CN-121614844-B - Non-invasive electricity load identification method based on neural network model

CN121614844BCN 121614844 BCN121614844 BCN 121614844BCN-121614844-B

Abstract

The invention relates to the technical field of load identification, in particular to a non-invasive electricity load identification method based on a neural network model. The method comprises the steps of obtaining current signals in different detection periods of an electric appliance of a user, inputting the current signals into a neural network to obtain a feature map, obtaining parameter vectors of each feature channel in each detection period based on gray distribution of different feature channels in the feature map in each detection period, obtaining relative correlation coefficients according to distribution differences of the parameter vectors of the same feature channel in each detection period and the adjacent detection period, evaluating channel information redundancy expression of each feature channel according to integral distribution differences and parameter vectors of the relative correlation coefficients of load detection data of each detection period in a single feature channel, introducing an ECA mechanism, determining feature importance weights of the feature channels, and judging load running states. The invention improves the accuracy of load running state identification.

Inventors

CAI CHANGQING
CHEN SIYAO
JIANG CHUNNING
YUAN HAO
CAO GUANGYU
KONG FANPING
LV YANSONG
SHEN BOYI
PENG HAICHAO
JU ZHENHE

Assignees

长春工程学院

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (6)

1. A non-invasive electrical load identification method based on a neural network model, the method comprising the steps of: acquiring current signals of the user in different detection periods of the electric appliance; The current signal is input into a neural network model to obtain a characteristic diagram, a parameter vector of each characteristic channel in each detection period is obtained based on the gray distribution of different characteristic channels in the characteristic diagram in each detection period, and a relative correlation coefficient of load detection data of each characteristic channel in each detection period is obtained according to the distribution difference of the parameter vectors of the same characteristic channel of the previous detection period adjacent to the detection period; Evaluating the channel information redundancy performance of each characteristic channel according to the overall distribution difference of the relative correlation coefficient of the load detection data of each characteristic channel in all detection periods and the parameter vector; introducing an ECA mechanism into the neural network model, and determining the characteristic importance weight of a characteristic channel by utilizing the redundant representation of channel information; the obtaining the relative correlation coefficient of the load detection data of each characteristic channel in each detection period comprises the following steps: The method comprises the steps of respectively marking the extreme differences of elements in reference vectors of channels to be analyzed in each period as first characteristic values of the channels to be analyzed in each period; calculating a first difference between the candidate period and a first eigenvalue of an expected analysis channel of the previous week of the candidate period; According to the first difference, obtaining a load steady-state coefficient of a candidate expected analysis channel; calculating a second difference between the mean value of the elements in the reference vector of the candidate week number expected analysis channel and the mean value of the elements in the reference vector of the previous week number expected analysis channel of the candidate period; Calculating a second ratio between the load steady-state coefficient of the candidate expected analysis channel and the second difference; Determining the normalization result of the second ratio as a relative correlation coefficient of load detection data of a channel to be analyzed in a candidate period; the candidate period is any detection period, and the channel to be analyzed is any characteristic channel; the evaluating the channel information redundant representation of each characteristic channel includes: Obtaining the load influence factors of the channels to be analyzed in each detection period according to the relative correlation coefficient of the load detection data of the channels to be analyzed in each detection period and the modular length of the parameter vector of the channels to be analyzed in each detection period; Obtaining channel information redundancy expression of the channel to be analyzed according to the difference between the load influence factors of the expected analysis channels of each detection week and the average value of the load influence factors of all the expected analysis channels of the detection week; the obtaining the load steady-state coefficient of the candidate expected analysis channel according to the first difference comprises the following steps: obtaining the maximum value of the first characteristic values of the channels to be analyzed in all periods; and calculating a first ratio between the first difference and the maximum value, and taking a negative correlation mapping result of the first ratio as a load steady-state coefficient of the channel to be analyzed of the candidate period.
2. The method for recognizing the non-invasive electrical load based on the neural network model according to claim 1, wherein the obtaining the parameter vector of each characteristic channel in each detection period based on the gray distribution of different characteristic channels in the characteristic map in each detection period comprises arranging gray values of all pixel points in each characteristic channel in the characteristic map in each detection period, respectively, to obtain the parameter vector of each characteristic channel in the characteristic map in each detection period.
3. The method for recognizing a non-invasive electrical load based on a neural network model according to claim 1, wherein the obtaining the load influence factor of the channel to be analyzed in each detection period according to the relative correlation coefficient of the load detection data of the channel to be analyzed in each detection period and the modular length of the parameter vector of the channel to be analyzed in each detection period comprises: And respectively taking the ratio between the modular length of the parameter vector of the expected analysis channel of each detection cycle and the relative correlation coefficient of the load detection data of the channel to be analyzed in the same detection cycle as the load influence factor of the expected analysis channel of each detection cycle.
4. The non-invasive electrical load recognition method based on neural network model according to claim 1, wherein the obtaining the redundant representation of the channel information of the channel to be analyzed according to the difference between the load influence factor of each test cycle expected analysis channel and the average value of the load influence factors of all test cycle expected analysis channels comprises: respectively recording a negative correlation mapping value of the difference between the load influence factors of the expected analysis channels of each detection cycle and the average value of the load influence factors of all the expected analysis channels of each detection cycle as a second characteristic value of the channel to be analyzed of each detection cycle; And taking the average value of the second characteristic values of all expected analysis channels in the detection period as the redundant representation of the channel information of the channel to be analyzed.
5. The neural network model-based non-intrusive electrical load identification method of claim 1, wherein the determining the feature importance weights of the feature channels using the redundant representation of channel information comprises: Calculating the accumulated sum of the channel information redundancy expressions of all the characteristic channels; and determining the ratio of the channel information redundancy expression of each characteristic channel to the accumulated sum as the characteristic importance weight of each characteristic channel.
6. The neural network model-based non-intrusive electrical load identification method of claim 1, wherein the determining a load operating state based on the feature importance weights comprises: Carrying out channel-by-channel multiplication operation on the feature importance weight and the original input feature map to obtain a feature map with channel attention; The load operation states are classified based on the characteristic map with the channel attention, and the load operation states include single load, double load, three load, four load and five load operation states.

