A multi-label load identification method of a multi-group mobile atlas fast clipping network

Abstract

The application provides a multi-label load identification method of a multi-group mobile atlas fast clipping network, which comprises the following steps: firstly, obtaining load information of a to-be-detected user, i.e., load amounts at different time points in a whole day; secondly, drawing a load curve atlas of the user, and then performing multi-group mobile fast clipping and convolution operation on the atlas; establishing a training set and training a Nasnet-deep network; outputting a user load state prediction value through the Nasnet-deep network; setting a threshold value, comparing the threshold value with the load state prediction value, and then outputting a running state of the load, so as to realize a multi-label identification function; and the multi-group mobile atlas fast clipping network method can solve the problem of low efficiency of user-side load supervision, realize a load type identification function, optimize a user-side supervision system, and improve the efficiency of bidirectional interaction of electric energy demand.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power systems and artificial intelligence, and relates to a multi-group mobile graph fast clipping network method, which is suitable for identification of multi-label loads of power systems. BACKGROUND

[0002] The power grid has always paid more attention to user side management, but the current user side power consumption behavior monitoring technology cannot understand the user side power consumption mode to give the corresponding energy saving optimization management mode. The improvement of power grid energy efficiency urgently needs a technology to obtain the load information of the user side to realize green two-way interaction. The existing multi-label load identification method, such as extracting user side current high harmonic additional features and power features, realizes the type identification of the load based on the mixed feature matrix, and the disadvantage is that the current high harmonic component has the characteristics of strong interference, and has a great requirement on the anti-interference ability of the equipment, which brings great difficulty to the large-scale investment and use of the monitoring equipment.

[0003] Generally, the feature graph identification load type method is adopted, the user side voltage and power data are collected and the feature graph is drawn, and the type of the load is output through the fuzzy artificial neural network. Since a large number of blank areas and non-key areas are contained in the feature graph, the important information of the graph cannot be extracted by the traditional intelligent algorithm, the training time of the neural network is prolonged, the multi-label load identification efficiency is greatly reduced, and the useless information which is not removed also causes the reduction of the identification accuracy.

[0004] Therefore, a multi-label load identification method of multi-group mobile graph fast clipping network is proposed to solve the problem of low efficiency of user side load supervision. SUMMARY

[0005] The application proposes a multi-label load identification method of multi-group mobile graph fast clipping network, which combines knowledge graph and neural network prediction for multi-label load identification; the steps in the use process are:

[0006] Step (1): obtaining the load information of the user to be tested, i.e. the load at different time points in a whole day;

[0007] Obtain the load information S of the user, which is composed of the power consumption at different time points in a whole day;

[0008] S=[S1,S2,S3,…,S t ,…,S 1440 ] t=1,2,3,…,1440; (1)

[0009] Among them, S1 is the power consumption of the user at the first time point, S2 is the power consumption of the user at the second time point, S3 is the power consumption of the user at the third time point, and S tS 1440 is the power consumption of the user at the 1440th moment;

[0010] Q i is the start-stop state quantity of the i-th type of load representing the running state of the load of the user;

[0011]

[0012] wherein i = 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15;

[0013] A load state vector of the user is introduced, and the load state vector Q is composed of 15 start-stop state quantities Q i representing the running states of different types of loads:

[0014] Q = [Q1, Q2, Q3, …, Q i , Q 15 ] (3)

[0015] wherein i = 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, Q1 is the start-stop state quantity of the first type of load, Q2 is the start-stop state quantity of the second type of load, Q3 is the start-stop state quantity of the third type of load, Q i is the start-stop state quantity of the i-th type of load, and Q 15 is the start-stop state quantity of the 15th type of load;

[0016] Step (2): draw a load curve atlas of the user, and then perform convolution operation on the atlas after multiple groups of moving and quick cutting:

[0017] take the moment t as the horizontal coordinate, and take the load quantity S t as the vertical coordinate, to generate a load curve atlas G with a pixel size of 1300x409; move the load curve atlas G in eight directions to obtain eight same atlases G 1 , G 2 , G 3 , G 4 , G 5 , G 6 , G 7 , and G 8 :

[0018]

[0019] wherein G 1 is the atlas obtained by moving G to the left-up direction, G 2 is the atlas obtained by moving G to the up direction, G 3The atlas obtained by moving G to the upper right direction, G 4 The atlas obtained by moving G to the left direction, G 5 The atlas obtained by moving G to the right direction, G 6 The atlas obtained by moving G to the lower left direction, G 7 The atlas obtained by moving G to the lower direction, G 8 The atlas obtained by moving G to the lower right direction;

[0020] The load curve atlas G and the atlas G 1 ,G 2 ,G 3 ,G 4 ,G 5 ,G 6 ,G 7 and G 8 are arranged and spliced to form a new atlas G', and the pixel size of G' is 3900x1227:

[0021]

[0022] The atlas G' is convolved using a convolution kernel, and the convolution kernel is:

[0023]

