TWI427547B - An Adaptive Non - invasive Method for Extracting Load from Artificial Intelligence - Google Patents

Description

Adaptive non-intrusive load characteristic extraction method for artificial intelligence technology

The present invention relates to a transient analysis of a total power waveform using a non-intrusive load monitoring device, and a technique for employing artificial intelligence and a related method for use in a load identification use strategy.

The non-intrusive load monitoring device only needs to know the start/stop operation state of each load by analyzing the total power waveform measured at the power inlet. The device of this monitoring device is nothing more than the desire that each load on the demand side can eventually be implemented for load management. The load monitoring method and the sensing signal processing of the non-intrusive load monitoring device are different from the traditional intrusive load monitoring device.

Compared with the traditional intrusive load monitoring device, the non-intrusive load monitoring device only needs to combine the voltage and/or current sensor with the user meter at the power supply inlet, and the monitoring device at the power inlet is analyzed by the analysis device. The measured total power waveform can effectively know the power usage of each load. Since only two sensors need to be installed at the power inlet, the monitoring device has the advantages of simple installation and disassembly, easy maintenance, low cost and economic benefit. However, the non-intrusive load monitoring device is more complicated and more difficult than the traditional intrusive load monitoring device in signal analysis and load identification methods.

The basic monitoring concept of a non-intrusive load monitoring device is based on the unique power characteristics that can be mined based on the power waveform generated by each load during start/stop. For example, real power and reactive power. That is, these power features mined from the total power waveform hide the message of each load operation. When the load operation changes, the monitoring device can identify the load by using some methods.

Prior art 1, using fuzzy logic theory for pattern recognition. This prior art processes the known load transient waveform into a home function similar to that proposed by the fuzzy theory and stores it in the database of the monitoring device. When the monitoring device performs load identification, the monitoring device calculates the maximum approaching degree of the load waveform to be identified and each known waveform stored in the database.

Prior to Technology 2, the Electric Power Research Institute (EPRI) developed a non-intrusive load monitoring system based on steady state analysis. The system uses a variation plane composed of effective power and reactive power characteristics (△P-△Q plane) and follows Edge Detection, Cluster Analysis, Cluster Matching, and exception analysis. Anomaly Resolution) and Load Identification are five steps to achieve load monitoring tasks. It is worth noting that in the "exceptional analysis" step, the system combines the effective power of each load with the arrangement of the reactive power to add a combination of the effective power and the ineffective power to guess the paired clusters to infer the possible different load combination optimal solutions. But when the system is faced with a large load of monitoring types ( n ), all possible permutations of the load operation will exhibit exponential growth (2 ⁿ ). Therefore, the effectiveness of the monitoring system remains to be assessed. In addition, if the monitoring system only uses the effective power of the load and the reactive power as the basis for identifying the load, the monitoring system will not be able to identify different kinds of equivalent loads.

In the prior art 3, all possible permutation combinations of the monitoring device for the load operation situation are proposed by using a different training algorithm for the inverse transfer neural network, Learning Vector Quantization (LVQ) and Probabilistic Neural (Probabilistic Neural). Network, PNN) is identified on the actual load test and the Electro-Magnetic Transient Program (EMTP) simulation. Among them, the input of the load identifier is steady-state energy, effective power, reactive power, current/voltage harmonic distortion, total current/voltage harmonic distortion (total current/ Voltage harmonic distortion) and the initiation of turn-on transient energy. In addition, the load monitored by the proposed monitoring device also includes an equivalent load. In this prior art, the transient energy signature is activated by the identification result to greatly influence the quality of the load result, and also improves the calculation time required by the load identifier to identify multiple loads. In addition, if the load recognizer only uses the effective power and the invalid power as the basis for identifying the equivalent load, the identification result of the load recognizer will be less than ideal.

It can be seen that there are still many shortcomings in the above-mentioned household items, which is not a good designer and needs to be improved.

In view of the shortcomings derived from the above-mentioned conventional technologies, the inventors of the present invention have improved and innovated, and after years of painstaking research, finally succeeded in research and development of an adaptive non-intrusive load characteristic extraction method of artificial intelligence technology. And its monitoring device.

The monitoring method of the adaptive non-intrusive load monitoring device of the present invention is shown in FIG. The monitoring method of the monitoring device is composed of three steps of "pre-data collection processing step" 11, "event detection and load current pattern extraction step" 12 and "feature identification step" 13. The monitoring device of the invention is designed by combining the load current pattern capturing method and the artificial intelligence identification technology, and the adaptive capability can be achieved through the optimization strategy.

The artificial intelligence identification technology includes: k-Nearest Neighbor Rule (k-NNR), inverse transfer neural network, artificial immune algorithm; the monitoring device uses k-nearest neighbor rule and Reverse the neural network to identify the start-up and shutdown-down of the actual load. It has the ability to adapt by using the Artificial Immune System (AIS). The following are introduced separately:

1. nearest neighbor rule (k-Nearest Neighbor Rule, k-NNR)

The k-nearest neighbor discriminator is a statistical classification method for nonparametric techniques. The recognizer is an instance-based learning method, and the method does not require any pre-hypothesis on the distribution of samples. The recognizer is constructed based on the concept of "objects are gathered together". Aggregation means that the distances of vectors (samples) of the same category from each other in the feature space should be relatively close. For vectors that classify unknown categories, the k-nearest neighbor rule first finds the vector that is closest to the vector k in the data set of the known category. Then, the category labels of the k nearest vectors are passed through a majority voting mechanism to determine the category of the vector. The identification of the k-nearest neighbor rule is shown in Figure 2.

Suppose the training data set {(x1, xd1), (x2, xd2), ..., (xm, xdm), ..., (xn, xdn)} is a known vector in n feature spaces and its corresponding category. (xm, xdm) is the data order pair m, and xm is a vector whose known class is xdm. When the training data set of the k-nearest neighbor rule is collected, the training phase is also completed. That is, the k-nearest neighbor recognizer does not need to be trained by any training algorithm. Regarding the vector y that classifies an unknown class, the squared Euclidean distance of the vector and the vector of each known class in the training data set can be calculated by the formula (1.1). These distances are sorted after the distance is calculated. The purpose of sorting is to find the vectors of the k known categories that are closest to the vector of the unknown category in the training data set. Finally, the category of the unknown class vector can be determined via the majority voting mechanism using equation (1.2). That is, these k known categories of vectors are used to determine the class of the vector of the unknown category.

Where: 1 represents the dimension of the feature space, M is the total dimension of the feature space, and y and xm represent the vector of the unknown category and the vector of the known category, respectively.

k-NNR( y )= y ^c (1.2)

Where: y ^c represents the category identified for the vector y of the unknown category, , and vot [‧] indicates a majority voting mechanism.

