CN113420436A - Non-invasive load switch event detection method based on GRU and ARIMA-T detection - Google Patents

Abstract

The invention relates to a non-intrusive load switch event detection method based on GRU and ARIMA-T detection, which specifically comprises the following steps: firstly, inputting load U, I data of a certain specific area, preprocessing the data, enhancing the number of virtual samples by using a Mixup method for less sample category data to obtain a balanced sample data set, calculating the input U, I data to obtain active power characteristics, then performing median filtering processing on the active power, and performing stationarity processing on the filtered data by using differential operation to obtain an input data sequence; respectively inputting the input data sequence into a GRU model and an ARIMA model to obtain the result of the load event; and optimizing the weight coefficient by using a differential evolution algorithm, and performing linear weighted combination by using the optimal weight coefficient to obtain a load event result based on GRU and ARIMA-T test. The method of the invention combines the GRU model and the ARIMA-T inspection model, improves the identification precision and generalization capability of the load switch event of the model, has stronger global search capability and capability of resisting local convergence, can better identify the load event, and has certain application value.

Description

Non-invasive load switch event detection method based on GRU and ARIMA-T detection

Technical Field

The invention relates to the field of non-invasive load detection, in particular to a non-invasive load switch event detection method based on GRU and ARIMA-T detection.

Background

Load monitoring is an important ring for realizing intelligent power utilization, power consumption conditions of all electrical appliances can be known by users in time through load monitoring, Non-invasive power load monitoring is a development trend and direction of a future monitoring technology, and as an important load monitoring mode, household energy consumption data can be classified into single equipment utilization data by applying an NILM (Non-intrusive load monitor) technology, so that the power consumption of the users can be measured in a dispersed and concentrated mode.

The effect based on event detection is directly influenced by the precision of a detection algorithm, if the running state of an electric appliance is changed, the switching event is detected by the algorithm, feature extraction and load identification can be completed by the algorithm, and at present, the event detection has several methods, such as a Cumulative Sum algorithm (cumulant Sum, CUSUM), a Log Likelihood Ratio (LLR) and a broad Log likelihood ratio, but the method has certain defects, for example, the accuracy of detecting the on-off state is reduced based on the similar waveforms of the Log likelihood ratio. The non-invasive load switch event detection method based on GRU and ARIMA-T detection can effectively improve the event detection precision and reduce the false detection rate and the missing detection rate.

Disclosure of Invention

The invention mainly aims to provide a non-intrusive load switch event detection method based on GRU and ARIMA-T tests.

The method comprises the following steps:

step1, inputting load U, I data of a specific area, and preprocessing the data;

step2, calculating the preprocessed data to obtain active power characteristics, then carrying out median filtering processing on the active power, and carrying out stationarity processing on the filtered data by utilizing differential operation to obtain an input data sequence;

step3, inputting the input data sequence into the GRU model to obtain a load event result 1;

step4, inputting the input data sequence into the ARIMA model, and obtaining a load event result 2 through T test;

step5, optimizing weight coefficients by using a differential evolution algorithm, and performing linear weighted combination by using the optimal weight coefficients to obtain a load event result based on GRU and ARIMA-T test;

the invention relates to a non-invasive load switch event detection method based on GRU and ARIMA-T detection, which comprises the steps of firstly inputting load U, I data of a certain specific area, preprocessing the data, and enhancing virtual data of few sample types by using a Mixup method

Drawings

For a more intuitive description of the embodiments of the present invention, reference will now be made in detail to the accompanying drawings of embodiments of the present patent:

FIG. 1 is a flow chart of a non-intrusive load switch event detection method based on GRU and ARIMA-T tests, in accordance with an embodiment of the present invention;

FIG. 2 is a flow chart of the GRU and ARIMA-T test performed in accordance with the present invention.

Detailed Description

The invention mainly aims to provide a non-intrusive load switch event detection method based on GRU and ARIMA-T tests.

The method comprises the following steps:

step1, inputting load U, I data of a specific area, and preprocessing the data;

step2, calculating the preprocessed data to obtain active power characteristics, then carrying out median filtering processing on the active power, and carrying out stationarity processing on the filtered data by utilizing differential operation to obtain an input data sequence;

step3, inputting the input data sequence into the GRU model to obtain a load event result 1;

step4, inputting the input data sequence into the ARIMA model, and obtaining a load event result 2 through T test;

step5, optimizing weight coefficients by using a differential evolution algorithm, and performing linear weighted combination by using the optimal weight coefficients to obtain a load event result based on GRU and ARIMA-T test;

the Step1 inputs load U, I data of a certain specific area, and the concrete steps of preprocessing the data are as follows:

after load U, I data of a certain specific area are input, the number of samples of certain categories is far less than that of other categories, the load samples are balanced by a data enhancement method Mixup, the Mixup is a data enhancement method based on the neighborhood risk minimization principle, new data samples are obtained through linear interpolation, and the method has the advantages of low calculation cost, capability of reducing the memory of a model to a damaged label, and model robustness and stability enhancement.

