Non-intrusive industrial load identification method and system considering multi-state features

Abstract

The invention discloses a non-intrusive industrial load identification method and system considering multi-state characteristics. The method comprises the following steps: acquiring voltage, current and power data of all electrical loads in industrial multi-type states; analyzing the power data by using an accumulation and event detection algorithm, capturing a switching event of each electrical load, and obtaining a U-I trajectory diagram of each electrical load according to corresponding voltage and current data; inputting the U-I trajectory diagrams of all the electrical loads into a pre-trained convolutional neural network to obtain a load preliminary classification result and comprehensive performance and accuracy thereof; and for the electrical load of which the comprehensive performance or accuracy of the preliminary classification result does not meet the requirement, screening an optimal feature combination by using an mRMR algorithm, and carrying out feature classification based on an SVM algorithm to obtain a corresponding electrical load classification result. According to the invention, the multi-state load identification accuracy can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load identification, and relates to a non-intrusive industrial load identification method and system considering multi-state characteristics. Background Art

[0002] Non-intrusive load monitoring (NILM) technology aims to break down the total energy consumption of a household into the specific usage of each appliance through single-channel blind source separation (BSS) to help residents understand and reduce energy consumption. Although studies have shown that providing classified information can effectively reduce energy consumption by 5-15%, traditional electricity meters can only display the overall power consumption, requiring machine learning tools to infer specific appliance usage patterns from a single measurement.

[0003] Since energy decomposition is inherently uncertain and is affected by complex factors such as data noise, device power consumption similarity, and simultaneous operation, it has become a research hotspot for artificial intelligence and machine learning applications. Commonly used techniques include factor hidden Markov models (FHMM) and signal processing methods. In recent years, there have been recognition technologies for multi-state electrical appliances, such as a two-stage method that combines power clustering and convolutional neural networks (CNN) to identify UI trajectories. However, when dealing with multi-state electrical appliances such as industrial load equipment, these methods ignore the states with long working cycles and many power consumption events, resulting in recognition accuracy rates lower than the overall average. Since the UI trajectories of each working state of multi-state industrial loads are similar, selecting only one state for modeling is likely to lead to classification confusion and low load recognition accuracy. Therefore, a non-invasive industrial load recognition method and system that considers multi-state characteristics is needed to improve recognition accuracy. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a non-intrusive industrial load identification method and system taking into account multi-state characteristics.

[0005] The present invention adopts the following technical solution.

[0006] A first aspect of the present invention provides a non-intrusive industrial load identification method considering multi-state characteristics, comprising:

[0007] Obtain voltage, current and power data of all electrical loads in various industrial states;

[0008] Analyze the power data using a cumulative and event detection algorithm to capture the switching events of each power load, and obtain a UI trajectory diagram of each power load based on the corresponding voltage and current data;

[0009] The UI trajectory diagrams of all power loads are input into the pre-trained convolutional neural network to obtain the preliminary load classification results and their comprehensive performance and accuracy;

[0010] For power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, the mRMR algorithm is used to screen the optimal feature combination and feature classification is performed based on the SVM algorithm to obtain the corresponding power load classification results.

[0011] Preferably, the step of obtaining the voltage, current and power data of all power loads in the industrial multi-type state further includes: performing missing value filling, smoothing and data normalization preprocessing on the acquired data; wherein the data normalization is implemented by a weighted z-score standardization method, and the formula is:

[0012]

[0013] Among them, X i is the original data; μ w is the weighted average; w i is the weight of the i-th data point; σ w is the square root of the average of the squares of the deviations of all data points from the weighted mean adjusted according to their respective weights; n is the number of data points; X novm The data is standardized by weighted z-score.

[0014] Preferably, the power data is analyzed by using a cumulative and event detection algorithm to capture switching events of each power load, as follows:

[0015] Use a sliding window to track the power data and calculate the cumulative value C of the data points in the sliding window. t :

[0016]

[0017] Where t is the distance between the time step and the rightmost side of the sliding window; W is the sliding window size; o(i) is the weight of the i-th data point, which is decayed by the exponential decay function o(i) = e -αi Calculation; α is the attenuation coefficient; X novm is the standardized power data;

[0018] In C t >T, the switching event of the corresponding power load is captured, and T is the set threshold.

[0019] Preferably, the UI trajectory diagram of each power load is obtained according to the corresponding voltage and current data, and the specific steps are as follows:

[0020] First, the voltage and current data in a steady-state period after the switching moment are extracted, and then mapped to the zero matrix according to the following formula to draw the UI trajectory graph of the power load;

[0021]

[0022] N(u mn ,i mn )=1

[0023] Where: u m (t) and i m (t) are the voltage and current data at time t respectively;

[0024] u mn and i mn are the row and column indices of the n×n zero matrix N respectively;

[0025] The floor symbol.

[0026] Preferably, the convolutional neural network includes an input layer, two convolutional layers, a maximum pooling layer, two convolutional layers, a maximum pooling layer, three convolutional layers, a maximum pooling layer, three convolutional layers, a fully connected layer, a Dropout layer, a fully connected layer, a Dropout layer, a fully connected layer and an output layer connected in sequence.

[0027] Preferably, the calculation formula of the comprehensive performance is:

[0028]

[0029] Among them, F 1 is the comprehensive performance of the preliminary classification results; P and R are the precision and recall of the preliminary classification results respectively; T p is a true positive, which means the number of samples whose true category is positive and whose predicted category is also positive; F p F is a false positive, which means the number of samples whose true category is negative and whose predicted category is positive; N is a false negative, which means the number of samples whose true class is positive but whose predicted class is negative;

[0030] The calculation formula of the accuracy is:

[0031]

[0032] Among them, a is the number of samples classified correctly; A is the total number of samples.