In addition, when the result of the voting occurs, the recognizer can represent the weight value of the voting based on the distance of the distance. This is called the distance weighted k-NNR.

2. Inverted neural network

The neural network of the biological neural network is composed of many large numbers of artificial nerve cells. Artificial nerve cells 2 are also referred to as artificial neurons, neuron-like or processing elements. The output of each processing unit is fanned out and these fanned out outputs will be inputs to other processing units. Regarding the relationship of the output of the processing unit to the input, the processing unit is generally expressed by taking the input weighted product sum and converting via an activation function. As shown in the artificial neural cell model in Figure 3, among them:

x _j : Input signal of biological nerve cells.

w _ij : the bonding strength of the signal transmission path between biological cells. It is called the so-called junction weights. The weighting value represents the degree of coupling of the jth processing unit (input) to the i th processing unit.

f(‧): Input weighted product and function 21 of artificial nerve cells. This function value is often referred to as a net input.

a (‧): activation function 22 of artificial nerve cells. It is a function that converts the net input value into the output value of the processing unit. Commonly used activation functions include linear functions, log sigmoid functions, hyperbolic tangent sigmoid functions, symmetric hard limit functions, and saturated linear functions. Saturating linear function).

θ _i : Threshold of artificial nerve cells. f(‧), a(‧) and θ _i are signal processing mechanisms that mimic the nucleus.

y _i : The output signal of the biological nerve cells.

The reverse-transfer-like neural network is the most representative and most widely used network learning model in various fields. The inverted delivery network is a multi-layer feedforward network of supervised learning. The architecture of the network includes an input layer, a hidden layer, and an output layer, and the hidden layer may have more than one layer. By introducing a hidden layer, the inverted transit network can map the feature space to a high dimension so that the network can solve complex classification problems. In addition, the inverted transfer network can also be used as an arbitrary function approximator according to the universal approximation theorem.

The learning process of the inverse transfer neural network is divided into two parts: forward propagation and back propagation. The network can utilize different training methods, such as steepest descent technique, variable learning rate back-propagation, and one-step secant algorithm. The iteratively minimizes the error function of the network, and the weighting and biasing parameter values of the network are adjusted during the minimization of the error function. The adjusted parameter values cause the error between the output value of the network and the desired outputs to be reduced. Commonly used error functions have a mean squared error function and a sum squared error function.

The architecture of the inverted transit neural network is shown in Figure 4, which includes:

Input layer ( x _j ): is the input variable of the network. The number of processing units of this layer depends on the problem.

Hidden layer ( z _q ): The middle layer that affects the interaction between processing units. There is no standard way to determine the number of layers and the number of processing units. The optimal number usually needs to be determined experimentally. Of course, the number of processing units can also be determined by adopting a pruning method or a constructing method.

Output layer ( y _i ): is the output variable of the network. The number of processing units of this layer depends on the problem.

The processing unit of the inverse transfer type neural network first multiplies each input signal (the bias weight is regarded as the input signal equal to 1) by the corresponding weighting values, and then takes the summed results to obtain the sum. Net input value. The net input value obtained is converted via the activation function as an input to the next processing unit. These input signals are forward-transferred in such a way that the output signal of the output layer processing unit is subtracted from the target value to obtain an error signal. Finally, all error signals are passed back to the ground based on the training algorithm to change the weight of the network. Through repeated learning procedures, the network's error function value gradually converges to an acceptable error value.

The mathematical derivation of the training process of the inverted transmission network is as follows:

Hidden layer processing unit input is

Where: v _qj represents the joint weighting value of the qth processing unit of the hidden layer and the jth processing unit of the input layer.

The hidden layer processing unit is converted to an output by an activation function

Output layer processing unit input is

Where: w _iq represents the joint weighting value of the i-th processing unit of the output layer and the qth processing unit of the hidden layer.

Define the error function as

The weight between the output layer and the hidden layer is updated to

Where: η represents the learning rate.

Using the chain rule and (1.5)-(1.7)

The δ _oi of (1.9) is the error signal of the output layer processing unit, and the error signal can be expressed as

Similarly, the weight between the hidden layer and the input layer is updated to

Using (1.3)-(1.7)

Reuse (1.10), can be rewritten (1.12)

(1.13) is the error signal of the hidden layer processing unit, and the error signal can be expressed as

In addition, in order to make the network have a faster convergence speed, the weight update formula between the layers of the network can introduce a momentum term. In addition to allowing the network to respond to local gradient changes, momentum also responds to the latest trends in error surfaces.

3. Artificial immune algorithm

The algorithm flow of the artificial immune algorithm corresponding to the extension of the human immune algorithm is shown in Figure 5.

The steps of the artificial immune algorithm shown in Figure 5 are described as follows: Step 1. Antigen Definition 31:

Define the issues to be optimized in real life. That is, a fitness function that defines the real problem to be optimized. This step contains the encoding and decoding methods that determine the solution. The way the coding is performed determines the expression of the antibody and its length.

Step 2. Initial (g=0) antibody population 32:

The pop_size initial antibodies (collectively referred to as the initial antibody population) are generated in a random manner according to the coding format determined in the antigen definition step. Each of these antibody populations represents a solution to a real problem. The artificial immune algorithm will generate the best feasible solution through these randomly generated initial solutions.

Step 3. Antibody adaptation calculation 33:

Each antibody (solution) in the antibody group is substituted into an adaptive function to perform an antibody adaptation degree calculation. That is, the affinity of each antibody for the antigen is calculated. Usually the adaptive function values can be mapped or directly optimized. The adaptation functions defined will vary depending on the problem.

Step 4. Cell Decision 34:

The degree of adaptation of each antibody is resolved and the antibodies are sorted in a sorted manner. That is, an antibody having a higher antigen affinity is separated into a plasma cell and a memory cell.

Antibodies with high antibody affinity for antigen will be separated into plasma cells from the antibody population (the number of plasma cells pla_size = distribution ratio const1 * pop_size). Better antibodies in the plasma cells are then replicated into memory cells (memory cell number mem_size = distribution ratio const2*pla_size). In this step, the memory cell antibody is currently an antibody with higher antibody affinity to the antigen (i.e., the preferred solution found so far). In each generation, memory cells are antibodies that are highly antibody-affinitive to antigen, and memory cells can reduce the time to search for the best solution. On the one hand, the artificial immune algorithm uses this solution as the tentative optimal solution. On the other hand, it uses this tentative optimal solution to find out whether there is a better solution in the solution space. That is, the artificial immune algorithm is based on the deep search of the antibody for antigen affinity by the low cell antibody instead of the memory cell. The number of plasma cells and the number of memory cells are similar to the plasma cells and memory cells proliferated by the colony selection of the human immune algorithm. Through constant evolution, antibodies can completely destroy antigens.

