A power load identification method based on dual-state feature matching and cloud-edge collaboration

Abstract

The present invention relates to the technical field of electric load identification, and specifically to an electric load identification method based on dual-state feature matching and cloud-edge-end collaboration, including: constructing a subset of load event startup and shutdown event features, establishing a load event and power-consuming equipment correlation model as a first model; using the node order optimization strategy of the Gini index, developing a load identification method of a directed acyclic graph support vector machine as a second model; constructing a load identification algorithm based on Mediod shift clustering and steady-state feature analysis as a third model; integrating the first model, the second model and the third model for training to construct a deep neural network load identification model based on a deep learning Bi-LSTM algorithm on the cloud, and establishing a cloud-edge-end collaborative identification mechanism. The present invention effectively judges the start-stop process of different electrical appliances, thereby achieving more accurate and efficient electric load identification.

Description

Technical Field

[0001] The present invention relates to the technical field of power load identification, and specifically to a power load identification method based on binary feature matching and cloud-edge-end collaboration. Background Art

[0002] With the rapid development of smart grid and smart home technologies, energy management and smart power systems are moving towards a more efficient, automated and intelligent direction. At the same time, non-invasive power load identification technology has gradually become one of the core technologies in energy management and smart home systems. Traditional load identification methods face the challenges of multiple electrical appliances in parallel, complex power environments and differentiated characteristics between different devices. For example, to monitor each electrical appliance separately, it is usually necessary to physically access the device or install additional sensors to obtain real-time power data. This method not only increases the installation cost of the equipment, but also limits its application in complex environments. In particular, when multiple electrical appliances are running in parallel, it is difficult to accurately identify the switch status of different electrical appliances.

[0003] Existing technologies often rely on the analysis of a single feature, and it is difficult to maintain high accuracy and efficiency in a variety of power usage scenarios, especially in the process of load event detection and feature extraction, which is prone to misjudgment. For example, residential users have a wide variety of electrical equipment and different brands, and electrical equipment may usually have multiple working states, which makes the load sample feature library label and the electrical equipment have a certain degree of irrelevance, thereby affecting the subsequent load identification performance of different categories. This shows that traditional load identification methods can no longer meet the needs of modern power systems for high-precision and high-efficiency load intelligent identification. Therefore, there is an urgent need for a new load identification method that can adapt to the parallel use of multiple electrical appliances and effectively respond to changes in complex power environments to meet the needs of smart grids and smart home systems in practical applications. Summary of the invention

[0004] In order to solve the technical problem of low load identification performance of the above-mentioned traditional load identification method, the purpose of the present invention is to provide a power load identification method based on dual-state feature matching and cloud-edge collaboration. The technical solution adopted is as follows:

[0005] An embodiment of the present invention provides a method for identifying power loads based on binary feature matching and cloud-edge-end collaboration, the method comprising the following steps:

[0006] S1, based on the optimized CUSUM composite sliding window accumulation and change point detection algorithm, the feature detection of load events is carried out, the feature subsets of load event startup and shutdown events are constructed, and the correlation model between load events and power-consuming equipment is established as the first model;

[0007] S2, combined with the labeling feature library of transient event operation characteristics, using the node order optimization strategy of the Gini index, develops a load identification method based on a directed acyclic graph support vector machine as the second model;

[0008] S3, combined with the labeling feature library of steady-state event operation characteristics, uses time characteristics to refine the classification of load events, and constructs a load identification algorithm based on Mediod shift clustering and steady-state feature analysis as the third model;

[0009] S4, integrates the first model, the second model and the third model for training to build a deep neural network load identification model based on the cloud-based deep learning Bi-LSTM algorithm, and establishes a distributed resource perception technology and cloud-edge-end collaborative identification mechanism with multi-feature matching and cloud-edge-end collaboration.

[0010] Furthermore, the statistical function in the optimized CUSUM composite sliding window accumulation and change point detection algorithm in step S1 is:

[0011] 1) Assume that a change point occurs at the kth moment in the random time series, and initialize the statistical function in the non-parametric bilateral CUSUM algorithm, as shown in the first and second formulas below:

[0012] ;

[0013] In the first formula, and represent the 0th moment, the kth moment, and the The positive CUSUM value at the moment, max means taking the maximum value function, It represents the positive offset of the detection point at the kth moment after removing the influence of the noise value; ; In the calculation formula of the positive offset, represents the data at the kth moment in a random time series, represents the average value of the data before the kth moment in the random time series. represents the noise value, k is a positive integer;

[0014] ;

[0015] In the second formula, and represent the 0th moment, the kth moment, and the The reverse CUSUM value at the moment, It represents the negative offset of the detection point at the kth moment after removing the influence of the noise value; ;

[0016] 2) Reduce the accumulated weight ratio of the detection window start segment, as shown in the third formula below:

[0017] ; In the third formula, represents the weight of sampling point i in the detection window, Indicates the middle value of the maximum width of the detection window, represents a constant greater than 0, Represents the position of sampling point i in the detection window, and its calculation formula is: , ; In the calculation formula, Indicates the width of the detection window, represents the width of the mean window, Indicates the position of the sampling point where the average window starts, i represents 0 to A positive integer between ;

[0018] right Normalize as shown in the fourth formula below:

[0019] ; In the fourth formula, Represents the normalized weight of sampling point i in the detection window;

[0020] 3) By detecting the normalized weights of the sampling points in the window, the statistical function in the non-parametric CUSUM algorithm based on the improved Sigmoid function is determined. The formula is as follows:

[0021] ; .

[0022] Furthermore, the construction of the load event startup and shutdown event feature subsets and the establishment of a load event and power equipment association model includes:

[0023] 1) In the definition of transient characteristics, the power within one cycle is its effective value, as shown in the fifth formula below:

[0024] ; In the fifth formula, Indicates the intensity of the current, N indicates the number of sampling points in one cycle, Indicates the current value of the nth sampling point in a cycle;

[0025] 2) According to the intensity of the current at each sampling point, the separation process of the transient process is set to obtain the separated transient waveform, and then the number of cycles of the current transient intensity change greater than the transient start judgment threshold is sampled and recorded, and the number of cycles of the current transient intensity change less than the transient end judgment threshold is sampled and recorded;

[0026] 3) Steady-state feature extraction based on particle resampling, taking the feature at each moment as a particle, and recalculating the importance of each particle and resampling it according to the distribution of particles before and after the load event point, until the sum of the variances of all particles is minimized, and finally combining the transient features to obtain the feature subsets of the load event startup and shutdown events;

[0027] 4) Establish a load event bipartite graph model, where is an undirected simple graph with vertices , edge , divide the vertex V into two disjoint subsets X and Y, satisfying , so that the two endpoints of each edge in graph G are located in X and Y respectively, is a bipartite graph, and its associated cost matrix is ​​expressed as: ; In the expression of the associated cost matrix, Represents a vertex and The association cost parameter between them, n represents the number of rows and columns of the matrix;

[0028] 5) Based on the Hungarian algorithm, for bipartite graphs Seeking the optimal match, the load matching mathematical model is shown in the sixth formula below:

[0029] ; In the sixth formula, min represents the minimum value function, Represents the cost parameter the sum of It represents the matching degree between different jump powers, n represents the number of load closing events, and m represents the number of load opening events;

