Non-invasive Load Identification Method Based on BP Neural Network Model

Abstract

The present invention discloses a non-intrusive load identification method based on a BP neural network model, which relates to the field of power technology and includes the following steps: A. Acquisition of feature analysis samples, B. Acquisition of a feature sample set, C. Optimization of the BP neural network, and D. Output of the load identification result. The present invention collects the operation data of different household loads through the metering core of an electric energy meter, extracts the transient and steady-state change processes of the household loads as load features, and combines the BP neural network model to achieve the identification of household loads, thereby improving the accuracy and reliability of load identification on the residential side.

Description

Technical Field

[0001] The present invention relates to the field of power technology, and in particular, to a non-intrusive load identification method based on a BP neural network model. Background Art

[0002] Load identification methods mainly include two categories: intrusive load identification and non-intrusive load identification. Although the identification results of intrusive load identification methods are relatively accurate, they are not very popular due to high costs and other reasons. The non-intrusive load identification method (non-intrusive load monitoring, NILM) has low costs and strong practicability, so NILM has become a hot spot in the field of intelligent metering in today's power system. NILM installs an embedded non-intrusive power identification module on the household electricity meter, and then uses a load identification algorithm to detect the load working conditions in the building. The problems that need to be solved in load identification are to establish a feature library of known electrical appliances and compare the load features extracted from the collected data with the known feature library to identify the components of the total load and achieve load identification. The load identification based on pattern recognition is essentially to achieve the goal of load identification by learning the load features (transient, steady state, etc.) of various electrical appliances. There are many load identification algorithms based on pattern recognition, including the chicken swarm algorithm, hidden Markov model, support vector machine, etc. However, the types of loads processed by these algorithms are relatively simple, and the use of a BP neural network model can identify multiple electrical loads in complex scenarios, with good application prospects. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a non-intrusive load identification method based on a BP neural network model, analyze the user's power consumption situation, strengthen the prediction and management of the residential user load side, awaken the dormant adjustment ability on the user side, reasonably guide and encourage residential users to actively participate in the power demand side response, improve the efficiency of power use, achieve the purpose of peak shaving and valley filling, and enhance the reliability, flexibility and economy of the new power system.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is as follows: including the following steps:

[0005] A. Acquisition of feature analysis samples: A.1. Real-time collection of the instantaneous voltage value and instantaneous current value at the power inlet of the household to obtain a voltage data set U s i and a current data set I s i ; A.2. For the voltage data set U s i and the current data set Is i Process to obtain the set of instantaneous active power values of the load P s i and the set of instantaneous reactive power values Q s i ; A.3. According to the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i Perform data interception to obtain the feature analysis samples;

[0006] B. Acquisition of the feature sample set: B.1. Process the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i to form a feature data set composed of feature values by processing the time-domain feature values and frequency-domain feature values; B.2. Use the clustering analysis method to perform similarity measurement classification on the obtained feature data set and obtain the feature sample set composed of the optimal load feature values;

[0007] C. Optimization of the BP neural network: Train the BP neural network through the feature sample set and optimize the training result by the steepest descent method;

[0008] D. Output the load recognition result: For the load to be recognized, input the optimal load feature value in the load to be recognized into the optimized BP neural network model to obtain the load recognition result.

[0009] The beneficial technical effects of the present invention are: By collecting the operation data of different household loads through the electric energy meter metering core, extracting the transient and steady-state change processes of the household loads as load features, and combining with the BP neural network model to realize the identification of household loads, thereby improving the accuracy and reliability of load identification on the residential side.

[0010] The present invention will be described in detail below with reference to the accompanying drawings. Description of the Drawings

[0011] Figure 1 is the training flow chart of the present invention;

[0012] Figure 2 is the recognition flow chart of the present invention.

[0013] See attached Figure 1 and 2, the present invention provides a non-intrusive load identification method based on a BP neural network model, including the following steps.

[0014] A. Acquisition of feature analysis samples.

[0015] A.1. Instantaneous voltage values and instantaneous current values at the electricity inlet end are collected in real time to obtain a voltage data set U s i and a current data set I s i .

[0016] A.2. The voltage data set U s i and the current data set I s i are processed to obtain a set of instantaneous active power values of the load P s i and a set of instantaneous reactive power values Q s i .

[0017] Specifically, the window sliding method is used to perform FFT transformation on the data within each period of the voltage data set U s i and the current data set I s i to obtain the initial phase of the fundamental wave of the voltage and the initial phase of the fundamental wave of the current signal in each acquisition cycle Øu and the initial phase of the fundamental wave of the current signal Øi . According to the formula

[0018]

[0019] the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i are calculated respectively.

[0020] A.3. According to the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i data interception is performed to obtain feature analysis samples.