Step 5. Antibody affinity calculation 35:

The antibody and the remaining antibody (the number of antibodies aff_size=pop_size-pla_size) remaining after the separation of the plasma cells and the plasma cells are subjected to antibody affinity calculation of the antibody to the antibody. The cells produced by this mechanism are called suppressor cells (the number of suppressor cells is sup_size=pop_size*const2-pla_size*const2).

In practical problems, the suppression of cells is like a dissimilar information to each other. This mechanism is a very unique mechanism for artificial immune algorithms. The purpose of calculating antibody affinity for antibodies is to discriminate the degree of agreement of antibodies between the two antibodies. This filtering mechanism first calculates the information entropy and average entropy, and then maps the average entropy to the interval [0, 1] to obtain the affinity value. The artificial immunization algorithm can also maintain a quantitative antibody against antigen affinity by a low antibody in addition to an antibody having a high antibody-to-antibody affinity in the antibody population by the obtained affinity value. The calculation formulas of information entropy and average entropy are shown in (1.15) and (1.16), respectively, and the affinity values can be calculated using (1.17). In summary, through this screening mechanism, artificial immune algorithms maintain a certain degree of spur between antibodies. In this way, artificial immune algorithms can do a variety of breadth search. This breadth of searchability is feasible to avoid the artificial immune algorithm to specifically bias towards certain solution combinations in the process of searching for the best solution, which causes the algorithm to prematurely converge (falling into the regional optimal solution), which can also avoid the infeasibility. The search for solutions reduces the time it takes for the algorithm to search for the best solution.

Wherein: j represents the j-th antibody information bit of the two antibodies, the binary code N=2, k is the number of different characters contained in the two antibodies, and P _{i , j} represents the similarity index of the j-th antibody information bit of the two antibodies .

Where: q represents the length of the antibody information bit of the two antibodies.

Where: v and w represent two antibodies that are calculated for affinity values.

Of course, the artificial immune algorithm also has a mechanism of concentration concept to combine the antibody to the antibody to the antigen affinity of the antibody to regulate the proliferation and inhibition of the antibody. In contrast, the flow of the immune algorithm is also different.

Step 6. Antibody reproduction, crossover, and mutation 36: The second step is 6.1 Antibody replication

The plasma cells and the suppressor cells perform antibody replication operations in accordance with a replication strategy such as a roulette wheel method in an antibody-adapted manner. Replicating antibodies in a roulette method treats each antibody in the antibody population as a block on the disc. As shown in the schematic diagram of the six-wheel method, the size of the block in the figure is proportional to the degree of adaptation of the antibody. That is, the meaning of the pk of (1.18) means that the higher the degree of adaptation of the antibody, the larger the area of the block. The higher the degree of adaptation of the antibody, the greater the chance of being selected into the mating pool. However, using the roulette method, it is necessary to notice whether there is a superantibody production. The production of superantibodies will cause premature convergence of artificial immune algorithms. In the artificial immune algorithm, the plasma cells and the suppressing cells respectively rotate the wheel pla_size times and sup_size times. That is, a total of pla_size+sup_size antibodies in the mating pool. The antibodies in these mating pools will be subjected to the mating operation of the antibodies.

Where: qk represents the cumulative probability of the kth antibody vk in the antibody population, , eval ( v _k )= f ( x ^k ) and f (.) denote the defined adaptation function.

6.2 Antibody mating

The mating operation of the antibody is performed by the antibody in the mating cell. The progeny antibodies produced by this operation represent the solutions in the newly searched solution space. Single point mating of antibodies is shown in the Figure 7 antibody mating schematic.

6.3 Antibody mutation

The mating antibody is subjected to a mutation operation of the antibody. If the antibody only performs the mating operation, the artificial immune algorithm is not able to completely search for the solution space. That is, when the artificial immune algorithm introduces a mutation operation of the antibody, the algorithm will have an opportunity to perform a "jumping" search in the solution space. This hopping search avoids the immune algorithm to obtain the best solution for the region. Single point mutations of antibodies are shown in the schematic diagram of the eight-body mutation.

In summary, the artificial immune algorithm can add new antibodies to expand the new feasible solution region by the operation of step 6: recombining the antibody population using the replication operation; generating a new solution using the mating operation; and applying the mutation operation to escape the solution space of the existing search.

Step 7. Second generation (g=g+1) antibody group 37:

At this point, the pla_size+sup_size antibodies in the initial antibody population have been subjected to the above steps to generate a new search antibody solution. This step couples the resulting new antibody to the mem_size memory cells. That is, the number of secondary antibody populations will be equal to the number of initial antibody populations (pop_size), and the secondary antibody population will also be evolved over and over until the termination conditions are established.

Step 8. Destroy the antigen 38:

When the termination condition is established, the artificial immune algorithm proposes an optimized result. This step is equivalent to the production of antibodies by human lymphoid tissue to destroy the antigen. Table 1.1 shows a comparison table between the artificial immune algorithm and the actual problem.

The various steps of the adaptive non-intrusive load monitoring device are described below.

Step 1: Pre-data collection

The main purpose of data pre-processing is to obtain the measured load power data, and the obtained data is filtered and processed into clean data through a filter. The filtered data will be provided to the monitoring device for subsequent analysis. The steps of "waveform recording" and "filtering" in this step are described below.

1.1. Waveform capture

The measured total power waveform (including voltage and current signals, but only for current signals in the present invention) is recorded by the microprocessor.

1.2. Filtering

The measured total power waveform is filtered via a low pass filter. Filter the program to eliminate load power data noise.

Step 2: Event detection and load current pattern capture

The filtered total current waveform can be calculated by using the formula (1.19), and the change value of current intensity can be calculated. When the change value is greater than a predetermined threshold value α, the monitoring device determines that an event with load start occurs. When the monitoring device determines that a load-starting event occurs, the filtered total current waveform is differentially processed based on the parallel overlap characteristics of the circuit. The waveform of this differential amount can be calculated by using the formula (1.20) to change the rate of current intensity. If the rate of change is greater than a predetermined threshold γ (and it is established within δ cycles), the monitoring device performs a capture of the load transient current pattern. Therefore, when a load start event is detected, the monitoring device can capture the transient current pattern of the load by using the (1.20) equation and the capture rule. The focus of the load transient current pattern is that when the load current changes from transient to steady state, the rate of change of current intensity tends to a stable minimum. Ideally, the value is zero.

Δ I _intensity =( I _intensity ) _{k +1} -( I _intensity ) _k (1.19)

among them: i(j) represents the load current at the jth current sampling point in each cycle, N is the total number of sampling points, mean(i) is the average of the load current, and k is the kth period.

Δ I ^' _intensity = ( I ^' _intensity ) _{k +1} -( I ^' _intensity ) _k (1.20)

among them: And ε is the adjustment parameter.

