sin(ωt+θ k )

[0111] A1,A2,A k is the amplitude of each harmonic, θ1, θ2, θ k is the phase angle of each harmonic.

[0112] The harmonic amplitude and phase are numerically large, and fusing them with the 3D VI trace with time characteristics requires data processing of the voltage, current, and amplitude phase. To normalize the selected data, you need to do the following:

[0113]

[0114] K represents the original value, K min Indicates the minimum value of the cycle, K max represents the maximum value of the cycle. K' represents the normalized value.

[0115] Normalized data are rounded as follows:

[0116] K sure =Floor(K'×(N-1))

[0117] A' f =Floor(A f )÷N=x f...y f

[0118] K sure and A' f represents the final eigenvalue, A f represents the basic amplitude, x f represents the quotient, y f Represents the remainder, Floor(A f ) refers to the basic amplitude A f Take the floor (round down).

[0119] In step 4, an image recognition model based on transfer learning is constructed, including but not limited to ECA-ResNet34, and then the three-dimensional color VI trajectory image with time characteristics is input into the recognition model for training and testing, including the following sub-steps:

[0120] Step 4.1: Build the ResNet34 image recognition network as follows Figure 2 As shown;

[0121] The ResNet34 residual network is easy to optimize and can improve accuracy by increasing the depth. Its internal residual blocks use skip connections to alleviate the gradient vanishing problem caused by increasing the depth in the deep neural network. Its network structure is shown in the figure below:

[0122] Step 4.2: Add the attention mechanism ECA, use a 1×1 convolution layer directly after the global average pooling layer, and remove the fully connected layer; this module avoids dimensionality reduction and effectively captures cross-channel interactions, involving only a few parameters to achieve good results.

[0123] Step 4.3: Implement cross-domain transfer learning. The convolutional and fully connected layers of the pre-trained ECA-ResNet34 model are considered as a cascade of feature extractors, assuming that the first few layers can be transferred between different tasks. By converting the load features into visual representations and replacing the last fully connected layer with the same size as the class of the electrical load, it can be generalized to any other dataset and implement cross-domain transfer learning.

[0124] Example:

[0125] A non-invasive load monitoring method based on time-enhanced multidimensional feature visualization. The method aims to identify loads with similar VI trajectories and multiple states. First, on the basis of the two-dimensional gray VI trajectory, the voltage and current change rate, power factor, third harmonic and its dynamic characteristics are fused, and combined with the HSV color coding technology to form a three-dimensional spatiotemporal color VI trajectory to make up for the lack of time information. Then, the ECA-ResNet34 network model based on transfer learning is used for load identification, which avoids the problems of network degradation and training difficulties caused by the excessive depth of traditional convolutional neural networks (CNNs), and realizes efficient monitoring of household loads. Using the method described in the present invention, the formation process of the VI trajectory can be intuitively observed, and similar loads that are difficult to identify in traditional methods can be effectively distinguished, which solves the problem of insignificant differences between similar loads in load identification, and has practical value in the field of NILM.

[0126] The specific steps include:

[0127] Step 1: Two public datasets, PLAID and WHITED, are used to validate the proposed load identification and machine learning methods. The PLAID dataset includes 1074 current and voltage records of 11 different device types from more than 55 households in Pittsburgh, Pennsylvania, USA, and the WHITED dataset is a global household and industrial transient energy dataset, and contains 1259 records of 54 different device types in different regions around the world. The PLAID dataset shows richer intra-device variations, while the WHITED dataset covers a wider range of inter-device differences. Both are recognized high-resolution appliance measurement datasets for non-intrusive load monitoring (NILM) evaluation. 80% of each device type is extracted as a training set, and the remaining 20% ​​is used as a test set.

[0128] Step 2: The results on the PLAID dataset show that the overall load identification accuracy reaches 98.4%. Among them, the accuracy of compact fluorescent lamps, heaters, and laptops reaches 100%. The VI trajectories of heaters and hair dryers are very similar, with accuracy rates of 100% and 98.7%, respectively. The accuracy rates of multi-state loads such as air conditioners, refrigerators, microwave ovens, and washing machines are 93.7%, 93.8%, 98.6%, and 96.3%, respectively. The evaluation is carried out using precision (P), recall (R), the harmonic mean of precision and recall (F1), and the total accuracy (A) as evaluation indicators. Most of the P, R, F1, and A of various loads remain above 0.95, with average values ​​of 0.969, 0.977, 0.973, and 0.997, respectively.