[0021] Specifically, according to the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i , the difference method is used to search for and intercept the time period of active power and reactive power for 3 cycles containing transient startup as a sample, and the sample is optimized through a sliding window and a two-sided CUSUM change point detection method. Optimizing the sample through a sliding window and a two-sided CUSUM change point detection method includes: using the slip method to perform slip grouping on the sets P s i and Q s i by the number of cycles. The cycle is T, the slip window is N, and the mean value of each group is calculated S i . If there is ∆η i ≥H within the connection time t, then the load is in the startup state during this time period, and the active power and reactive power data for 3 cycles within this time period are intercepted as a sample. Among them, ∆η i = S i -S i-1 , H is the minimum power change difference, and its value is an empirical value.

[0022] B. Acquisition of the characteristic sample set.

[0023] B.1. Process the time-domain eigenvalues and frequency-domain eigenvalues of the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i to form a characteristic data set composed of eigenvalues.

[0024] Specifically, in step B.1, the characteristic data set includes characteristics in the time domain and characteristics in the frequency domain. The characteristics in the time domain include the maximum value, minimum value, average value, peak-to-peak value, absolute value average, variance, standard value, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, and margin factor. The characteristics in the frequency domain include the frequency mean, center frequency, frequency root mean square, frequency standard deviation, fundamental wave phase and amplitude, and the phases and amplitudes of the 1st, 2nd, 3rd, 4th, and 5th harmonics.

[0025] B.2. Use the clustering analysis method to measure and classify the similarity of feature quantities in the obtained feature data set, and obtain a feature sample set composed of optimal load characteristic values.

[0026] Denote the value of the eigenvalue N j as

[0027] .

[0028] The correlation coefficient between N j and N k can be used as the similarity measure between the two feature quantities.

[0029] Specifically, through

[0030]

[0031] calculate the similarity between the feature quantities of N j and N k , where for all j, k, | r jk | ≤ 1, r jk =r kj .

[0032] | r jk | The closer it is to 1, the more correlated or similar N j and N k are. | r jk | The closer it is to zero, the weaker the similarity between N j and N k .

[0033] Obtain the correlation coefficients between each feature and further perform clustering. Select the fundamental wave amplitude, 1st, 2nd, 3rd, 4th, 5th harmonic amplitudes, waveform factor, and margin factor from the clustering results to form a feature sample set.

[0034] In the variable set clustering analysis, here use the maximum coefficient method to define the distance between two types of variables as

[0035]

[0036] At this time, R(G1, G2) is equal to the similarity measure value between the two most similar variables in the two classes.

[0037] C. Optimization of the BP neural network: Train the BP neural network through the feature sample set, and optimize the training results by the steepest descent method.

[0038] Specifically, select the sigmoid function as the activation function of the BP neural network model, and its form is:

[0039]

[0040] α > 0, the slope can be controlled; the state of the neural network output unit Q s i (i = 1, 2, … n) is

[0041]

[0042] where m is the number of neurons in the selected hidden unit layer, w ij is the weight from the middle layer to the output layer, v jk is the weight from the input layer to the middle layer, W s is the input layer of the neural network, and the value is I s k ([[]] k = 1, 2, … 8), 8 is the number of features selected for each sample, s is the total number of feature samples of all loads, i, j, k correspond to the output layer, the middle layer, and the input layer respectively.

[0043] Output unit Q s i and the ideal output T s i The difference is denoted as:

[0044]

[0045] where, X is any set of weight values w ij and v jk of ρ the solution of the latitude vector, and the weights corresponding to the minimum value of E are obtained using the steepest descent iteration algorithm w ij and the weight v jk .

[0046] D. Output the load recognition result: For the load to be recognized, input the optimal load feature value in the load to be recognized into the optimized BP neural network model to obtain the load recognition result.

[0047] Taking the collection of several commonly used household appliances as a sample training library and using this training model to randomly identify a certain type of electrical appliance as an example. The training electrical appliances included are: induction cooker, water heater, hot water kettle, air conditioner (since the cooling and heating states of the air conditioner are quite different and are regarded as two electrical appliances here), rice cooker, microwave oven, washing machine, a total of 7 electrical appliances. The identified electrical appliances are the random combined operating states of these 7 electrical appliances.

[0048] Step 1: Obtaining the feature analysis samples.

[0049] 1-1: Separately collect the instantaneous voltage values and instantaneous current values of the induction cooker, water heater, hot water kettle, air conditioner in the cooling state, air conditioner in the heating state (since the cooling and heating states of the air conditioner are quite different and are regarded as two electrical appliances here), rice cooker, microwave oven, washing machine from the start-up - steady-state operation - shutdown time period, and record them as U s i and I s i . (Where s is from 1 to 8, respectively representing the above-mentioned several sample states, and i is the total number of sampling points for each electrical appliance; the sampling frequency is 6.4k, that is, 128 sampling points per cycle).