Description

Non-invasive electricity load identification method based on neural network model Technical Field The invention relates to the technical field of load identification, in particular to a non-invasive electricity load identification method based on a neural network model. Background The non-invasive load refers to that under the condition that the system operation is not required to be changed or an additional sensor is added, corresponding electric operation data is obtained after the electric loads of different types are acquired and processed through real-time data, the load decomposition and identification are realized through methods such as data analysis processing or machine learning, the actual use condition, energy consumption and other energy consumption information of each electric equipment in a line are monitored, the load identification is core content of the non-invasive load identification, a common method is to analyze the electric operation states of different users through a clustering algorithm, and characteristic curves corresponding to different loads are obtained, so that the decomposition and identification of abnormal conditions are realized. The method for constructing the load database based on the steady state and transient state characteristics establishes a corresponding load identification model, introduces a multi-label classification technology into load identification research, and breaks through the limitation of the original ammeter on non-invasive load identification. The load mixed electricity data comprises all electricity consumption of the load, the load action changes the electricity consumption condition of the load, and the electricity consumption condition is represented in the mixed electricity data based on the additivity of the current. In non-invasive load identification, as the operating state of the load changes, the state of the load signature changes. The existing non-invasive load identification method still has obvious limitations in actual household application scenes, importance differences among different characteristic channels cannot be fully considered in the actual processing process, so that the characteristic response among the channels has obvious information redundancy problem, and the accuracy of the non-invasive load identification is lower. Disclosure of Invention In order to solve the problem of lower accuracy in the non-invasive load identification of the existing method, the invention aims to provide a non-invasive power utilization load identification method based on a neural network model, and the adopted technical scheme is as follows: the invention provides a non-invasive electricity load identification method based on a neural network model, which comprises the following steps: acquiring current signals of the user in different detection periods of the electric appliance; The current signal is input into a neural network model to obtain a characteristic diagram, a parameter vector of each characteristic channel in each detection period is obtained based on the gray distribution of different characteristic channels in the characteristic diagram in each detection period, and a relative correlation coefficient of load detection data of each characteristic channel in each detection period is obtained according to the distribution difference of the parameter vectors of the same characteristic channel of the previous detection period adjacent to the detection period; Evaluating the channel information redundancy performance of each characteristic channel according to the overall distribution difference of the relative correlation coefficient of the load detection data of each characteristic channel in all detection periods and the parameter vector; Introducing an ECA mechanism into the neural network model, determining the characteristic importance weight of the characteristic channel by utilizing the redundant representation of the channel information, and determining the load running state based on the characteristic importance weight. Preferably, the obtaining the parameter vector of each feature channel in each detection period based on the gray distribution of different feature channels in the feature map in each detection period includes respectively arranging gray values of all pixel points in each feature channel in the feature map in each detection period to obtain the parameter vector of each feature channel in the feature map in each detection period. Preferably, the obtaining the relative correlation coefficient of the load detection data of each characteristic channel in each detection period according to the distribution difference of the parameter vector of the same characteristic channel of each detection period and the adjacent previous detection period includes: The method comprises the steps of respectively marking the extreme differences of elements in reference vectors of channels to be analyzed in each period as first