[0129]

[0130] Table 1 Confusion matrix evaluation indicators of PLAID dataset

[0131] Step 3: To verify that the addition of time information can significantly improve the recognition accuracy, the experiment was conducted again without the time information under the same model and conditions as in this paper. The overall load recognition accuracy is 90.5%, and it is 98.4% after adding the time characteristics, an increase of 7.9%; for multi-state loads such as air conditioners, refrigerators, and washing machines, the accuracy is 83.9%, 46.5%, and 72.7%, and it is 93.7%, 93.8%, and 96.3% after adding the time characteristics, an increase of 9.8%, 47.3%, and 23.6%; the accuracy of hair dryers and heaters is 87% and 80%, and due to the similarity of the trajectories, there is confusion between them. After adding the time characteristics, it is 100% and 98.7%, an increase of 13% and 18.7%. The evaluation is carried out with the precision (P), recall (R), the harmonic mean of the precision and recall (F1), and the total accuracy (A) as evaluation indicators. The average values ​​of P, R, F1, and A for each type of load are 0.853, 0.856, 0.850, and 0.983, respectively, which are all lower than 0.969, 0.977, 0.973, and 0.997 when the time characteristics are added.

[0132]

[0133] Table 2 Confusion matrix evaluation indicators on the PLAID dataset without time information

[0134] Step 4: To verify the feasibility of transfer learning, experiments were conducted on the WHITED dataset. The ECA-ResNet34 model trained on the PLAID dataset was used to extract all layers except the last fully connected layer, and the last fully connected layer of the original model was replaced with a new fully connected layer, and applied to the new device classification task. The results showed that all loads had good classification effects, with F1-score mostly maintained above 99% and F-macro at 98.8%.

Claims

1. A non-intrusive load monitoring method based on time-enhanced multi-dimensional feature visualization, characterized in that: The following steps are involved: Step 1: Extract the voltage and current signals before and after the load state changes, and organize the VI traces according to the time labels; Step 2: Construct a three-dimensional spatiotemporal VI trajectory with time labels; Step 3: Encode the 3D VI trajectory into a color image with a time stamp; Step 4: Build a recognition model based on transfer learning, and then input the three-dimensional color VI trajectory image with time characteristics into the recognition model for training and testing; Through the above steps, non-intrusive load monitoring is achieved; In step 3, in order to improve the feature expression of the three-dimensional VI trajectory with timestamp, the three-dimensional VI trajectory is encoded into a color image with timestamp using hue, saturation and brightness based on the temporal feature; It includes the following sub-steps: Step 3.1: Use the hue H to identify the direction of movement of the trajectory and record it as: The function atan2(·) is a four-quadrant inverse tangent function, which calculates the phase angle between two consecutive points in the trajectory, and the angle range is from 0° to 360°, V j+1 Refers to the voltage value at point j+1, V j Refers to the voltage value at point j, I j+1 Refers to the current value at point j+1, I j refers to the current value at point j, T j+1 Refers to the timestamp of point j+1, T j refers to the timestamp of point j, v max Refers to the maximum voltage, i max Refers to the maximum current, t max Refers to the maximum value of time; Define a new time-series tone matrix H(n x ,n y ), whose elements are obtained by grid (n x ,n y ) is calculated by the hue average value of the trajectory segment, expressed as Among them, A is a set that contains all the The index of the path point; A = {a|aThe path point passes through the grid |·| is the cardinality of a set; It contains temporal features, that is, it is calculated based on the rate of change of the trajectory segment in the time series, and provides a time-varying hue value for each grid point in a 2N×2N matrix; Among them, n x Refers to the number of rows in the temporal tone matrix H, n y Refers to the number of columns of the temporal tone matrix H, refers to the row index in the grid, refers to the column index in the grid, H j It refers to the phase angle between the jth pair of consecutive points in the trajectory; Step 3.2: Use saturation S to represent the power factor of the load, that is, the ratio of active power to apparent power, specifically: N is the total number of sampling points, P active is the active power, P apparent is the apparent power, V rms , I rms are the effective values ​​of load voltage and current, respectively, v n Refers to the voltage value of the nth sampling point, i n Refers to the current value at the nth sampling point; Step 3.3: Brightness is represented by harmonic characteristics, and the steady-state current signal is decomposed using Fourier transform FFT: I=A1sin(ωt+θ1)+A2sin(ωt+θ2)+...+A k sin(ωt+θ k ) A1,A2,A k is the amplitude of each harmonic, θ1, θ2, θ k is the phase angle of each harmonic; The harmonic amplitude and phase are numerically large, and fusing them with the 3D VI trace with time characteristics requires data processing of the voltage, current, and amplitude phase; to normalize the selected data, the following operations need to be performed: K represents the original value, K min Indicates the minimum value of the cycle, K max represents the maximum value of the cycle, and K' represents the normalized value; Normalized data are rounded as follows: K sure =Floor(K'×(N-1)) A' f =Floor(A f )÷N=x f ...y f K sure and A' f represents the final eigenvalue, A f represents the basic amplitude, x f represents the quotient, y f Represents the remainder, Floor(A f ) refers to the basic amplitude A f Take the lower limit.