[0030] Define the cost function The seventh formula is: ; In the seventh formula, represents the grey relational coefficient of the pth characteristic component, ,satisfy , , Indicates that m load start events are retrieved within a period of time t and n closing events ;

[0031] Defining the grey correlation coefficient The eighth formula is: ; In the eighth formula, represents the reference value of the jth eigencomponent, Indicates The dimensionless attribute value of the feature component after normalization, Represents the resolution coefficient of grey relational degree;

[0032] The decision matrix is ​​established as shown in the ninth formula below:

[0033] ;

[0034] In the ninth formula, R represents the decision matrix, represents the reference value of the characteristic component, Indicates feature components;

[0035] 6) Improve the Hungarian matching algorithm and expand the characteristic component p at time t+Δt, as shown in the following tenth formula:

[0036] ; In the tenth formula, It means that k loads are turned on or off at the same time at time t. Represents a period of time, express The first derivative with respect to time, represents the characteristic component p at time t+Δt; ; In the formula, represents the value of the p characteristic components of the nth load, and k represents the number of loads;

[0037] The remaining part is equivalent to other possible load characteristic components, rewritten as: ; In the formula, represents the value of the p characteristic components of the nth load, Represents the value of the p characteristic components of the mth load;

[0038] 7) Modify the original cost matrix and add the original zero elements , and replace the non-zero elements with The value of The calculation formula is as follows: ; In the formula, represents the normalized cost matrix, represents the maximum element in the cost matrix C, Represents the minimum element in the cost matrix C;

[0039] 8) Use the Hungarian algorithm to solve the new matrix , the expression of the new matrix is:

[0040] ;

[0041] 9) Adopt the transient feature template matching clustering algorithm based on proximity to carry out association modeling and obtain the correlation model between load events and power equipment;

[0042] The definition of the posting progress is shown in the following eleventh formula: ; In the eleventh formula, n represents the dimension of the vector, represents the difference between vector x and vector y, and Represents the normalized vector x and vector y;

[0043] The calculation formula based on the progress of the amplitude is: ; In the calculation formula of the progress of posting, and Represents vectors With vector The maximum element value in represents the vector dimension, .

[0044] Furthermore, the separation implementation process of the transient process is set according to the intensity of the current at each sampling point to obtain the separated transient waveform, including:

[0045] The first step is to import the transient waveform extracted by the detection algorithm;

[0046] The second step is to calculate the current intensity before and after the transient process occurs. If the current before the transient process occurs is less than or equal to the current after the transient process occurs, the third step is entered, otherwise the fourth step is entered.

[0047] The third step is to subtract the waveform of the last steady-state current cycle before the transient state occurs from the transient waveform in the first step by cycle.

[0048] The fourth step is to subtract the waveform of the first steady-state current cycle after the transient state occurs from the transient waveform in the first step by cycle;

[0049] Step 5: Output the separated transient waveform.

[0050] Furthermore, the steady-state feature extraction based on particle resampling takes the feature at each moment as a particle, and recalculates the importance of each particle and resamples it according to the distribution of particles before and after the load event point until the sum of the variances of all particles is minimized, including:

[0051] In the first step, assume that the load input change point is t, set k = 0, randomly obtain N particles after the current time t and before the next change point, and initialize each particle i so that it has an equal weight, where i is a positive integer;

[0052] The second step is importance sampling. For each particle i, the importance weight is estimated according to the degree of deviation from its center. , as shown in the twelfth formula below: ; In the twelfth formula, represents a particle set, Represents particles and a particle in the particle set The Euclidean distance between Normalize to get ;

[0053] The third step is resampling. According to the importance weight in the second step, particles with importance weights less than the threshold are discarded, and replacement sampling is performed based on the discarded particles.

[0054] The fourth step is to increase k by 1 and repeat the process of the second and third steps until the variance change in the particle group is maximized.

[0055] The fifth step is to perform differential analysis on the load characteristics in the currently obtained particle set and the previously recorded load characteristics to extract the load change characteristics.

[0056] Furthermore, the node order optimization strategy of the Gini index is used to develop the basic principles of the load identification method of the directed acyclic graph support vector machine, including:

[0057] 1) Assume that the sample set is as shown in the following formula 13: ;

[0058] In the thirteenth formula, represents sample data, represents the sample category label, N represents the number of samples, and n represents the sample dimension. represents n-dimensional real number space;

[0059] 2) Construct the optimal hyperplane to separate samples of different categories, and the optimization problem is shown in the following fourteenth formula:

[0060] ; , ;

[0061] In the fourteenth formula, W represents the normal vector of the hyperplane, b represents the offset of the hyperplane, and C represents the penalty factor. is the slack variable, represents the transpose of the hyperplane normal vector, Represents the optimization objective function The minimum value of , i represents the sample index, and m represents the number of samples;

[0062] 3) Construct the Lagrangian function, as shown in the fifteenth formula below:

[0063] ;

[0064] In the fifteenth formula, and is the Lagrange multiplier;

[0065] 4) Solve the optimization problem and make Find the partial derivatives and get the optimal dual problem:

[0066] ; The constraints are: ;

[0067] By solving the optimization problem, the classification decision surface is obtained;

[0068] 5) Lagrange multipliers obtained based on the solution of the dual problem , the sixteenth formula for determining the classification decision surface is as follows:

[0069] ; In the sixteenth formula, The corresponding samples are called support vectors, and sign represents the positive and negative sign function;

[0070] 6) The seventeenth formula using particle swarm optimization is as follows:

[0071] ; In the seventeenth formula, represents the velocity of the particle in the dth dimension, represents the inertia weight, and represents the learning factor, and Represents a random number, represents the best historical position of the particle, represents the global optimal position, Indicates the current position of the particle. represents the constraint factor;

[0072] 7) The sample set is divided into two subsets by the node and , the eighteenth formula of the Gini index of the sample after this division is as follows:

[0073] ; ; In the eighteenth formula, Represents a sample set The Gini index, Indicates The square of the class sample proportion, Represents a sample set Divided into and The Gini index after

[0074] The resulting nineteenth formula for the reduction in the Gini index is as follows: ; In the nineteenth formula, Indicates the degree of reduction of impurity before and after the sample set is divided;

[0075] 8) Select the node with the largest Gini index difference as the root node, then regard D1 and D2 as independent sets respectively, select the classifier with the largest Gini index difference as the current node in the next layer, and so on, to build a directed acyclic graph SVM structure.

[0076] Furthermore, the load identification algorithm based on Mediod shift clustering and steady-state characteristic analysis includes:

[0077] 1) Given sampling points , the twentieth formula for the probability density estimate at point x is as follows:

[0078] ; In the 20th formula, represents the probability density estimate at the sampling point x, represents the Gaussian kernel function, h represents the fixed bandwidth, and d represents the dimension;

[0079] 2) The cluster center is updated by minimization, as shown in the following formula 21:

[0080] ; In the twenty-first formula, k represents the number of iterations, Indicates the current center point. Represents the next step center point obtained by iteration and is selected from the sample set The sampling points that satisfy the update conditions are: represents the weight function;

[0081] 3) Segment the load operation time characteristics and divide a day into b time periods. The calculation formula of the objective function is as follows: ; In the calculation formula of the objective function, Indicates The number of time periods occupied by a load, Indicates the number of load types;

[0082] 4) Load operation period through sparse binary matrix Indicates that, assuming Represents 8 load environments, and the matrix expression is as follows:

[0083] ; In the expression, Indicates the load status of the first load environment in the first time period, where 1 indicates that the load exists and 0 indicates that the load does not exist;

[0084] 5) If and Moment (where ) is the split time point, then The calculation formula is as follows:

[0085] ; In the formula, , ;

[0086] 6) Based on Gaussian and density statistical methods, the one-dimensional characteristics of the operating time of all steady-state loads are analyzed, and load events are divided into two categories: regular and irregular;

[0087] 7) Use Mediod Shift clustering to refine the classification of load events.