[0050] 1-2: For the U s i and I s i obtained in step 1-1, adopt the window sliding difference method with a window of 1, and perform FFT transformation on each cycle (128 data points) to obtain the fundamental wave initial phase of the voltage Øu and the initial phase of the fundamental wave of the current Øi , and calculate the set of instantaneous active power values

[0051]

[0052] of the above 8 electrical appliances and the set of instantaneous reactive power values P s i and Q s i respectively according to the formula (where T is the number of sampling points in each cycle, which is 128 here).

[0053] 1-3: Through the sets P s i and Q s i respectively determine the start-up time period of the electrical appliance: First, respectively take the sets P s iand Q s i Group by the sliding window method according to the number of periods, with a window of 5 and a number of periods of 128, and calculate the grouped average value S i , when it is judged that there are ∆η i ≥H during the continuous time t (where ∆η i =S i -S i-1 , H is the minimum power change difference, and its value is an empirical value), then the electrical device is in the starting transient state during this time period, and the sets P s i and Q s i The data of 3 periods within this time period are used as a feature sample. Finally, a new set of instantaneous active power values P s j and the set of instantaneous reactive power values Q s j (where j ∈ i; it is the time period of 3 starting cycles).

[0054] Step 2: Obtaining the feature sample set.

[0055] 2-1: For the P s j and Q s j obtained in 1-3, calculate the time-domain features of each electrical device according to a period of 128: maximum value, minimum value, average value, peak-to-peak value, absolute value average, variance, standard value, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, and margin factor; and frequency-domain features: frequency mean, center frequency, frequency root mean square, frequency standard deviation, fundamental wave phase and amplitude, 1st, 2nd, 3rd, 4th, and 5th harmonic phases and amplitudes.

[0056] 2-2: Optimize the load feature sample set: Use variable clustering analysis to calculate the similarity distance between every two feature values calculated in 2-1 above (using the maximum coefficient to define the distance between two types of features: R(G1,G2)=max{r jk}), and the calculation method is:

[0057]

[0058] | rjk |The closer to 1, the more j related or similar to N. k | r jk |The closer to zero, the weaker the j similarity between N k and N (N is a set of characteristic quantities).

[0059] Select representative eigenvalues from the clustering results. Here, the fundamental wave amplitude, the amplitudes of the 1st, 2nd, 3rd, 4th, and 5th harmonics, as well as the waveform factor and the margin factor are further used to form the final characteristic sample set P s j and Q s j (where P s j and Q s j , s is 1 - 8 types of electrical appliances; j is the total number of final samples formed for each type of electrical appliance, and each sample of each electrical appliance is an 8 - dimensional vector, representing 8 characteristics of the sample).

[0060] Step 3: According to the final sample feature set obtained in Step 2 - 2, use a BP neural network for modeling and optimization, and use the steepest descent method for training optimization. Here, the sigmoid function is selected as the activation function of the neural network model, and its functional form is as follows:

[0061]

[0062] α > 0, which can control its slope;

[0063] The state of the neural network output unit Q s i ( i = 1, 2, … n) is:

[0064]

[0065] where m is the number of neurons in the selected hidden unit layer. Here, the number of hidden layer neurons is taken as 20 w ij is the weight from the middle layer to the output layer, v jk is the weight from the input layer to the middle layer, W s is the input layer of the neural network, and its value is I s k (k (i = 1, 2, …, 8), k is the number of features selected for each sample, and s is the total number of feature samples of all loads. i, j, k They respectively correspond to the output layer, the middle layer, and the input layer.

[0066] Output unit Q s i And the ideal output T s i The difference is denoted as:

[0067] .

[0068] Wherein, X is any set of weight values w ij and v jk of ρ The latitude vector solution. Use the steepest descent iterative algorithm to find the weights corresponding to the minimum value of E w ij and weights v jk .

[0069] Step 4: Use the BP neural network model to discriminate the types of electrical appliances.

[0070] 4-1: According to Step 1, obtain the set of instantaneous active power values P s j within a certain operating time of an unknown electrical appliance and the set of instantaneous reactive power values Q s j . From Step 2, obtain P s j and Q s j The fundamental wave amplitude, the 1st, 2nd, 3rd, 4th, and 5th harmonic amplitudes, as well as the waveform factor and the margin factor, are further used to form the final feature sample set P 终 s j and Q 终 s j . Input the finally obtained sample set into the BP neural network model optimized in Step 3, and output the result. Let h > 0.7 in the recognition result be recognized as a corresponding type of electrical appliance, and h < 0.7 be not that electrical appliance. (Where h is the recognition result of this unknown electrical appliance corresponding to 8 training electrical appliance types respectively, 0 < h < 1).