Similarly, the filtered total current waveform can be calculated by using the equation (1.19). When the change value is less than a predetermined threshold value -β, the monitoring device detects that the load stop event occurs. When the monitoring device detects that the load stop event occurs, the monitoring device performs the difference processing and extraction of the total current waveform filtered before and after the load is stopped. This is the so-called load stop current style capture. The steps of event detection and current capture are shown in Figure 1.

The event detection and load current pattern capture of the monitoring device is based on time domain analysis. In addition, the values of the parameters α and β of the monitoring device are easily determined compared to other parameters.

Step 3: Feature Identification

The quality of the feature extraction and the number of features can have a great impact on the result of the identifier recognition load. Excessive features not only increase the amount of calculation of the recognizer but also affect the speed of recognition. It also causes errors in recognition due to the inclusion of redundant features. The number of features extracted by feature extraction represents the dimension of the feature space. In the present invention, the extracted load current pattern will be subjected to feature extraction. In addition, the features extracted from the captured transient current pattern will be used to identify the start of the load; the features extracted from the captured stop current pattern will be used to identify the stop of the load.

In terms of identifying the start of the load, the extracted load transient current patterns are converted to the maximum value (Ipeak), the average value (average value, Iavg), and the root mean square value using (1.21)-(1.23), respectively. , Irms). The time (Et) experienced by these converted eigenvalues and load current transients is used as the input to the load recognizer. In addition, the four numerical features obtained are also based on time domain analysis.

In terms of identifying the stop of the load, the captured load stop current pattern is also converted to a numerical characteristic using the (1.21)-(1.23) equation as the input of the load recognizer.

The main purpose of the load identification process of the monitoring device is to construct an identifier to identify the start/stop of the load. The recognizer can identify the category of the unknown category vector based on a data set of a known category. The invention adopts the k-nearest neighbor rule and the inverse transfer type neural network proposed by the pattern recognition theory as a load identifier implementation method of the monitoring device.

In the foregoing step 2, for a robust non-intrusive load monitoring device, the monitoring device has an adaptive capability when the monitoring device monitors different types of loads (and the type of load being monitored increases). As shown in Figure 1, even if the monitoring device is monitoring different types of (and newly added unmonitored) loads, the parameters γ, δ, and ε of the monitoring device as illustrated in Figure 1 and Equation (1.20) are artificially immunized. The optimum value can be automatically set under the goal of lifting the device identification rate. The artificial immune algorithm flow used in the present invention is shown in FIG.

In an adaptive non-intrusive load monitoring device, the captured load current pattern (in the case of a certain γ, δ, and ε parameter value) is converted into a numerical feature by feature extraction to become a sample (vector) distributed in the feature space. ). If the samples distributed in the feature space are in a coupled state, the monitoring performance of the monitoring device will decrease. Conversely, if the samples distributed in the feature space are able to present a cluster distribution as much as possible, the monitoring performance of the monitoring device will be improved.

Therefore, the optimization of the parameter value of the monitoring device means finding the optimal parameter value so that the samples distributed in the feature space are as clustered as possible. In this way, the monitoring performance of the monitoring device is further improved. Therefore, the parameter optimization problem for the monitoring device is equivalent to the distribution form of the sample in the feature space.

Therefore, the above-mentioned parameter values of the monitoring device can be adaptively set to an optimal value by the optimization strategy in improving the monitoring performance target of the monitoring device.

In the artificial immune algorithm, the real problem is defined as the adaptive function to be optimized. Through the foregoing description, the present invention uses the weighted sum form of the Fisher criterion as an adaptive function of the artificial immune algorithm. That is, the artificial immune algorithm calculates the affinity of each antibody for the antigen based on the formulas (1.24)-(1.26).

(1.24) is the Fisher's criterion for each dimension (feature axis) of categories i and j in the feature space. The main spirit of the Fisher specification is to maximize the sample distance between different categories on each feature axis, while the distance between samples of the same category is as small as possible. For the Fisher class of the p-category (p>2), it can be decomposed into k= Two categories of Fisher's guidelines. For example, when p=4, the four categories of Fisher's criteria are broken down into six two-two categories of Fisher's criteria.

Where: s represents the sample, the numerator is the distance between category i and category j, the denominator is the variance of category i and category j, n is the number of samples of category i, m is the number of samples of category j, and 1 Represents the dimension of the feature space.

(1.25) is the expression of the degree of adaptation of each dimension in the feature space.

Where: f represents the vector of k two categories of Fisher's criterion, and ρ is the empirical constant.

In order to improve the worst case distribution of the artificial immune algorithm in the evolution process, (1.25) takes the minimum value of f. The average value of this formula f is the overall sample distribution considering the p category. In addition, the constant ρ of the formula (1.25) allows the fitness value of each dimension to appropriately accept the contribution from the average value of f, and it can also avoid too large an average value causing the degree of adaptation of each dimension to be excessively affected. .

Finally, the (1.26) formula integrates the fitness values of each dimension in the feature space into a single fitness value in a weighted sum manner. The integrated fitness values are used to assess the affinity of each antibody for the antigen for the artificial immune algorithm.

Where: w ¹ + w ² +...+ w ^M =1, and fitness ^l represents the fitness value of each dimension.

Through the k two categories of Fisher's criteria, samples of each category will be clustered on each feature axis. Indirectly, the samples of each category are also clustered in the feature space (ie, in the form of a distribution of clusters).

The antibody coding format of the artificial immune algorithm used in the present invention is binary. This coding form maps real variables to a binary string, and the length of the string can be determined according to the required accuracy. Suppose the domain of a variable xj is between [aj,bj] (and the accuracy required after mapping to the fourth decimal place), the domain of this variable should be at least split into ( b _j - a _j ) ×10 ⁴ size range. The number of bits mj required for a real variable encoding can be obtained by using the equation (1.27).

In the decoding phase, the artificial immune algorithm can directly restore a binary string to the corresponding real number by using (1.28).

Where: decimal ( substring _j ) represents the decimal value of the substring _j of the real variable x _j .

Please refer to FIG. 9 , which is a schematic diagram of a physical device of an overall power and monitoring device according to the present invention; the adaptive non-intrusive load monitoring device 4 of the present invention first uses a voltage and current sensor 41 to supply power to the load terminal 6 (each load 61 ). Signal measurement is performed at the office. Then, the measured signal is signal-converted and filtered by the signal converter 42 and the filter 43, and then transmitted to the microprocessor 44 at the back end for related numerical calculation processing (feature extraction, load identifier, manual Immunization algorithm).

In the actual measurement one and two, the monitoring device is used to identify the start and stop of the actual load. The load recognizer implementation is the k-nearest neighbor rule and the inverse transfer class neural network method. The types of loads monitored include Fans, Fluorescent Lights, and Radios.