[0024] The convolution kernel is moved from left to right and from top to bottom on the atlas G', and the convolution result is the average value of the pixel points in the area covered by the convolution kernel on the atlas G'; after all the pixel points on the picture G' are covered by the convolution kernel, the convolved atlas G'' is obtained;

[0025] The rows and columns of pixel points in the convolved atlas are all zero or all 255, and are cropped and removed; the picture G S only contains valuable load curve information and removes meaningless blank information;

[0026] Step (3): Establish a training set and train the Nasnet-deep network:

[0027] Randomly collect load information S of N users, and calculate the load state vector of the N users according to the load usage, and then establish a training set; the training set is composed of N load information S pictures G S obtained by step (2) and the load state vectors of the N users; the atlas G S is input into the Nasnet-deep network, and the start-stop state quantity Q i of the i-th type of load is output from the Nasnet-deep network, and the i-th Nasnet-deep network model is trained to predict the start-stop state quantity of the i-th type of load;

[0028] Step (4): Output the prediction value of the user load state through the Nasnet-deep network The start-stop state of the i-th load is represented as:

[0029] The Nasnet-deep network model is composed of a Nasnet layer, three fully connected layers, and a regression layer: the Nasnet layer is stacked by five Cells; one Cell is composed of five blocks; the (j-1)-th feature map h j-1 and the j-th feature map h j are respectively taken as the input of the hidden layer A and the hidden layer B in the j-th Cell, j = 1, 2, 3, 4, 5; then, the feature maps of the hidden layer A and the hidden layer B are respectively subjected to a 3x3 convolution and a 5x5 max-pooling operation, and the results of the two operations are fused in a summation manner; finally, all the feature maps generated by the five blocks are spliced to form a complete feature map; the output feature map of the Nasnet layer is G N ; the output feature map of the Nasnet layer is G N After passing through the three fully connected layers, the prediction value of the start-stop state of the i-th load is finally output by the regression layer

[0030] Step (5): Set a threshold value and compare it with the prediction value of the load state to output the running state of the load:

[0031] The start-stop state threshold value of the i-th load is calculated according to the load state vectors of N users collected in step (3)

[0032]

[0033] wherein Q i,n is the start-stop state quantity of the i-th load of the n-th user, n = 1, 2, 3, …, N;

[0034] The prediction value of the start-stop state of the i-th load is compared with the start-stop state threshold value When is greater than or equal to the start-stop state threshold value , it indicates that the i-th load is in a start state; when is less than the start-stop state threshold value , it indicates that the i-th load is in a stop state; finally, the running states of the 15 load types are output. ​

[0035] The present application has the following advantages and effects relative to the prior art:

[0036] (1) The present application proposes a method for fast cutting of multiple groups of moving graphs, which moves, splices and cuts the original graph containing the load curve, thereby eliminating a large amount of blank information and useless information in the original graph and increasing the proportion of important load data information, thereby reducing the training time while improving the accuracy;

[0037] (2) The present application proposes a Nasnet-deep network model, which replaces the "full connection layer, softmax layer and classify layer" in the Nasnet network structure with "full connection layer, full connection layer, full connection layer and regression layer", adds a regression layer so that the output of the Nasnet-deep network is a specific start-stop state prediction value, and the stacking of three full connection layers can improve the depth and accuracy of the network model;

[0038] (3) The present application sets a start-stop state threshold based on the prediction value of the network model, and determines the type of load by comparing the prediction value and the start-stop state threshold, which converts the traditional multi-label recognition problem into a quantitative prediction problem, can increase the output types of multi-label recognition, and can improve the recognition accuracy of multi-label recognition. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is the network flow chart of the fast cutting of multiple groups of moving graphs of the method of the present application.

[0040] Figure 2 is the Cell structure diagram of the Nasnet-deep network of the method of the present application. DETAILED DESCRIPTION

[0041] The multi-label load recognition method of the fast cutting network of multiple groups of moving graphs proposed by the present application is described in detail as follows in combination with the drawings:

[0042] Figure 1 is the network flow chart of the fast cutting of multiple groups of moving graphs of the method of the present application.

[0043] The specific steps of the fast cutting network method of multiple groups of moving graphs of the method of the present application are as follows:

[0044] Step (1): Obtain the load information of the user to be tested, i.e. the load at different times of the day;

[0045] Step (2): Draw the load curve graph of the user, and then perform multiple group moving fast cutting and convolution operation on the graph;

[0046] Step (3): Establish a training set and train the Nasnet-deep network;

[0047] Step (4): output the user load state prediction value through the Nasnet-deep network;

[0048] Step (5): set a threshold and compare it with the load state prediction value to output the running state of the load.

[0049] Figure 2 is the Cell structure diagram of the Nasnet-deep network of the method of the present application. One Cell is composed of 5 blocks; the (j-1)th feature map h j-1 and the jth feature map h j respectively serve as the input of the hidden layer A and the hidden layer B in the jth Cell, j=1, 2, 3, 4, 5; then, the feature map of the hidden layer A and the hidden layer B is respectively subjected to a 3x3 convolution and a 5x5 maximum pooling operation, and the results of the two operations are fused in a summation manner; finally, all the feature maps generated by the 5 blocks are spliced to form a complete feature map.