2. The method according to claim 1, characterized in that In step 1, the following sub-steps are included: Step 1.1: Obtain aggregated power, current and voltage information through smart sensors and construct a VI curve with time tags; assume that only one load changes its working state at a certain moment, extract the current and voltage signals before and after the load state changes, and obtain the VI trajectory of the load; Step 1.2: Extract the steady-state current and voltage waveforms of T cycles before and after the event, denoted as v on 、v off 、i on and i off , taking the fundamental voltage phase angle as a reference, the periodic voltage waveform is fast Fourier transformed, the first intersection of the fundamental voltage and the horizontal axis is taken as the initial sampling point, and the load voltage and current with time tags are defined as: v(t)=(v on +v off ) / 2 i(t)=(i off -i on ) / 2 The voltage and current in one cycle are used as the horizontal axis and the vertical axis respectively to form a VI track with a time label.

3. The method according to claim 1, characterized in that In step 2, a three-dimensional VI trajectory curve with time labels is constructed, including the following sub-steps: Step 2.1: Construct the three-dimensional coordinates of voltage V, current I and time t, taking the time axis as the reference, so that each voltage and current data corresponds to a time point; Step 2.2: Based on the time axis, connect any two consecutive points in the sequence to draw a three-dimensional trajectory graph.

4. The method according to claim 3, characterized in that In step 2.1, constructing a three-dimensional coordinate system with a time axis includes the following sub-steps: Step 2.1.1: Define the voltage V, current I and time t in the three-dimensional coordinate system to correspond to the X-axis, Y-axis and Z-axis respectively, and normalize the voltage and current data so that they can be compared in a unified scale; Step 2.1.2: Use the np.arange function to create a sequence of equally spaced timestamps starting from 0 to the total acquisition time T, specifically: t = np.arange(0,T,Δt) Among them, the time interval Δt is determined according to the sampling frequency, T is the total acquisition time length, and Δt is the sampling time interval, ensuring that each voltage and current data point has a corresponding time point; Next, each voltage and current reading will be mapped to a point in three-dimensional space, where each point represents the voltage and current value at a specific point in time, specifically: P i =[V norm (i),I norm (i),t(i)] P i Represents a point in three-dimensional space, represented by the normalized voltage V norm (i), normalized current I norm (i) and the corresponding timestamp t(i).

5. The method according to claim 4, characterized in that In step 2.2, consider for each point: P i =[V norm (i),I norm (i), t(i)] in three-dimensional space, construct a VI curve with time characteristics, for any continuous point in the sequence [V norm (i),I norm (i),t(i)] and [V norm (i+1),I norm (i+1),t(i+1)], construct a line segment connecting these two points. For the line segment between two consecutive points, the parameterized equation is: λ is a parameter between 0 and 1, representing the i To P i+1 Interpolation of V norm (i+1) refers to the normalized voltage of the next point in the sequence, I norm (i+1) refers to the normalized current of the next point in the sequence, the subscript norm refers to the normalization process, t(i) refers to the timestamp of the current point, and t(i+1) refers to the timestamp of the next point; The entire trajectory is formed by connecting the line segments in sequence. The mathematical representation of the trajectory is regarded as a set of equations of these line segments, specifically: Where n is the total number of data points, Line i (λ) refers to the fact that the entire trajectory is a set of line segment equations connecting every two adjacent points in the sequence, and i refers to the index of each pair of adjacent points.

6. The method according to claim 1, characterized in that In step 4, an image recognition model based on transfer learning is constructed, and the model includes ECA-ResNet34.

Images