[0088] Furthermore, the use of Mediod Shift clustering to refine and classify load events includes:

[0089] The first step is to set the cluster bandwidth based on the load power characteristic statistical model. The threshold at which the mean value stops drifting ;

[0090] The second step is to calculate the power mean drift vector of each load event sample in the interval ;

[0091] The third step is to translate the position of each sample point. By calculating the drift vector, as shown in the following formula 22:

[0092] ; In the 22nd formula, represents the drift vector;

[0093] Step 4: Repeat steps 2 and 3 until Stop when.

[0094] Furthermore, the deep neural network load identification model based on the Bi-LSTM algorithm includes:

[0095] 1) When the Bi-LSTM algorithm is forward calculated, the output of the hidden layer The hidden layer output at the previous moment The influence of; when calculating in reverse, the output of the hidden layer The hidden layer output at the next moment The output layer is at time Output The output of the forward and backward hidden layers and Decide jointly, as shown in Formula 23 below:

[0096] ; In the twenty-third formula, is the activation function of the output layer; is the weight matrix from the hidden layer to the output layer during forward calculation, It is the weight matrix from the hidden layer to the output layer during reverse calculation;

[0097] 2) Output of the forward hidden layer The output of the hidden layer at the previous moment and the current input Decide jointly, as shown in Formula No. 24 below:

[0098] ; In the twenty-fourth formula, is the activation function of the hidden layer neurons, is the weight matrix of the hidden layer from the previous moment to the next moment during forward calculation, It is the weight matrix from the input layer to the hidden layer during forward calculation;

[0099] 3) Output of the reverse hidden layer of the Bi-LSTM algorithm The output of the hidden layer at the next moment and the current input Decide jointly as shown in Formula No. 25 below:

[0100] ; In the twenty-fifth formula, is the weight matrix of the hidden layer from the next moment to the previous moment during reverse calculation, It is the weight matrix from the input layer to the hidden layer during reverse calculation.

[0101] Furthermore, a distributed resource perception technology and cloud-edge-end collaborative identification mechanism with multi-feature matching and cloud-edge-end collaboration are established, including:

[0102] 1) First, the Bi-LSTM model is pre-trained in the cloud using an offline dataset;

[0103] 2) Secondly, the first model, the second model, and the third model are deployed to the user side, and new training samples are constructed using the total table data collected online on the side, and the parameters of the fully connected layer of the model are adjusted;

[0104] 3) Model parameter update process based on group weighted cross entropy, acting on the cross entropy loss of a single sample The expression is as follows:

[0105] ; In the expression, , used to indicate whether the current sample is a positive or negative label, , represents the probability of positive labels output by the Softmax layer of the model;

[0106] 4) Define group weighted cross entropy as the batch loss function when updating model parameters, as shown in the following formula 26:

[0107] ; In the twenty-sixth formula, Represents the number of samples in a batch. The subscripts P and N indicate that the current variable is only for positive samples and negative samples in the batch, respectively. and represents the sample number in a batch, and w represents the weight coefficient of a single sample;

[0108] 5) All the confident positive samples on the client are uploaded to the cloud and stored in the data set, and the pre-trained model on the cloud is updated regularly;

[0109] 6) Implement cloud-edge-end collaborative identification mechanism.

[0110] The present invention has the following beneficial effects:

[0111] The present invention provides an electric load identification method based on dual-state feature matching and cloud-edge-end collaboration. The method first detects the features of load events based on an optimized CUSUM composite sliding window accumulation and change point detection algorithm, constructs a subset of load event startup and shutdown event features, and establishes a load event and power-consuming equipment correlation model as a first model; then, in combination with a transient event operation characteristic labeling feature library, a node order optimization strategy of the Gini index is used to develop a load identification method of a directed acyclic graph support vector machine as a second model; secondly, in combination with a steady-state event operation characteristic labeling feature library, time features are used to refine and classify load events, and a load identification algorithm based on Mediod shift clustering and steady-state feature analysis is constructed as a third model; finally, the first model, the second model and the third model are integrated for training to construct a deep neural network load identification model based on a cloud-based deep learning Bi-LSTM algorithm, and a distributed resource perception technology and a cloud-edge-end collaborative identification mechanism of multi-feature matching and cloud-edge-end collaboration are established. The present invention helps to cope with complex changes in the power environment by adapting to the parallel use of multiple electrical appliances, and can effectively judge the start and stop processes of different electrical appliances, thereby achieving more accurate and efficient power load identification to meet the needs of smart grids and smart home systems in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0113] Figure 1 This is a flow chart of a method for identifying power load based on dual-state feature matching and cloud-edge collaboration of the present invention;

[0114] Figure 2 Schematic diagram of a detection process using a composite sliding window-based CUSUM algorithm in an embodiment of the present invention;

[0115] Figure 3 Schematic diagram of the cumulative sum of the sliding window of the average value in an embodiment of the present invention;

[0116] Figure 4 is a schematic diagram of the cumulative sum of the composite sliding window in an embodiment of the present invention;

[0117] Figure 5 is a schematic diagram of an original sliding window detection process in an embodiment of the present invention;

[0118] Figure 6 is a schematic diagram of a composite sliding window detection process in an embodiment of the present invention;

[0119] Figure 7 It is a framework diagram of an association model in an embodiment of the present invention;

[0120] Figure 8 The figure is a flow chart of a load identification algorithm based on a Bi-LSTM model in an embodiment of the present invention. DETAILED DESCRIPTION

[0121] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0122] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0123] In order to overcome the problem that it is difficult to effectively judge the start and stop process of different household appliances based on a single appliance identification feature quantity when there are many types of household appliances with different working principles, the present invention proposes a power load identification method based on dual-state feature matching and cloud-edge collaboration. Figure 1 As shown, the following steps are included:

[0124] S1, based on the optimized CUSUM composite sliding window accumulation and change point detection algorithm, the feature detection of load events is carried out, the feature subsets of load event startup and shutdown events are constructed, and the correlation model between load events and power-consuming equipment is established as the first model.

[0125] As a specific implementation method, an optimization algorithm model is performed for each feature detection of the electrical equipment. The specific steps of optimizing the CUSUM change point detection algorithm include:

[0126] 1) Examine random time series. Assume that a change point occurs at the kth moment in the random event sequence, and initialize the statistical function defined in the non-parametric bilateral CUSUM algorithm: , , where the statistical function and They respectively represent the positive and negative offsets of the current detection point after removing the influence of the noise value.