[0071] The present invention samples in real time the voltage and current transient or steady-state waveform data of different household loads of users from an electricity meter metering core through SPI communication and sends them to a load identification module. The load identification module performs algorithm processing on the sampled data of multiple cycles taken, uses this as a data set for clustering and training, generates a feature library, and combines it with a BP neural network model to achieve the identification of household loads, thereby improving the accuracy and reliability of load identification on the residential side.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that it is still possible to modify the specific implementation manners of the present invention or perform equivalent replacements on some technical features; without departing from the spirit of the technical solutions of the present invention, they should all be covered within the scope of the technical solutions claimed by the present invention.

Claims

1. A non-intrusive load identification method based on a BP neural network model, characterized in that, Including the following steps: A. Acquisition of characteristic analysis samples: A.

1. Instantaneously collect the voltage instantaneous value and current instantaneous value at the electricity access end to obtain the voltage data set U s i and the current data set I s i ; A.

2. Process the voltage data set U s i and the current data set I s i to obtain the set P of instantaneous active power values of the load s i and the set Q of instantaneous reactive power values s i ; A.

3. According to the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i perform data interception to obtain a feature analysis sample; B. Acquisition of characteristic sample sets: B.

1. Process the set of instantaneous active power values \(P\) s i and the set of instantaneous reactive power values \(Q\) s i to form a feature data set consisting of eigenvalues through processing of time-domain eigenvalues and frequency-domain eigenvalues; The characteristic data set includes characteristics in the time domain and characteristics in the frequency domain. The characteristics in the time domain include maximum value, minimum value, average value, peak-to-peak value, absolute average value, variance, standard value, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, and margin factor. The characteristics in the frequency domain include frequency mean, center frequency, root mean square of frequency, standard deviation of frequency, fundamental wave phase and amplitude, and phases and amplitudes of the 1st, 2nd, 3rd, 4th, and 5th harmonics. B.

2. Using the clustering analysis method to conduct similarity measurement classification on the obtained characteristic data set and obtaining a characteristic sample set composed of optimal load characteristic values. By calculating N j and the similarity between N k and the characteristic quantities, where , for all j, k; , for all j, k; Obtaining the correlation coefficients between various characteristics and further clustering, and selecting the fundamental wave amplitude, amplitudes of the 1st, 2nd, 3rd, 4th, and 5th harmonics, as well as the waveform factor and margin factor from the clustering results to form a characteristic sample set. C. Optimization of the BP neural network: Training the BP neural network through the characteristic sample set and optimizing the training result by the steepest descent method. D. Outputting the load identification result: For the load to be identified, inputting the optimal load characteristic values in the load to be identified into the optimized BP neural network model to obtain the load identification result.

2. The non-invasive load identification method based on the BP neural network model according to claim 1, characterized in that In step A.2, the voltage data set U is respectively processed by means of window slip s i and the current data set I s i The data within each period is subjected to FFT transformation to obtain the initial phase Øu of the fundamental wave of the voltage and the initial phase Øi of the fundamental wave of the current signal under each acquisition cycle. According to the formula Calculate the set of instantaneous values of active power P separately s i and the set of instantaneous values of reactive power Q s i .

3. The non-intrusive load identification method based on the BP neural network model according to claim 1, wherein In step A.3, according to the set of instantaneous active power values P s i and the set of instantaneous reactive power values Q s i , the progressive difference method is used to search for and intercept a time period of active power and reactive power that contains 3 cycles of transient start as a sample, and the sample is optimized by means of a sliding window and a double-sided CUSUM change point detection method.

4. The non-intrusive load identification method based on the BP neural network model according to claim 3, characterized in that, Optimizing the sample through a sliding window and a double-sided CUSUM change point detection method includes: using a slip method for set P s i and Q s i to perform slip grouping by cycle number. The cycle is T, the slip window is N, and the mean value S of each group is calculated i . If there is ∆η i ≥H within the connection time t, then the load is in the starting state during this time period, and the active power and reactive power data of 3 cycles within this time period are intercepted as a sample. Among them, ∆η i =S i -S i-1 , and H is the minimum power change difference, and its value is an empirical value.

5. The non-intrusive load identification method based on the BP neural network model according to claim 1, wherein In step C, the activation function of the BP neural network model selects the sigmoid function, and its form is: , where α>0 can control its slope; Neural network output unit state is , where m is the number of neurons in the selected hidden unit layer, is the weight from the middle layer to the output layer, is the weight from the input layer to the middle layer, is the input layer of the neural network, and the value is , 8 is the number of features selected for each sample, s is the total number of feature samples of all loads, and i, j, k correspond to the output layer, the middle layer, and the input layer respectively; Output unit The difference from the ideal output is denoted as: Among them, is any set of weights and of the weft vector solution. Use the steepest descent iterative algorithm to find the weights corresponding to the minimum value of E and weights .

Images