The formula for the calculation of Mixup is:

x％＝λx_i+(1-λ)x_j,y％＝λy_i+(1-λ)y_j

where λ Beta (α, α), α ∈ (0, ∞). (x)_i,y_i) And (x)_j,y_j) Are two sets of samples randomly decimated from the input voltage current data, and λ ∈ [0,1 ]]The hyper-parameter α limits the interpolation strength between feature objects, typically α is 0.5.

Calculating the cosine similarity of the original sample and the virtual sample by the following formula:

wherein A is_iAnd B_iData before and after enhancement, respectively. The larger the cosine similarity value is, the more similar the extended data sample of the Mixup method is to the original data sample.

Step2, calculating the preprocessed data to obtain active power characteristics, then performing median filtering processing on the active power, and performing stationarity processing on the filtered data by using differential operation to obtain an input data sequence, wherein the specific steps of the method are as follows:

calculating the preprocessed data to obtain the active power characteristic, for P_nSet y (y is typically an odd number) as (1,2, …, N) the length of the sliding window, which is a neighborhood (P)_x,P_x+1,…,P_x+y) The values in the neighborhood are arranged in order from small to large, with the median of the series replacing the value at its central position. Can take two numbers P in the middle of the sequence_x+i,P_x+jAverage value of (2)

As the median of the array, the noise points in the array can be effectively removed.

And then, carrying out median filtering processing on the calculated active power to remove errors caused by data fluctuation, and then further carrying out stationarity processing on the filtered data by adopting differential operation to obtain an input data sequence.

Step3, inputting the input data sequence into the GRU model to obtain a load event result 1, and the specific steps are as follows:

wherein, the calculation formulas of the GRU model are respectively as follows:

z_t＝σ(W_z·[h_t-1,x_t])

r_t＝σ(W_r·[h_t-1,x_t])

after the input data sequence is input, pass h_tObtaining a load event result 1

Step4 inputs the input data sequence into the ARIMA model, and obtains a load event result 2 through T test, which comprises the following specific steps:

in ARIMA, the future values of variables are assumed to be linear functions of the past observations and random errors. That is, the basic process of generating a time series has the following form:

y_t＝θ₀+φ₁y_t-1+φ₂y_t-2+…+φ_py_t-p +ε_t-θ₁ε_t-1-θ₂ε_t-2-…-θ_qε_t-q

in the mode, phi is included in the model_i、θ_jTwo parameters, where i ═ 1,2, …, p, j ═ 1,2, …, q; the actual value, the random error of independent identity distribution and the order of the model are respectively expressed as y_t、ε_tP and q;

the basic principle of the T-test is as follows: two groups of samples are respectively recorded to respectively satisfy the following conditions:

and X₁＝(x₁₁,x₁₂,L,x_1i,L,x_1n)，X₂＝(x₂₁,x₂₂,…,x_2j,…,x_2m) Then, the mean and variance of the two groups of samples are respectively:

simultaneously defining: t is the statistic of two groups of data samples, F is the variance ratio thereof, i.e.

The active power sequence characteristics of the ARIMA decomposition load are applied, the three decomposed time sequences are used as analysis objects, residual error items in the three time sequences are extracted and used as new data sequences, the primary variance and the secondary variance are calculated at the same time, and the result of the T test can be used as a load event detection result 2.

Step5 optimizes the weight coefficient by using a differential evolution algorithm, and performs linear weighted combination by using the optimal weight coefficient to obtain a load event result based on GRU and ARIMA-T test, wherein the method comprises the following specific steps:

and optimizing a weight coefficient for the load event result 1 of the GRU model and the load event result 2 of the ARIMA-T test model by using a differential evolution algorithm, and performing linear weighted combination by using the optimal weight coefficient to obtain a load event result based on GRU and ARIMA-T test.

The patent has certain universality, various equivalent transformations are carried out on the technical scheme of the invention within the technical idea scope of the invention, and the direct or indirect application in other related technical fields is within the patent protection scope of the invention.