[0127] The first and second formulas are as follows:

[0128] ;

[0129] In the first formula, Indicates the positive CUSUM value at time 0, represents the positive CUSUM value at the kth moment, represents the positive CUSUM value at the k-1th moment, and max represents the maximum value function. It represents the positive offset of the detection point at the kth moment after removing the influence of the noise value; ; In the calculation formula of the positive offset, represents the data at the kth moment in a random time series, It represents the average value of the data before the kth moment in the random time series. It is usually considered is known, Represents the noise value, a random event sequence Changes smaller than the noise value will be ignored. k is a positive integer.

[0130] ;

[0131] In the second formula, Indicates the reverse CUSUM value at time 0, Represents the reverse CUSUM value at the kth moment, Represents the reverse CUSUM value at the k-1th moment, It represents the negative offset of the detection point at the kth moment after removing the influence of the noise value; .

[0132] The schematic diagram of the detection process using the CUSUM algorithm based on the composite sliding window is as follows: Figure 2 As shown, in Figure 2 middle, is the width of the mean window, is the width of the detection window. When the interval between two transient events exceeds + , the two events can be detected separately.

[0133] The weight calculation for each sampling point in the average value window includes:

[0134] First, Gaussian normal distribution can be used to assign weights. The weight calculation formula is as follows:

[0135] ; In the formula, represents the weight of sampling point i in the average window, represents the position of sampling point i within the average window, represents the position of the sampling point s in the middle of the detection window, e represents the natural constant, Represents a constant greater than 0.

[0136] Secondly, by adjusting the Gaussian window, a compromise is achieved between being too sensitive to changes in the previous and next data and being too dependent on the intermediate value. The calculation formula is as follows:

[0137] ; In the formula, Indicates the width of the average window; ; In the formula, The sample point where the averaging window begins.

[0138] Finally, the weight of the Gaussian normal distribution is normalized to obtain the weight, which is calculated as follows: ; In the formula, Represents the normalized weight of sampling point i within the average window.

[0139] 2) By analyzing the improved Sigmoid function, the accumulated weight ratio of the detection window start segment is reduced, and transient events are quickly identified, as shown in the third formula below:

[0140] ; In the third formula, represents the weight of sampling point i in the detection window, Indicates the middle value of the maximum width of the detection window, represents a constant greater than 0, Represents the position of sampling point i in the detection window, and its calculation formula is: , ; In the calculation formula, Indicates the width of the detection window, represents the width of the mean window, Indicates the position of the sampling point where the average window starts, i represents 0 to A positive integer between .

[0141] right Normalize as shown in the fourth formula below:

[0142] ; In the fourth formula, Represents the normalized weight of sampling point i in the detection window.

[0143] 3) By detecting the normalized weights of the sampling points in the window, the statistical function in the non-parametric CUSUM algorithm based on the improved Sigmoid function is determined. The formula is as follows:

[0144] ; .

[0145] So far, this embodiment has obtained the first model.

[0146] In order to verify the effect of the implementation method of the present invention, this embodiment puts the statistical function determination method into practical application, and selects modified real examples, including:

[0147] The actual load switching situation in this embodiment is as follows: The schematic diagram of the average sliding window cumulative sum is as follows Figure 3 As shown, the schematic diagram of the composite sliding window cumulative sum is as follows Figure 4 As shown by Figure 3 and Figure 4 It can be seen that at sampling point 105, an air conditioner was put into operation, at 139, an electric rice cooker was put into operation, at 231, an induction cooker was put into operation, and at 323, a kettle was put into operation, and then the electrical equipment put into operation was removed respectively. Among them, the switching behavior of the next load occurred after the previous equipment was put into operation and reached a steady state.

[0148] In order to avoid including multiple load events within the window size, the load event monitoring module average window and detection window width Both are set to 5. , To facilitate the description of the advantages of the algorithm in terms of cumulative effect, the width of the detection window is set to a fixed value, that is, the cumulative sum stops accumulating only when the sampling point reaches the window width.

[0149] When detecting a monitored object with a climbing event, The selection of the value needs to take into account both sensitivity and noise resistance, that is, at least Obviously, the selection range of reasonable threshold in the algorithm is wider. In order to facilitate the comparison of the sensitivity of the new and old algorithms, this embodiment uses the original sliding window H min and H max The threshold is selected based on the average value of , and the threshold H is calculated to be 400.

[0150] When a certain type of air conditioner load is put into use at sampling point 105, the original sliding window algorithm fails to reach the threshold value in the first detection window after the ramp event occurs, and the cumulative sum reaches the threshold value at the next detection window sampling point 113. The event occurrence time is reversed to sampling time 110, thus losing information at 5 sampling times. The schematic diagram of the original sliding window detection process is shown in the figure below. Figure 5 As shown, the schematic diagram of the composite sliding window detection process is as follows Figure 6 As shown by Figure 5 and Figure 6 It can be seen that the algorithm records the event occurrence time at sampling point 105, and completely retains the startup process information. It can be proved that the algorithm proposed by the present invention has a great advantage over the original algorithm in detecting climbing load events.

[0151] S2, combined with the labeling feature library of transient event operation characteristics, uses the node order optimization strategy of the Gini index to develop a load identification method based on a directed acyclic graph support vector machine as the second model.

[0152] In this embodiment, constructing a subset of load event startup and shutdown event features and establishing a load event and power equipment association model includes the following steps:

[0153] 1) In the definition of transient characteristics, the power within one cycle is its effective value, as shown in the fifth formula below:

[0154] ; In the fifth formula, Indicates the intensity of the current, N indicates the number of sampling points in one cycle, Indicates the current value at the nth sampling point in a cycle.

[0155] 2) According to the intensity of the current at each sampling point, the separation process of the transient process is set to obtain the transient waveform after separation, including:

[0156] The first step is to import the transient waveform extracted by the detection algorithm;

[0157] The second step is to calculate the current intensity before and after the transient process occurs. If the current before the transient process occurs is less than or equal to the current after the transient process occurs, the third step is entered, otherwise the fourth step is entered.

[0158] The third step is to subtract the waveform of the last steady-state current cycle before the transient state occurs from the transient waveform in the first step by cycle.

[0159] The fourth step is to subtract the waveform of the first steady-state current cycle after the transient state occurs from the transient waveform in the first step by cycle;

[0160] Step 5: Output the separated transient waveform.

[0161] 3) Steady-state feature extraction based on particle resampling, taking the feature at each moment as a particle, and recalculating the importance of each particle and resampling it according to the distribution of particles before and after the load event point, until the sum of the variances of all particles is minimized, and finally combining the transient features to realize the load event start-stop feature subset, that is, to obtain the load event start-up and shutdown event feature subsets. The specific process is as follows:

[0162] In the first step, assume that the load input change point is t, set k = 0, randomly obtain N particles after the current time t and before the next change point, and initialize each particle i so that it has an equal weight, where i is a positive integer, ;

[0163] The second step is importance sampling. For each particle i, the importance weight is estimated according to the degree of deviation from its center. , as shown in the twelfth formula below: ; In the twelfth formula, represents a particle set, Represents particles and a particle in the particle set The Euclidean distance between Normalize to get ;

[0164] The third step is resampling. According to the importance weight in the second step, particles with importance weights less than the threshold are discarded, and replacement sampling is performed based on the discarded particles.