Referring to FIG. 9, the main function of the power supply 51 (electric power supply terminal 5) is to give a DC voltage of 5 volts. The measuring device of the total power waveform measured at the power supply inlet of the load terminal 6 is a voltage and current sensor 41. In addition, although the actual measured and collected total power waveform of the present invention includes voltage and current, the monitoring device only focuses on the total current waveform in the analysis of the signal.

Referring to FIG. 9, in the actual measurement of the present invention, the high-frequency cutoff frequency of the low-pass filter of the filter 43 is set to 500 Hz.

Measured one: load start identification

The identification description of the measured load start identification, the transient current pattern extraction result and the identification result are analyzed as follows.

1.1. Identification instructions

The type of load identified by the monitoring device at this measurement is an electric fan, a provincial radio lamp, and a radio. Refer to Figure 1 and (1.19) and (1.20) for the steps and formulas used to load the transient current pattern. In terms of parameter setting of the monitoring device, the α value was set to 0.03, the γ value was 0.065, the δ value was 10.0, and the ε value was set to 1.0. The set parameter values are determined by observing all collected load current waveforms. When the load transient current pattern can be successfully captured, the captured patterns are converted into three numerical features using (1.21)-(1.23), respectively. These converted features, along with the time elapsed by the transient phenomenon, are input to the load recognizer. The features extracted from the captured transient current pattern are used to identify the start of the load. Based on this measurement, the sampling speed was 1 μs/Sample, and the waveform was recorded for 0.5 seconds.

According to the actual measurement and identification, the current data of each individual load is measured by 36 (a total of 108), and it is randomly divided into a training set of 69 and a test set of 39. In addition, 41 multiple load current measurements were used as tests.

In terms of parameter setting of the load identifier, the optimal k value of the k-nearest neighbor rule depends entirely on the collected data; the relevant parameters of the inverse transfer type neural network are often determined experimentally. It can be found from Fig. 11 that the recognition results of k values that are too large or too small are not good. Therefore, the k value is set to 7, and the individual load test identification rate of the k-nearest neighbor rule can reach 97.44%. In the parameter setting of the inverted transit type neural network, the network architecture is 4-6-3, the activation function of the hidden layer and the output layer is a logarithmic sigmoid function, and the training algorithm is the steepest slope of the batch with momentum. In the descending method, the learning rate and the momentum coefficient are sequentially set to 0.1 and 0.2, and the error function is the sum squared error function with the target of 1e-5, and the learning algebra is 3000 generations.

1.2. Transient current pattern capture results

The relationship between the load current and its intensity change rate is shown in Figure 12. The rate of change of the load current intensity in Figure 12 can be calculated using the equation (1.20).

The key focus of the load transient current pattern capture is the period of transient duration. When a load is initiated, the current transient will remain for a specific period of time. After this period of time, the load current will enter a stable state.

As shown in Figure 13 and Figure 14, the load transient current pattern can be effectively captured (the waveform in the left half of the figure is before the capture; the waveform on the right half is captured).

The feature space distribution of the load is shown in Figure 15 and Figure 16. The reproducibility and uniqueness of each load transient current and the enantiomorphic relationship of clustering in the feature space can also be clearly seen from these figures. These innate features can assist the load recognizer in identifying the load.

1.3. Analysis of identification results

Figure 17 shows the training convergence results of the measured and inverted neural network. The training and test set of the inverse transfer-like neural network is normalized to [-1, 1]. The training time and sum square error of the inverted transmission network are about 9.89 seconds and 0.06, respectively. The test error for individual loads is approximately 1.35. Tables 1.2 and 1.3 are the identification results of the load identifier k-nearest neighbor rule and the inverse transit type neural network. It can be seen from Table 1.2 and Table 1.3 that the total test identification rates of the two load identifiers of the monitoring device are 95% and 98.75%, respectively. The left side of the confusion matrix CM(i,j) table represents the data of the i-th class is classified into the j-th class (unit: pen); the right side is the corresponding recognition rate (unit: %).

It can be known from Table 1.2 that the identification rate of the electric fan is 100% (the load identifier does not have any misjudgment); the recognition rate of the provincial radio lamp is 90.32% (the load recognizer misidentifies the three provincial radio lamp data into radio) The radio identification rate is 95.24% (the load recognizer misjudges one radio data into a provincial radio light). Overall, the total test identification rate of the k-nearest neighbor rule is about 95%, and the k-nearest neighbor rule is more likely to be confused when identifying provincial radio lights and radios. Therefore, if the k-nearest neighbor rule wants to improve the recognition rate of the provincial radio light and the radio, additional power features can be added. Of course, the k-nearest neighbor rule can also be used to transform existing features into more representative ratios. As can be seen from Table 1.3, the total test identification rate of the inverted transmission network can reach 98.75% (the load recognizer only misidentifies one radio data into a provincial radio light.). The measured results show that the identification of the inverted transmission network is better than the identification of the k-nearest neighbor rule.

Actual measurement 2: load stop identification

The identification description of the measured two-load stop identification, the load stop current pattern extraction result and the identification result are analyzed as follows:

2.1. Identification instructions

The beta value is 0.016 in terms of parameter setting of the monitoring device. The set parameter values are determined by observing all collected load current waveforms. When the load stop current pattern is effectively captured, the captured patterns are converted into three numerical features using (1.21)-(1.23), respectively. These converted features are input to the load recognizer. Features extracted from the captured stop current pattern are used to identify the stop identification of the load. Based on this measurement, the sampling speed was 1 μs/Sample, and the waveform was recorded for 0.5 seconds.

In this actual measurement and identification test, the current data of each individual load was measured by 36 (total 108), and it was randomly divided into a training set of 69 and a test set of 39. In addition, 41 multiple load current measurements were used as tests.

In this actual measurement, the k value is set to 2, and the individual load test identification rate of the k-nearest neighbor rule is as high as 100%. In the parameter setting of the inverted transit neural network, the network architecture is 3-5-3, the activation function of the hidden layer and the output layer is a logarithmic sigmoid function, and the training algorithm is the batch steepest slope method with momentum. The learning rate and the momentum coefficient are sequentially set to 0.3 and 0.2, and the error function is the sum squared error function with a target of 1e-5, and the learning algebra is 3000 generations.

2.2. Load stop current style capture results

As shown in Fig. 18 and Fig. 19, the difference current pattern before and after the load is stopped can also be effectively captured (the waveform in the left half of the figure is before the capture; the waveform on the right half is after the capture.) .