Claims (6)

1. A non-intrusive load switch event detection method based on GRU and ARIMA-T detection is characterized in that: the GRU model and the ARIMA-T inspection model are combined, so that the load switch event identification precision and generalization capability of the model are improved, the global search capability and the capability of resisting local convergence are stronger, the load event can be identified better, and the method specifically comprises the following steps:

step1, inputting load U, I data of a specific area, and preprocessing the data;

step2, calculating the preprocessed data to obtain active power characteristics, then carrying out median filtering processing on the active power, and carrying out stationarity processing on the filtered data by utilizing differential operation to obtain an input data sequence;

step3, inputting the input data sequence into the GRU model to obtain a load event result 1;

step4, inputting the input data sequence into the ARIMA model, and obtaining a load event result 2 through T test;

step5 optimizes the weight coefficient by using a differential evolution algorithm, and performs linear weighted combination by using the optimal weight coefficient to obtain a load event result based on GRU and ARIMA-T test.

2. The method of inputting and pre-processing area-specific load U, I data according to claim 1, wherein:

load U, I data of a certain specific area are input, data are preprocessed, because the number of samples of a category is far less than that of data of other categories, a training flooding phenomenon can be generated, the feature training of the samples of the category is insufficient, load samples are balanced by adopting a Mixup algorithm, the Mixup is a data enhancement method based on a neighborhood risk minimization principle, new data samples are obtained through linear interpolation, the calculation cost is low, the memory of a model to damaged labels can be reduced, and the robustness and the stability of the model are enhanced.

3. The method of claim 1, wherein the step of computing the pre-processed data to obtain the active power characteristics, the step of performing median filtering on the active power, and the step of performing stationarity processing on the filtered data by using differential operation to obtain the input data sequence comprises:

calculating the preprocessed data to obtain the active power characteristic, for P_nSet y (y is typically an odd number) as (1,2, …, N) the length of the sliding window, which is a neighborhood (P)_x,P_x+1,…,P_x+y) Arranging the numerical values in the neighborhood from small to large, replacing the numerical value of the central position of the numerical sequence with the median of the numerical sequence, and taking two numbers P in the middle of the numerical sequence_x+i,P_x+jAverage value of (2)

As the median of the array, the noise points in the array can be effectively removed;

and then, carrying out median filtering processing on the calculated active power to remove errors caused by data fluctuation, and then further carrying out stationarity processing on the filtered data by adopting differential operation to obtain an input data sequence.

4. The method of inputting an input data sequence into a GRU model to obtain a load event result 1 as claimed in claim 1 wherein:

wherein, the calculation formulas of the GRU model are respectively as follows:

z_t＝σ(W_z·[h_t-1,x_t])

r_t＝σ(W_r·[h_t-1,x_t])

after the input data sequence is input, pass h_tLoad event result 1 is obtained.

5. The method of claim 1, further inputting the input data sequence into an ARIMA model and obtaining a load event result 2 by a T-test, wherein:

in ARIMA, the future values of variables are assumed to be linear functions of the past observations and random errors. That is, the basic process of generating a time series has the following form:

y_t＝θ₀+φ₁y_t-1+φ₂y_t-2+…+φ_py_t-p+ε_t-θ₁ε_t-1-θ₂ε_t-2-…-θ_qε_t-q

in the mode, phi is included in the model_i、θ_jTwo parametersWherein i is 1,2, …, p, j is 1,2, …, q; the actual value, the random error of independent identity distribution and the order of the model are respectively expressed as y_t、ε_tP and q;

the basic principle of the T-test is as follows: two groups of samples are respectively recorded to respectively satisfy the following conditions:

and X₁＝(x₁₁,x₁₂,L,x_1i,L,x_1n)，X₂＝(x₂₁,x₂₂,…,x_2j,…,x_2m) Then, the mean and variance of the two groups of samples are respectively:

simultaneously defining: t is the statistic of two groups of data samples, F is the variance ratio thereof, i.e.

F(n-1,m-1)

The active power sequence characteristics of the ARIMA decomposition load are applied, the three decomposed time sequences are used as analysis objects, residual error items in the three time sequences are extracted and used as new data sequences, the primary variance and the secondary variance are calculated at the same time, and the result of the T test can be used as a load event detection result 2.

6. The method as claimed in claim 1, wherein the weight coefficients are optimized by a differential evolution algorithm, and the optimal weight coefficients are used for linear weighted combination to obtain the load event result based on GRU and ARIMA-T tests, wherein:

and optimizing a weight coefficient for the load event result 1 of the GRU model and the load event result 2 of the ARIMA-T test model by using a differential evolution algorithm, and performing linear weighted combination by using the optimal weight coefficient to obtain a load event result based on GRU and ARIMA-T test.

Images