[0165] The fourth step is to increase k by 1 and repeat the process of the second and third steps until the variance change in the particle group is maximized.

[0166] The fifth step is to perform differential analysis on the load characteristics in the currently obtained particle set and the previously recorded load characteristics to extract the load change characteristics.

[0167] 4) To find the best match between load input and load removal, a load event bipartite graph model is established, where is an undirected simple graph with vertices , edge , divide the vertex V into two disjoint subsets X and Y, satisfying , so that the two endpoints of each edge in graph G are located in X and Y respectively, is a bipartite graph, and its associated cost matrix is ​​expressed as: ; In the expression of the associated cost matrix, Represents a vertex and The association cost parameter between them, n represents the number of rows and columns of the matrix.

[0168] 5) Based on the Hungarian algorithm, for bipartite graphs Seeking the optimal match, the load matching mathematical model is shown in the sixth formula below:

[0169] ; In the sixth formula, min represents the minimum value function, Represents the cost parameter the sum of It represents the matching degree between different jump powers, n represents the number of load closing events, and m represents the number of load opening events;

[0170] Define the cost function The seventh formula is: ; In the seventh formula, represents the grey relational coefficient of the pth characteristic component, ,satisfy , , Indicates that m load start events are retrieved within a period of time t and n closing events ;

[0171] Defining the grey correlation coefficient The eighth formula is: ; In the eighth formula, represents the reference value of the characteristic component, Indicates The dimensionless attribute value of the feature component after normalization, Represents the resolution coefficient of grey relational degree;

[0172] The decision matrix is ​​established as shown in the ninth formula below:

[0173] ;

[0174] In the ninth formula, R represents the decision matrix, represents the reference value of the characteristic component, Indicates feature components.

[0175] 6) Improve the Hungarian matching algorithm and expand the characteristic component p at time t+Δt, as shown in the following tenth formula:

[0176] ; In the tenth formula, It means that k loads are turned on or off at the same time at time t. Represents a period of time, express The first derivative with respect to time, represents the characteristic component p at time t+Δt; ; In the formula, represents the value of the p characteristic components of the nth load, and k represents the number of loads;

[0177] The remaining part is equivalent to other possible load characteristic components, rewritten as: ; In the formula, represents the value of the p characteristic components of the nth load, Represents the values ​​of the p characteristic components of the mth load.

[0178] 7) Modify the original cost matrix and add the original zero elements , and replace the non-zero elements with The value of The calculation formula is as follows: ; In the formula, represents the normalized cost matrix, represents the maximum element in the cost matrix C, Represents the minimum element in the cost matrix C.

[0179] 8) Use the Hungarian algorithm to solve the new matrix , the expression of the new matrix is:

[0180] ;

[0181] 9) In order to establish the correlation between load events and power equipment, a transient feature template matching clustering algorithm based on the progress schedule is used for correlation modeling to obtain the correlation model between load events and power equipment. The entire correlation model framework is shown in the figure below. Figure 7 shown.

[0182] The definition of the posting progress is shown in the following eleventh formula: ;

[0183] In the eleventh formula, n represents the dimension of the vector, represents the difference between vector x and vector y, express, Represents the normalized vector x and vector y;

[0184] The calculation formula based on the progress of the amplitude is: ; In the calculation formula of the progress of posting, and Represents vectors With vector The maximum element value in represents the vector dimension, .

[0185] In this embodiment, the transient characteristics of the load are used to identify the load, and the basic principle of the node order optimization SVM based on the Gini index is as follows:

[0186] 1) Assume that the sample set is as shown in the following formula 13: ;

[0187] In the thirteenth formula, represents sample data, represents the sample category label, N represents the number of samples, and n represents the sample dimension. Represents n-dimensional real number space.

[0188] 2) Construct the optimal hyperplane to separate samples of different categories, and the optimization problem is shown in the following fourteenth formula:

[0189] ; , ;

[0190] In the fourteenth formula, W represents the normal vector of the hyperplane, b represents the offset of the hyperplane, and C represents the penalty factor. is the slack variable, represents the transpose of the hyperplane normal vector, Represents the optimization objective function The minimum value of , i represents the sample index, and m represents the number of samples.

[0191] 3) In order to optimize the solution, construct the Lagrangian function, as shown in the fifteenth formula below:

[0192] ;

[0193] In the fifteenth formula, and is the Lagrange multiplier.

[0194] 4) Solve the optimization problem and make Find the partial derivatives and get the optimal dual problem:

[0195] ; The constraints are: ;

[0196] By solving the optimization problem, the classification decision surface is obtained.

[0197] 5) Lagrange multipliers obtained based on the solution of the dual problem , the sixteenth formula for determining the classification decision surface is as follows:

[0198] ; In the sixteenth formula, The corresponding samples are called support vectors, and sign represents the positive and negative sign function.

[0199] 6) The seventeenth formula using particle swarm optimization is as follows:

[0200] ; In the seventeenth formula, represents the velocity of the particle in the dth dimension, represents the inertia weight, and represents the learning factor, and Represents a random number, represents the best historical position of the particle, represents the global optimal position, Indicates the current position of the particle. Represents the constraint factor.

[0201] 7) The sample set is divided into two subsets by the node and , the eighteenth formula of the Gini index of the sample after this division is as follows:

[0202] ; ; In the eighteenth formula, Represents a sample set The Gini index, Indicates The square of the class sample proportion, Represents a sample set Divided into and The Gini index after

[0203] The resulting nineteenth formula for the reduction in the Gini index is as follows: ; In the nineteenth formula, Indicates the degree of reduction of impurity before and after the sample set is divided.

[0204] 8) Select the node with the largest Gini index difference as the root node, and then and They are regarded as independent sets respectively, and the classifier with the largest Gini index difference is selected as the current node in the next layer, and so on to build a directed acyclic graph SVM structure.

[0205] So far, this embodiment has obtained the second model.

[0206] S3, combined with the labeling feature library of steady-state event operation characteristics, uses time characteristics to refine the classification of load events, and constructs a load identification algorithm based on Mediod shift clustering and steady-state feature analysis as the third model.

[0207] In this embodiment, the load identification model based on the Mediod Shift clustering algorithm and steady-state characteristic analysis is as follows:

[0208] 1) Given sampling points , the twentieth formula for the probability density estimate at point x is as follows:

[0209] ; In the 20th formula, represents the probability density estimate at the sampling point x, represents the Gaussian kernel function, h represents the fixed bandwidth, and d represents the dimension.

[0210] 2) The cluster center is updated by minimization, as shown in the following formula 21:

[0211] ; In the twenty-first formula, k represents the number of iterations, Indicates the current center point. Represents the next step center point obtained by iteration and is selected from the sample set The sampling points that satisfy the update conditions are: represents the weight function.

[0212] 3) In order to improve the clustering accuracy by using the steady-state time characteristics, the load operation time characteristics are segmented and a day is divided into b time periods. The calculation formula of the objective function is as follows: ; In the calculation formula of the objective function, Indicates The number of time periods occupied by a load, Indicates the load type number.