2.3. Analysis of identification results

Figure 20 shows the training convergence results of the measured two-inverted neural network. The training and test set of the inverse transfer-like neural network is normalized to [-1, 1]. The training time and sum square error of the inverted transmission network are about 7.828 seconds and 0.004, respectively. The test error for individual loads is approximately 0.006. Comparing Figure 5.19 with Figure 5.16, it can be found that the error of the inverse transmission network of the second measurement converges below the error value in the earlier training algebra, but the error of the inverse transmission network of the measured one needs to be trained until after 1000 generations. Below the error value. Tables 1.4 and 1.5 are the identification results of the load identifier k-nearest neighbor rule and the inverse transit type neural network. It can be seen from Table 1.4 and Table 1.5 that the total test identification rate of the two load identifiers of the monitoring device is 100%.

Table 1.5 Inverted Transfer Neural Network Measured Two Identification Results Table

It can be seen from Table 1.4 and Table 1.5 that the total recognition rate of both load identifiers is as high as 100%. That is, the load recognizer did not have any misjudgment. Regarding this actual measurement, the identification result of the k-nearest neighbor rule is as good as the identification result of the inverted transmission network.

Embodiment 3: Parameter optimization and identification of monitoring device

The description of the parameterization optimization of the measured monitoring device, the optimization result, the description of the load identification, and the load identification result are analyzed as follows.

3.1. Monitoring device parameter optimization instructions

Based on this measurement, the sampling speed was 0.5 ms/Sample during the data collection process of the training and test set, and the recorded waveform was 5.0 seconds. In addition, the monitoring device uses only training data during the optimization of the artificial immune algorithm.

The corresponding load transient current pattern captures in Figure 21 to Figure 28 are the same measured data but the waveform acquisition results are performed under different device parameter values.

Case 1 - Monitoring device parameter values: α, γ, δ, and ε are 0.03, 0.05, 7.0, and 0.93, respectively, please refer to Figure 21 to Figure 24.

Case 2 - Monitoring device parameter values: α, γ, δ, and ε are 0.03, 0.07, 30.0, and 0.83, respectively, please refer to Figure 25 to Figure 28.

The following five points of discussion can be made through observations 21 through 28:

1. In both cases, the transient current pattern of the fan can be effectively captured.

2. Under the condition of the parameter value of Case 1, the transient current pattern of the provincial radio lamp and radio captured is incomplete due to the higher ε value and lower γ value.

3. Under the condition of the parameter value of case 2, the transient current pattern of the provincial radio lamp can be effectively captured due to the lower ε value and the higher γ value. Comparing Case 2 with the radio current pattern captured in Case 1, the styles taken are only different from Figure 5.29 and Figure 5.41.

4. The role of the adjustment parameter ε in the monitoring device is to assist in the setting of the gamma value. The monitoring device can be protected from the measured environment by the set ε value; further, it also contributes to the determination of the gamma value.

5. In both cases, the transient current pattern of the microwave oven being extracted is incomplete due to improper setting of the δ value. If the delta value is not set to an appropriate value, the load current changes from transient to steady state will cause the monitoring device to mistakenly assume that a load start/stop event occurs.

Based on the above discussion, how to determine the optimal parameter values of the monitoring device so that the samples distributed in the feature space can exhibit the distribution form of the cluster as much as possible.

Therefore, when the monitoring device of the present invention monitors different kinds of loads (and the type of load monitored increases), the parameters γ, δ, and ε of the monitoring device can be automatically found by the artificial immune device accompanied by the Fisher standard. Good value. Thus, when the monitoring device is introduced into the fourth monitoring load microwave oven, the monitoring device can be adaptive. The parameter optimization results of the measured three are discussed below.

3.2. Monitoring device parameter optimization results

According to the actual measurement, the parameter encoding form of the monitoring device is in the form of a binary string. The length of the string can be determined by using the formula (1.27). The parameter variable range and code length of the monitoring device are shown in Table 1.6. Table 1.7 shows the relevant parameter setting information of the artificial immune algorithm. The empirical constant ρ of the formula (1.25) is set to 0.07. (1.26) is (0.1, 0.25, 0.15, 0.5). Figure 29 shows the evolutionary trend of the artificial immune algorithm. Through evolution, the best values for monitoring device parameters (γ, δ, ε) are (0.0469, 120.0, 0.9681).

The results of the transient current pattern of each load after the parameters of the monitoring device are optimized are shown in FIG. 30 to FIG. The feature spatial distribution of each feature before and after optimization of the parameters of the monitoring device is shown in Figure 34 to Figure 44.

From Fig. 34 to Fig. 41, it can be found that the sample distribution of the characteristic Et is most obvious before and after the parameter optimization of the monitoring device, but the sample of the characteristic Ipeak is distributed in the parameter of the monitoring device. The difference between before and after optimization is very small. The difference referred to here means that the distance of samples between different classes on each feature axis can be as large as possible, and the distance between samples of the same category can be reduced as much as possible. Similarly, from Fig. 42 to Fig. 44, it can be found that the sample clustering effect of each load in the feature space is very obvious after the parameter optimization of the monitoring device. Therefore, when the monitoring device monitors different kinds of loads (and the type of load monitored increases), the parameter values of the monitoring device are accompanied by the spirit of Fisher's criterion by artificial immune algorithm (replacement by artificial means) The sample distribution is more clustered) is optimized.

3.3. Identification instructions

The types of loads identified by the monitoring device at this measurement are electric fans, provincial radio lamps, radios, and microwave ovens. Refer to Figure 1 and (1.19) and (1.20) for the flow and formula used for load transient current mode extraction. In terms of parameter setting of the monitoring device, the alpha value is equal to 0.03, the gamma value is 0.0469, the δ value is 120.0, and the ε value is set to 0.9681. The set parameter values are determined after the monitoring device is optimized. When the load transient current pattern can be successfully captured, the captured patterns are converted into three numerical features using (1.21)-(1.23), respectively. These converted features, along with the time elapsed by the transient phenomenon, are input to the load recognizer. The features extracted from the captured transient current pattern are used to identify the start of the load.

According to the actual measurement and identification, the current data of each individual load is measured by 36 (a total of 144), and it is randomly divided into a training set of 92 and a test set of 52. In addition, 42 multiple load current measurements were used as tests.

In terms of parameter setting of the load identifier, the optimal k value of the k-nearest neighbor rule depends entirely on the collected data; the relevant parameters of the inverse transfer type neural network are often determined experimentally. It can be found from Fig. 45 that the recognition result of the excessive k value is not good. Therefore, the k value is set to 5, and the individual load test identification rate of the k-nearest neighbor rule can reach 98.08%. In the parameter setting of the inverted transit neural network, the network architecture is 4-5-4, the activation function of the hidden layer and the output layer is a logarithmic sigmoid function, and the training algorithm is the batch steepest slope method with momentum. The learning rate and the momentum coefficient are sequentially set to 0.13 and 0.02, and the error function is the sum squared error function with a target of 0.001, and the learning algebra is 3000 generations. The transient current pattern capture results for multiple loads are shown in Figure 46.