[0050] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent flow transformation using the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1.A multi-label load identification method of a multi-group mobile atlas fast clipping network, characterized in that, The knowledge graph is combined with neural network prediction for multi-label load identification; the steps in the use process are: Step (1): obtaining the load information of the user to be measured, i.e. the load at different time points throughout the day; Obtaining the load information S of the user, which is composed of the power consumption at different time points throughout the day; S = [S1, S2, S3,..., S t ,…,S 1440 ] t = 1, 2, 3,..., 1440; (1) S1 is the power consumption of the user at the first time, S2 is the power consumption of the user at the second time, S3 is the power consumption of the user at the third time, S t S is the power consumption of the user at the t time, S 1440 S is the power consumption of the user at the 1440 time. To characterize the load operating state of the user, the start-stop state quantity Q of the i-th type of load is introduced i : Wherein, i = 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15; The load state vector of the user is introduced, and the load state vector Q is composed of 15 start-stop state quantities Q representing the running states of different types of loads i Composition: Q = [Q1, Q2, Q3,..., Q i ,…,Q 15 ] (3) wherein i = 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, Q1 is the start-stop state quantity of the first load type, Q2 is the start-stop state quantity of the second load type, Q3 is the start-stop state quantity of the third load type, Q i is the start-stop state quantity of the i-th load type, Q 15 is the start-stop state quantity of the 15-th load type; Step (2): drawing the load curve atlas of the user, and then performing multi-group moving quick cutting and convolution operation on the atlas: With time t as the horizontal coordinate, and load amount S as the vertical coordinate, load information S is generated into a load curve graph G, with a pixel size of 1300x409; the load curve graph G is moved in eight directions to obtain eight identical graphs G t , 1 , 2 , 3 , 4 , 5 , 6 , 7 , and G 8 : wherein G 1 is a graph obtained by moving G to the upper left direction, G 2 is a graph obtained by moving G to the upper direction, G 3 is a graph obtained by moving G to the upper right direction, G 4 is a graph obtained by moving G to the left direction, G 5 is a graph obtained by moving G to the right direction, G 6 is a graph obtained by moving G to the lower left direction, G 7 is a graph obtained by moving G to the lower direction, G 8 is a graph obtained by moving G to the lower right direction; The load curve atlas G and the atlas G 1 ,G 2 ,G 3 ,G 4 ,G 5 ,G 6 ,G 7 and G 8 are arranged and spliced to form a new atlas G', and the pixel size of G' is 3900x1227: Convolution kernel is used to perform convolution processing on the atlas G', and the convolution kernel is: The convolution kernel is moved and covered on the atlas G' from left to right and from top to bottom, and the convolution result is the average value of the pixel points on the area covered by the convolution kernel on the atlas G'; after all the pixel points on the picture G' are covered by the convolution kernel, the convolution atlas G'' is obtained; The rows and columns of the pixel points in the convolution atlas are all zero or all 255, and the picture G after the cutting and removing treatment S Only valuable load curve information is included, and meaningless blank information is removed. Step (3): establishing a training set and training the Nasnet-deep network: Randomly collect load information S of N users, and calculate the load state vector of the N users according to the load use condition, and then establish a training set; the training set is composed of N load information S and the picture G obtained through step (2) S and the load state vector of the N users; the atlas G S As the input of the Nasnet-deep network, the start-stop state quantity Q of the i-th type of load i As the output of the Nasnet-deep network, the i-th Nasnet-deep network model is trained for predicting the start-stop state quantity of the i-th type of load; Step (4): outputting a user load state prediction value through a Nasnet-deep network On-off state representing the i-th type of load: The Nasnet-deep network model is composed of a Nasnet layer, three full connection layers and a regression layer: the Nasnet layer is stacked by five Cells; one Cell is composed of five blocks; the (j-1)-th feature map h j-1 and the j-th feature map h j are respectively taken as the input of the hidden layer A and the hidden layer B in the j-th Cell, j = 1, 2, 3, 4, 5; then, the feature maps of the hidden layer A and the hidden layer B are respectively subjected to a 3x3 convolution and a 5x5 max-pooling operation, and the results of the two operations are fused in a summation manner; finally, all the feature maps generated by the five blocks are spliced to form a complete feature map; the output feature map G N of the Nasnet layer is G N After passing through the three full connection layers, the start-stop state prediction value of the i-th load is finally output by the regression layer Step (5): setting a threshold value and comparing it with the load state prediction value to output the running state of the load: Calculating the start-stop state threshold of the i-th type of load according to the load state vector of N users collected in step (3) wherein Q i,n is the start-stop state quantity of the nth user's ith load, n = 1, 2, 3,..., N; The start-stop state prediction value of the ith type of load is compared with a start-stop state threshold value When the start-stop state prediction value is greater than or equal to the start-stop state threshold value it indicates that the ith type of load is in a start state, and when the start-stop state prediction value is less than the start-stop state threshold value it indicates that the ith type of load is in a stop state; finally, the running states of the 15 types of loads are output.​​​

Images