[0213] 4) Load operation time period is calculated through sparse binary matrix Indicates that, assuming Represents 8 load environments, and the matrix expression is as follows:

[0214] ; In the expression, Indicates the load status of the first load environment in the first time period, where 1 indicates that the load exists and 0 indicates that the load does not exist.

[0215] 5) If and Moment (where ) is the split time point, then The calculation formula is as follows:

[0216] ; In the formula, , .

[0217] 6) Based on Gaussian and density statistical methods, the one-dimensional characteristics of the operating time of all steady-state loads are analyzed, and load events are divided into two categories: regular and irregular.

[0218] 7) Use Mediod Shift clustering to refine the classification of load events, including:

[0219] The first step is to set the cluster bandwidth based on the load power characteristic statistical model. The threshold at which the mean value stops drifting ;

[0220] The second step is to calculate the power mean drift vector of each load event sample in the interval ;

[0221] The third step is to translate the position of each sample point. By calculating the drift vector, as shown in the following formula 22:

[0222] ; In the 22nd formula, represents the drift vector;

[0223] Step 4: Repeat steps 2 and 3 until Stop when.

[0224] So far, this embodiment has obtained the third model.

[0225] S4, integrates the first model, the second model and the third model for training to build a deep neural network load identification model based on the cloud-based deep learning Bi-LSTM algorithm, and establishes a distributed resource perception technology and cloud-edge-end collaborative identification mechanism with multi-feature matching and cloud-edge-end collaboration.

[0226] As a specific implementation method, a load identification algorithm based on the Bi-LSTM model is provided. The flow chart of the load identification algorithm based on the Bi-LSTM model is as follows: Figure 8 As shown, the following steps are included:

[0227] 1) When the Bi-LSTM algorithm is forward calculated, the output of the hidden layer The hidden layer output at the previous moment The influence of; when calculating in reverse, the output of the hidden layer The hidden layer output at the next moment The output layer is at time Output The output of the forward and backward hidden layers and Decide jointly, as shown in Formula 23 below:

[0228] ; In the twenty-third formula, is the activation function of the output layer; is the weight matrix from the hidden layer to the output layer during forward calculation, It is the weight matrix from the hidden layer to the output layer during reverse calculation.

[0229] 2) Output of the forward hidden layer The output of the hidden layer at the previous moment and the current input Decide jointly, as shown in Formula No. 24 below:

[0230] ; In the twenty-fourth formula, is the activation function of the hidden layer neurons, is the weight matrix of the hidden layer from the previous moment to the next moment during forward calculation, It is the weight matrix from the input layer to the hidden layer during forward calculation.

[0231] 3) Output of the reverse hidden layer of the Bi-LSTM algorithm The output of the hidden layer at the next moment and the current input Decide jointly as shown in Formula No. 25 below:

[0232] ; In the twenty-fifth formula, is the weight matrix of the hidden layer from the next moment to the previous moment during reverse calculation, It is the weight matrix from the input layer to the hidden layer during reverse calculation.

[0233] In this embodiment, a distributed resource perception technology and a cloud-edge-end collaborative identification mechanism for multi-feature matching and cloud-edge-end collaboration are established, including:

[0234] 1) First, the Bi-LSTM model is pre-trained in the cloud using an offline dataset.

[0235] 2) Secondly, the first model, the second model and the third model are deployed to the user side, and new training samples are constructed using the total table data collected online on the side, and the parameters of the fully connected layer of the model are adjusted.

[0236] 3) Model parameter update process based on group weighted cross entropy, acting on the cross entropy loss of a single sample The expression is as follows:

[0237] ; In the expression, , used to indicate whether the current sample is a positive or negative label, , represents the probability of positive labels output by the Softmax layer of the model.

[0238] 4) Define group weighted cross entropy as the batch loss function when updating model parameters, as shown in the following formula 26:

[0239] ; In the twenty-sixth formula, Represents the number of samples in a batch. The subscripts P and N indicate that the current variable is only for positive samples and negative samples in the batch, respectively. and Represents the sample number in a batch, and w represents the weight coefficient of a single sample, which is used to adjust the contribution of positive and negative samples to the total loss.

[0240] 5) All the confident positive samples on the client are uploaded to the cloud and stored in the dataset, and the pre-trained model on the cloud is updated regularly.

[0241] 6) Implement cloud-edge-end collaborative identification mechanism.

[0242] Thus, this embodiment provides a model for accurately identifying loads of various electrical equipment, which can effectively improve the accuracy and robustness of online load identification.

[0243] In order to verify the effect of the method of the present invention and conduct practical application, a modified real example is selected, specifically including:

[0244] The characteristic data of some typical electrical appliances were collected and extracted from a household user. The data included 10 kinds of electrical appliances with a total of 13 operating states (electric heating 1, 2 and 3, electric stew pot, foot bath bucket, hair dryer cold air, hot air, induction cooker, pressure cooker, printer, rice cooker, soybean milk machine, TV), among which electric heating is a multi-state appliance and TV is a low-power appliance. Considering that one state corresponds to one output, the neural network has 13 outputs. The output of the neural network model is shown in Table 1.

[0245] Table 1

[0246]

[0247] In this embodiment, the output of each output layer neuron is a probability value, which indicates the probability that the input data belongs to a certain state, and its value range is between 0 and 1, where 0 indicates that the input data cannot belong to the state, and 1 indicates that the input data must belong to the state. In an output vector, the state corresponding to the output with the highest probability is regarded as the identification result of the input data.

[0248] Using precision (P), recall (R), accuracy (A) and F1 value as the performance evaluation indicators of Bi-LSTM neural network to compare the algorithm performance of neural network, it can be calculated as follows: Indicates the number of times a certain state exists and is detected at this time; Indicates the number of times a state does not exist at this time but the result indicates that it exists; Indicates the number of times a state does not exist at this time and the result indicates that it does not exist; Indicates the number of times a state exists but is not detected:

[0249] ; ; ; ;

[0250] Since the current harmonic components of the electrical equipment are also considered as characteristic components, the current harmonic components can be obtained by fast Fourier transforming the current. Therefore, 32 parameters, including active power, reactive power and 15th odd and even harmonics of the electrical equipment, are selected as the input variables of the neural network, that is, the input dimension of the neural network is 32. The neural network is trained and tested three times, and the accuracy, recall rate and F1 value of each state in the test results are calculated. The output results of the neural network model are shown in Table 2:

[0251] Table 2

[0252]

[0253] Table 2

[0254]

[0255] In Table 2, state 1 represents electric heating level 1, state 2 represents electric heating level 2, state 3 represents electric heating level 3, state 4 represents electric stew pot, state 5 represents foot bath bucket, state 6 represents hair dryer cold air level, state 7 represents hair dryer hot air level, state 8 represents induction cooker, state 9 represents pressure cooker, state 10 represents printer, state 11 represents rice cooker, state 12 represents soybean milk maker, and state 13 represents television.

[0256] The average precision, average recall and average F1 of the neural network are calculated as shown in Table 3:

[0257] Table 3

[0258]

[0259] As can be seen from Table 3, in this embodiment, Bi-LSTM is superior in terms of the four evaluation indicators of recognition accuracy, precision, recall rate and F1 value, that is, Bi-LSTM shows better performance in load identification.