3.4. Analysis of identification results

Figure 47 shows the training convergence results of the measured three-inverted neural network. The training and test set of the inverse transfer-like neural network is normalized to [-1, 1]. The training time and sum square error of the inverted transmission network are about 8.05 seconds and 4.01, respectively. The test error for individual loads is approximately 0.29. Table 1.8 to Table 1.10 are the identification results of the load identifier k-nearest neighbor rule and the inverse transit type neural network. It can be seen from Table 1.8 to Table 1.10 that the total test identification rates of the two load identifiers of the monitoring device are 97.87% and 88.30%, respectively.

As can be seen from Table 1.8, the identification rate of the electric fan and the microwave oven is 100% (the load identifier does not have any misjudgment.) The recognition rate of the provincial radio lamp is 96.67% (the load recognizer misjudges one provincial radio lamp data) Electric fan.); The radio identification rate is 95.49% (the load recognizer misidentifies one radio data into a provincial radio light.). Overall, the total test identification rate for the k-nearest neighbor rule is 97.87%.

It can be known from Table 1.10 that the total test identification rate of the inverted transmission network is 88.30% (the load recognizer misinterprets 10 provincial radio lamp data into radios; the load recognizer misjudges one radio data into provincial radio lights). Therefore, the inverted transmission network is more likely to be confused when it recognizes provincial radio lights and radios.

The measured results show that the identification result of k-nearest neighbor rule is better than that of inverted transfer network.

It can be found from Fig. 45 and Fig. 11 that the k-nearest neighbor rule has a better recognition rate than the actual one (in the case of any k value). It can be seen from Table 1.8 and Table 1.2 that the total test identification rate of the k-nearest neighbor rule is also higher than the total test identification rate of the measured one. In addition, the newly monitored load microwave oven can also be correctly identified. In addition, taking the small (larger) k value of Figure 45 and Figure 11 as an example, the k-nearest neighbor rule will not change the value of k as the measured rate of the measured one. The change was made while maintaining the recognition level of 98.08% (96.15%). As such, these results also mean that the sample clustering of the loads in the feature space is more pronounced after the parameters of the monitoring device are optimized.

When introducing a new monitoring load, the k-nearest neighbor rule recognizer of the monitoring device does not need to be retrained. On the other hand, when the new load is added, the reverse transfer network must be retrained. In terms of parameter setting of the load identifier, the k-nearest neighbor rule only needs to determine the size of the k value, but the reverse transmission network needs to consider many network factors.

Therefore, when the monitoring device monitors different types of loads (and the type of load being monitored increases), the parameter values (monitoring performance) of the monitoring device can be optimized by the artificial immune algorithm along with the Fisher criteria.

11. . . Pre-data collection processing steps

12. . . Event detection and load current pattern capture steps

13. . . Feature identification step

2. . . Artificial nerve cell

twenty one. . . Input weighted product and function of artificial nerve cells

twenty two. . . Activation function of artificial nerve cells

twenty three. . . Binding strength of signal transmission path between biological cells

31. . . Artificial immune algorithm process step 1

32. . . Artificial immune algorithm process step 2

33. . . Artificial immune algorithm process step 3

34. . . Artificial immune algorithm process step 4

35. . . Artificial immune algorithm process step 5

36. . . Artificial immune algorithm process step 6

37. . . Artificial immune algorithm flow step 7

38. . . Artificial immune algorithm flow step 8

4. . . Adaptive non-intrusive load monitoring device

41. . . Voltage and current sensor

42. . . Signal converter

43. . . filter

44. . . microprocessor

5. . . Power supply

51. . . Power Supplier

6. . . Load side

61. . . load

Figure 1 is a block diagram of an adaptive non-intrusive load monitoring device;

Figure 2 is a schematic diagram of the k-nearest neighbor rule identification;

Figure 3 is a schematic diagram of an artificial nerve cell model;

Figure 4 is a diagram of the inverted transmission neural network architecture;

Figure 5 is a flow chart of the artificial immune algorithm;

Figure 6 is a schematic diagram of the roulette method;

Figure 7 is a schematic diagram of single point mating of antibodies;

Figure 8 is a schematic diagram of single point mutation of antibody;

Figure 9 is a schematic diagram of the physical equipment of the overall power and monitoring device;

Figure 10 is a starting current diagram for loading phase angles of different supply voltages;

Figure 11 is a graph showing the relationship between a different k value and the individual load test identification rate;

Figure 12 is a graph showing the relationship between a measured load current and its intensity change rate;

Figure 13 is a graph showing the results of load transient current pattern capture for a different load case;

Figure 14 is a graph showing the results of load transient current pattern capture in a measured multi-load situation;

Figure 15 is a 3-D map of the feature space (Ipeak, Iavg, and Et);

Figure 16 is a feature space 1-D diagram (Irms);

Figure 17 is a convergence diagram of the measured and inverted neural network training;

Figure 18 is a graph showing the results of the actual current stop current pattern of the two individual loads;

Figure 19 is a graph showing the results of the measured two multiple load stop current patterns;

Figure 20 is a convergence diagram of the measured two-inverted neural network training;

Figure 21, 22, 23, and 24 are the actual situation of the electric fan, provincial radio, radio, microwave oven transient current pattern capture (the horizontal line in the lower left represents the threshold) γ);

Figure 25, 26, 27, and 28 are the measured values of the transient current pattern of the electric fan, provincial radio lamp, radio, and microwave oven in the second case (the horizontal line in the lower left diagram represents the threshold value). γ);

Figure 29 shows the evolution trend of the artificial immune algorithm (the evolution time of each generation is about 370.69 seconds);

Figure 30, 31, 32, and 33 are the current state diagrams of the electric fan, provincial radio lamp, radio, and microwave oven after the parameters of the three monitoring devices are optimized;

Figure 34 shows the characteristic space 1-D map of Ipeak optimization parameters (γ=0.05, δ=7.0, and ε=0.93);

Figure 35 shows the characteristic space 1-D map of Ipeak after optimization of the parameters of the monitoring device (γ=0.0469, δ=120.0, and ε=0.9681);

Figure 36 shows the characteristic space 1-D map of Iavg before the parameter optimization of the monitoring device (γ=0.05, δ=7.0, and ε=0.93);

Figure 37 shows the characteristic space 1-D map of Iavg after optimization of the parameters of the monitoring device (γ=0.0469, δ=120.0, and ε=0.9681);

Figure 38 shows the characteristic space 1-D map of Irms before optimization of the parameters of the monitoring device (γ=0.05, δ=7.0, and ε=0.93);

Figure 39 shows the characteristic space 1-D map of Irms after optimization of the parameters of the monitoring device (γ=0.0469, δ=120.0, and ε=0.9681);

Figure 40 shows the characteristic space 1-D map of Et before the parameter optimization of the monitoring device (γ=0.05, δ=7.0, and ε=0.93);

Figure 41 shows the characteristic space 1-D map of Et after optimization of the parameters of the monitoring device (γ=0.0469, δ=120.0, and ε=0.9681);

Figure 42 shows the 3-D map of the characteristic space of Irms, Iavg, and Et before the parameters of the monitoring device are optimized (γ=0.05, δ=7.0, and ε=0.93);

Figure 43 shows the 3-D map of the characteristic space of Iavg, Irms, and Et after the parameters of the monitoring device are optimized (γ=0.0469, δ=120.0, and ε=0.9681);

Figure 44 is a partial enlarged view of Figure 43;

Figure 45 shows the relationship between the three different k values and the individual load test identification rates.

Figure 46 shows the result of the multi-load transient current pattern capture;

Figure 47 shows the convergence diagram of the measured three-inverted neural network training;

11‧‧‧Pre-data collection steps

12‧‧‧Event detection and load current pattern capture steps

13‧‧‧Character identification step

Claims (8)

An adaptive non-intrusive load feature extraction method for artificial intelligence technology, wherein the feature extraction method comprises: pre-processing of data collection, obtaining the measured load power data, and the obtained data is filtered and processed into noise by a filter. For less data, the filtered data will be provided to the monitoring device for subsequent analysis. The analysis step also includes waveform recording and filtering. The waveform recording and the filtering step are described as follows: waveform extraction is performed by micro The processor records the measured total power waveform (only for the current signal is analyzed); filtering, filtering the measured total power waveform through a low pass filter, and filtering the program, To eliminate the load of power data noise; event detection and load current pattern capture, wherein the event detection and load current pattern capture step comprises: filtering the total current waveform through the use of Δ I _intensity = ( I _intensity ) _{k +1} -( I _intensity ) _k calculates the change in current intensity, where i(j) represents the load current in the jth current sampling point in each cycle, N is the total number of sampling points, mean(i) is the average value of the load current, and k is the kth period; when the current intensity When the change value is greater than a preset threshold value α, the monitoring device determines that a load start event occurs, and the filtered total current waveform is subjected to a difference amount processing based on the parallel overlap characteristic of the circuit, and the waveform of the difference amount is processed by using Formula Δ I ' _intensity =|( I ' _intensity ) _{k +1} -( I ' _intensity ) _k | Calculate the rate of change of current intensity, where: And ε is an adjustment parameter. If the rate of change is greater than a predetermined threshold γ and it is established within δ cycles, the monitoring device performs the capture of the load transient current pattern, wherein the parameters γ, δ of the monitoring device ε can be automatically set to an optimal value by the artificial immune algorithm under the target of the lifting device identification rate; when the current intensity change value is less than a preset threshold value -β, the monitoring device determines that the load is stopped. The event occurs, and the total current waveform filtered before and after the load is stopped is subjected to differential processing and extraction; and feature identification. The method for feature extraction according to claim 1, wherein the feature identification step comprises: in terms of starting the identification load, the extracted load transient current patterns are respectively utilized , , Converted into the eigenvalues of the maximum value, the average value and the rms value, the converted eigenvalues and the time E _t experienced by the load current transient are used as the input of the load recognizer; in terms of identifying the stop of the load, the capture The load stop current pattern is also utilized separately , , Converted to a numerical feature as an input to the load recognizer. An adaptive non-intrusive load feature extraction method of the artificial intelligence technology described in claim 1, wherein the event detection and load current pattern extraction step is The artificial immune algorithm steps include: antigen definition: an adaptation function defining a real problem to be optimized; this step includes determining the encoding and decoding mode of the solution, and the encoding and decoding method will perform the performance of the optimal solution of the optimization function. The form and its length are determined; the initial antibody population: the initial antibody population is generated in a random manner according to the coding format determined in the antigen definition step, and each antibody in the antibody population represents a solution to the real problem; antibody adaptation calculation: Each antibody in the antibody population, ie, the solution, is substituted into the adaptation function for antibody adaptation; the defined adaptation function will vary according to different issues; cell decision making; antibody affinity calculation; antibody replication, mating, and mutation; Sub-generation antibody population: coupling the generated new antibody solution to memory cells; and destroying the antigen: when the termination condition is established, the artificial immune algorithm proposes an optimized result. The method for adaptive non-intrusive load feature extraction of the artificial intelligence technology described in claim 2, wherein the load identifier uses the k-nearest neighbor rule or the inverse transfer type neural network method as the load of the monitoring device. Recognizer implementation. The method of adaptive non-invasive borrowing feature extraction according to the artificial intelligence technology described in claim 3, wherein the cell determining step is to distinguish the degree of adaptation of each antibody and classify the antibodies in a sorted manner, that is, , the best solution found so far, separated into plasma cells; better antibodies in the plasma cells, that is, better solution, will be copied again Memory cells; artificial immune algorithms based on memory cells to replace low-antibody antibodies for antigen affinity affinity search. The method for adaptive non-invasive borrowing feature extraction according to the artificial intelligence technology described in claim 3, wherein the remaining antibody and plasma cells are separated into memory cells after the separation step of the antibody affinity calculation step. The antibodies, ie, the information that is not similar to each other, will be subjected to the antibody affinity calculation of the antibody; the cells produced by this mechanism are called inhibitor cells; this screening mechanism first calculates the information entropy and the average entropy, and then The average entropy is mapped to the interval [0, 1] to obtain the affinity value; the information entropy is calculated as , wherein: j represents the j-th antibody information bit of the two antibodies, the binary code N=2, k is the number of different characters contained in the two antibodies, and P _i,j represents the similarity of the j-th antibody information bit of the two antibodies Indicator; and the average entropy calculation formula is , wherein: q represents the antibody information bit length of the two antibodies; and the affinity value can be utilized Calculated, where: v and w represent the two antibodies to which the affinity values were calculated. The method of adaptive non-invasive borrowing characteristic extraction according to the artificial intelligence technology described in claim 3, wherein the antibody replication, mating and mutation steps comprise: antibody replication: the degree of adaptation of plasma cells and inhibitor cells to antibodies The antibody is replicated in cooperation with a replication strategy such as the roulette method; antibody mating: the antibody in the mating pool is subjected to the mating operation of the antibody; the progeny antibody produced by this operation represents a solution in the newly searched solution space; the antibody mutation: Perform a leaping search in the solution space. The method for adaptive non-intrusive load-bearing feature extraction of the artificial intelligence technology described in claim 3, wherein the adaptation function uses the weighted sum form of the Fisher criterion as the adaptation of the artificial immune algorithm in the antigen definition step function.