[0260] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for identifying power load based on dual-state feature matching and cloud-edge collaboration, characterized in that: The following steps are involved: S1, based on the optimized CUSUM composite sliding window accumulation and change point detection algorithm, the feature detection of load events is carried out, the feature subsets of load event startup and shutdown events are constructed, and the correlation model between load events and power-consuming equipment is established as the first model; S2, combined with the labeling feature library of transient event operation characteristics, using the node order optimization strategy of the Gini index, develops a load identification method based on a directed acyclic graph support vector machine as the second model; S3, combined with the labeling feature library of steady-state event operation characteristics, uses time characteristics to refine the classification of load events, and constructs a load identification algorithm based on Mediod shift clustering and steady-state feature analysis as the third model; S4, integrates the first model, the second model and the third model for training to build a deep neural network load identification model based on the cloud-based deep learning Bi-LSTM algorithm, and establishes a distributed resource perception technology and cloud-edge-end collaborative identification mechanism with multi-feature matching and cloud-edge-end collaboration.

2. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: The statistical function in the optimized CUSUM composite sliding window accumulation and change point detection algorithm in step S1 is: 1) Assume that a change point occurs at the kth moment in the random time series, and initialize the statistical function in the non-parametric bilateral CUSUM algorithm, as shown in the first and second formulas below: ； In the first formula, and represent the 0th moment, the kth moment, and the The positive CUSUM value at the moment, max means taking the maximum value function, It represents the positive offset of the detection point at the kth moment after removing the influence of the noise value; ; In the calculation formula of the positive offset, represents the data at the kth moment in a random time series, represents the average value of the data before the kth moment in the random time series. represents the noise value, k is a positive integer; ； In the second formula, and represent the 0th moment, the kth moment, and the The reverse CUSUM value at the moment, It represents the negative offset of the detection point at the kth moment after removing the influence of the noise value; ; 2) Reduce the accumulated weight ratio of the detection window start segment, as shown in the third formula below: ; In the third formula, represents the weight of sampling point i in the detection window, Indicates the middle value of the maximum width of the detection window, represents a constant greater than 0, Represents the position of sampling point i in the detection window, and its calculation formula is: , ; In the calculation formula, Indicates the width of the detection window, represents the width of the average window, Indicates the position of the sampling point where the average window starts, i represents 0 to A positive integer between ; right Normalize as shown in the fourth formula below: ; In the fourth formula, Represents the normalized weight of sampling point i in the detection window; 3) By detecting the normalized weights of the sampling points in the window, the statistical function in the non-parametric CUSUM algorithm based on the improved Sigmoid function is determined. The formula is as follows: ； 。 3. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: The step of constructing a subset of load event startup and shutdown event characteristics and establishing a load event and power-consuming equipment association model includes: 1) In the definition of transient characteristics, the power within one cycle is its effective value, as shown in the fifth formula below: ; In the fifth formula, Indicates the intensity of the current, N indicates the number of sampling points in one cycle, Indicates the current value of the nth sampling point in a cycle; 2) According to the intensity of the current at each sampling point, the separation process of the transient process is set to obtain the separated transient waveform, and then the number of cycles of the current transient intensity change greater than the transient start judgment threshold is sampled and recorded, and the number of cycles of the current transient intensity change less than the transient end judgment threshold is sampled and recorded; 3) Steady-state feature extraction based on particle resampling, taking the feature at each moment as a particle, and recalculating the importance of each particle and resampling it according to the distribution of particles before and after the load event point, until the sum of the variances of all particles is minimized, and finally combining the transient features to obtain the feature subsets of the load event startup and shutdown events; 4) Establish a load event bipartite graph model, where is an undirected simple graph with vertices , edge , divide the vertex V into two disjoint subsets X and Y, satisfying , so that the two endpoints of each edge in graph G are located in X and Y respectively, is a bipartite graph, and its associated cost matrix is ​​expressed as: ; In the expression of the associated cost matrix, Represents a vertex and The association cost parameter between them, n represents the number of rows and columns of the matrix; 5) Based on the Hungarian algorithm, for bipartite graphs Seeking the optimal match, the load matching mathematical model is shown in the sixth formula below: ; In the sixth formula, min represents the minimum value function, Represents the cost parameter the sum of It represents the matching degree between different jump powers, n represents the number of load closing events, and m represents the number of load opening events; Define the cost function The seventh formula is: ; In the seventh formula, represents the grey relational coefficient of the pth characteristic component, ,satisfy , , Indicates that m load start events are retrieved within a period of time t and n closing events ; Defining the grey correlation coefficient The eighth formula is: ; In the eighth formula, represents the reference value of the jth eigencomponent, Indicates The dimensionless attribute value of the feature component after normalization, Represents the resolution coefficient of grey relational degree; The decision matrix is ​​established as shown in the ninth formula below: ； In the ninth formula, R represents the decision matrix, represents the reference value of the characteristic component, Indicates feature components; 6) Improve the Hungarian matching algorithm and expand the characteristic component p at time t+Δt, as shown in the following tenth formula: ; In the tenth formula, It means that k loads are turned on or off at the same time at time t. Represents a period of time, express The first derivative with respect to time, represents the characteristic component p at time t+Δt; ; In the formula, represents the value of the p characteristic components of the nth load, and k represents the number of loads; The remaining part is equivalent to other possible load characteristic components, rewritten as: ; In the formula, represents the value of the p characteristic components of the nth load, Represents the value of the p characteristic components of the mth load; 7) Modify the original cost matrix and add the original zero elements , and replace the non-zero elements with The value of The calculation formula is as follows: ; In the formula, represents the normalized cost matrix, represents the maximum element in the cost matrix C, Represents the minimum element in the cost matrix C; 8) Use the Hungarian algorithm to solve the new matrix , the expression of the new matrix is: ； 9) Adopt the transient feature template matching clustering algorithm based on proximity to carry out association modeling and obtain the correlation model between load events and power equipment; The definition of the posting progress is shown in the following eleventh formula: ; In the eleventh formula, n represents the dimension of the vector, represents the difference between vector x and vector y, and Represents the normalized vector x and vector y; The calculation formula based on the progress of the amplitude is: ; In the calculation formula of the progress of posting, and Represents vectors With vector The maximum element value in represents the vector dimension, .

4. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 3 is characterized in that: The separation implementation process of the transient process is set according to the intensity of the current at each sampling point to obtain the separated transient waveform, including: The first step is to import the transient waveform extracted by the detection algorithm; The second step is to calculate the current intensity before and after the transient process occurs. If the current before the transient process occurs is less than or equal to the current after the transient process occurs, the third step is entered, otherwise the fourth step is entered. The third step is to subtract the waveform of the last steady-state current cycle before the transient state occurs from the transient waveform in the first step by cycle. The fourth step is to subtract the waveform of the first steady-state current cycle after the transient state occurs from the transient waveform in the first step by cycle; Step 5: Output the separated transient waveform.

5. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 3 is characterized in that: The steady-state feature extraction based on particle resampling takes the feature at each moment as a particle, and recalculates the importance of each particle according to the distribution of particles before and after the load event point and resamples until the sum of the variances of all particles is minimized, including: In the first step, assume that the load input change point is t, set k = 0, randomly obtain N particles after the current time t and before the next change point, and initialize each particle i so that it has an equal weight, where i is a positive integer; The second step is importance sampling. For each particle i, the importance weight is estimated according to the degree of deviation from its center. , as shown in the twelfth formula below: ; In the twelfth formula, represents a particle set, Represents particles and a particle in the particle set The Euclidean distance between Normalize to get ; The third step is resampling. According to the importance weight in the second step, particles with importance weights less than the threshold are discarded, and replacement sampling is performed based on the discarded particles. The fourth step is to increase k by 1 and repeat the process of the second and third steps until the variance change in the particle group is maximized. The fifth step is to perform differential analysis between the load characteristics in the currently obtained particle set and the previously recorded load characteristics to extract the load change characteristics.

6. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: The basic principles of the load identification method of directed acyclic graph support vector machine are developed by using the node order optimization strategy of Gini index, including: 1) Assume that the sample set is as shown in the following formula 13: ; In the thirteenth formula, represents sample data, represents the sample category label, N represents the number of samples, and n represents the sample dimension. represents n-dimensional real number space; 2) Construct the optimal hyperplane to separate samples of different categories, and the optimization problem is shown in the following fourteenth formula: ； ， ； In the fourteenth formula, W represents the normal vector of the hyperplane, b represents the offset of the hyperplane, and C represents the penalty factor. is the slack variable, represents the transpose of the hyperplane normal vector, Represents the optimization objective function The minimum value of , i represents the sample index, and m represents the number of samples; 3) Construct the Lagrangian function, as shown in the fifteenth formula below: ； In the fifteenth formula, and is the Lagrange multiplier; 4) Solve the optimization problem and make Find the partial derivatives and get the optimal dual problem: ; The constraints are: ; By solving the optimization problem, the classification decision surface is obtained; 5) Lagrange multipliers obtained based on the solution of the dual problem , the sixteenth formula for determining the classification decision surface is as follows: ; In the sixteenth formula, The corresponding samples are called support vectors, and sign represents the positive and negative sign function; 6) The seventeenth formula using particle swarm optimization is as follows: ; In the seventeenth formula, represents the velocity of the particle in the dth dimension, represents the inertia weight, and represents the learning factor, and Represents a random number, represents the best historical position of the particle, represents the global optimal position, Indicates the current position of the particle. represents the constraint factor; 7) The sample set is divided into two subsets by the node and , the eighteenth formula of the Gini index of the sample after this division is as follows: ; ; In the eighteenth formula, Represents a sample set The Gini index, Indicates The square of the class sample proportion, Represents a sample set Divided into and The Gini index after The resulting nineteenth formula for the reduction in the Gini index is as follows: ; In the nineteenth formula, Indicates the degree of reduction of impurity before and after the sample set is divided; 8) Select the node with the largest Gini index difference as the root node, then regard D1 and D2 as independent sets respectively, select the classifier with the largest Gini index difference as the current node in the next layer, and so on, to build a directed acyclic graph SVM structure.

7. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: Load identification algorithm based on Mediod shift clustering and steady-state characteristic analysis, including: 1) Given a sampling point , the twentieth formula for the probability density estimate at point x is as follows: ; In the 20th formula, represents the probability density estimate at the sampling point x, represents the Gaussian kernel function, h represents the fixed bandwidth, and d represents the dimension; 2) The cluster center is updated by minimization, as shown in the following formula 21: ; In the twenty-first formula, k represents the number of iterations, Indicates the current center point. Represents the next step center point obtained by iteration and is selected from the sample set The sampling points that satisfy the update conditions are: represents the weight function; 3) Segment the load operation time characteristics and divide a day into b time periods. The calculation formula of the objective function is as follows: ; In the calculation formula of the objective function, Indicates The number of time periods occupied by a load, Indicates the number of load types; 4) Load operation period through sparse binary matrix Indicates that, assuming Represents 8 load environments, and the matrix expression is as follows: ; In the expression, Indicates the load status of the first load environment in the first time period, where 1 indicates that the load exists and 0 indicates that the load does not exist; 5) If and Moment (where ) is the split time point, then The calculation formula is as follows: ; In the formula, , ; 6) Based on Gaussian and density statistical methods, the one-dimensional characteristics of the running time of all steady-state loads are analyzed, and load events are divided into two categories: regular and irregular. 7) Use Mediod Shift clustering to refine the classification of load events.

8. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 7 is characterized in that: The use of Mediod Shift clustering to refine the classification of load events includes: The first step is to set the cluster bandwidth based on the load power characteristic statistical model. The threshold at which the mean value stops drifting ; The second step is to calculate the power mean drift vector of each load event sample in the interval ; The third step is to translate the position of each sample point. By calculating the drift vector, as shown in the following formula 22: ; In the 22nd formula, represents the drift vector; Step 4: Repeat steps 2 and 3 until Stop when.

9. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: The deep neural network load identification model based on the Bi-LSTM algorithm includes: 1) When the Bi-LSTM algorithm is forward calculated, the output of the hidden layer The hidden layer output at the previous moment The impact of; when calculating in reverse, the output of the hidden layer The hidden layer output at the next moment The output layer is at time Output The output of the forward and backward hidden layers and Decide jointly, as shown in Formula 23 below: ; In the twenty-third formula, is the activation function of the output layer; is the weight matrix from the hidden layer to the output layer during forward calculation, It is the weight matrix from the hidden layer to the output layer during reverse calculation; 2) Output of the forward hidden layer The output of the hidden layer at the previous moment and the current input Decide jointly, as shown in Formula No. 24 below: ; In the twenty-fourth formula, is the activation function of the hidden layer neurons, is the weight matrix of the hidden layer from the previous moment to the next moment during forward calculation, It is the weight matrix from the input layer to the hidden layer during forward calculation; 3) Output of the reverse hidden layer of the Bi-LSTM algorithm The output of the hidden layer at the next moment and the current input Decide jointly as shown in Formula No. 25 below: ; In the twenty-fifth formula, is the weight matrix of the hidden layer from the next moment to the previous moment during reverse calculation, It is the weight matrix from the input layer to the hidden layer during reverse calculation.

10. The power load identification method based on dual-state feature matching and cloud-edge collaboration according to claim 1 is characterized in that: Establish distributed resource perception technology and cloud-edge-end collaborative identification mechanism with multi-feature matching, including: 1) First, the Bi-LSTM model is pre-trained in the cloud using an offline dataset; 2) Secondly, the first model, the second model, and the third model are deployed to the user side, and new training samples are constructed using the total table data collected online on the side, and the parameters of the fully connected layer of the model are adjusted; 3) Model parameter update process based on group weighted cross entropy, acting on the cross entropy loss of a single sample The expression is as follows: ; In the expression, , used to indicate whether the current sample is a positive or negative label, , represents the probability of positive labels output by the Softmax layer of the model; 4) Define group weighted cross entropy as the batch loss function when updating model parameters, as shown in the following formula 26: ; In the twenty-sixth formula, Represents the number of samples in a batch. The subscripts P and N indicate that the current variable is only for positive samples and negative samples in the batch, respectively. and represents the sample number in a batch, and w represents the weight coefficient of a single sample; 5) All the confident positive samples on the client are uploaded to the cloud and stored in the data set, and the pre-trained model on the cloud is updated regularly; 6) Realize cloud-edge-device collaborative identification mechanism.

Images