[0033] Preferably, for the power load whose comprehensive performance or accuracy of the preliminary classification result does not meet the requirements, the mRMR algorithm is used to screen the optimal feature combination, including:

[0034] For the power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, obtain their load feature sets, including power features, current features, UI trajectory features, and harmonic features;

[0035] The mRMR algorithm is used to find the optimal feature combination from the load feature set. The process is as follows:

[0036] The data of each feature in the load feature set constitutes a sub-feature set S, and the maximum correlation maxD(S,c) and minimum redundancy minR(S) of the sub-feature set S are calculated; where D(S,c) represents the correlation of the sub-feature set S, R(S) represents the redundancy of the sub-feature set S, and c is the feature category variable;

[0037] The importance comprehensive score is calculated based on the maximum relevance maxD(S,c) and the minimum redundancy: φ = maxD(S,c)-minR(S);

[0038] Select several sub-feature sets S with the largest comprehensive importance scores to form the optimal feature combination.

[0039] Preferably, the power characteristics include active power P and reactive power Q;

[0040] The current characteristics include the current root mean square effective value I, the current span I te ;

[0041] The UI trajectory features include trajectory area Ar, trajectory symmetry score A asy , maximum Y-axis intercept Ymax, minimum Y-axis intercept Ymin;

[0042] The harmonic characteristics include the total harmonic voltage distortion rate UTHD, U3, U5, U7, U9, U11, and the total harmonic current distortion rate ITHD, I3, I5, I7, I9, I11, wherein U3, U5, U7, U9, U11 represent the total distortion rates of the 3rd, 5th, 7th, 9th, and 11th harmonic voltages, and I3, I5, I7, I9, I11 represent the total distortion rates of the 3rd, 5th, 7th, 9th, and 11th harmonic currents.

[0043] Preferably, the trajectory symmetry score A asy The calculation formula is:

[0044]

[0045] Where U i is the voltage value of the ith data point; I i is the current value of the ith data point; is the average value of the voltage; is the average value of the current; N is the total number of data points in the time window.

[0046] Preferably, the calculation formulas of D(S,c) and R(S) are:

[0047]

[0048] Among them, x i is the value of the i-th data point in the sub-feature set S, c is the feature category variable, S is the sub-feature set in the load feature set; p(x i ,c) is x i The joint probability distribution of p(x i ) is x i The marginal probability distribution of is, p(c) is the marginal probability distribution of c; p(x i ,x j ) is x i With x j The joint probability distribution of j ) is x j The marginal probability distribution of |S| is the total number of data points in the sub-feature set S;

[0049] Preferably, the feature classification based on the SVM algorithm includes: inputting the optimal feature combination screened by the mRMR algorithm into a pre-trained SVM classifier for feature classification to obtain a corresponding power load classification result, wherein the SVM classifier uses a one-to-many method to construct k classifiers for pre-training, wherein k is the total number of working states, and the classifier uses a linear kernel function, and the C value of the linear kernel function is 1, and the random state is 0.

[0050] A second aspect of the present invention provides a non-intrusive industrial load identification system considering multi-state characteristics, comprising:

[0051] A data acquisition module is used to obtain the voltage, current and power data of all power loads in various industrial states;

[0052] A data analysis module, used to analyze the power data using a cumulative and event detection algorithm, capture switching events of each power load, and obtain a UI trajectory diagram of each power load based on corresponding voltage and current data;

[0053] The preliminary classification module is used to input the UI trajectory diagrams of all power loads into the pre-trained convolutional neural network to obtain the preliminary classification results of the loads and their comprehensive performance and accuracy;

[0054] The feature classification module is used to select the optimal feature combination using the mRMR algorithm and perform feature classification based on the SVM algorithm for power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, so as to obtain the corresponding power load classification results.

[0055] A third aspect of the present invention provides a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0056] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program implements the steps of the method when executed by a processor.

[0057] Compared with the prior art, the beneficial effects of the present invention include at least:

[0058] The present invention performs weighted Z-Score standardization on the acquired industrial load data, and by introducing a weight mechanism, makes the standardization process of multi-state characteristic data of industrial users more flexible and controllable, and can better adapt to the characteristics and requirements of multi-state data sets.

[0059] The cumulative and event detection algorithm of the present invention uses a sliding window to continuously track the recent activity data of the device, and gives a higher weight to the recent data through weighted accumulation, which can more accurately reflect the current state changes of the equipment, quickly capture the switching events of each power load, obtain the UI trajectory diagram of each power load, improve the detection accuracy, and through dynamic adjustment of weights, enhance the ability to distinguish the importance of data at different time points, making load identification more sensitive and reliable.

[0060] The present invention adopts a two-stage recognition method: in the first stage, the VGG16 convolutional neural network is used to identify the UI trajectory features. In the second stage, the mRMR algorithm is used to filter out feature combinations with high information content and low redundancy, and then the SVM algorithm is used to identify the filtered feature combinations to achieve refined load recognition.

[0061] In the second stage of identification, the mRMR feature selection algorithm is used to screen the types of load characteristics, which greatly reduces the redundancy of load characteristic information and the amount of training calculation, and can effectively improve the identification accuracy.

[0062] In the second stage of identification, feature identification is extracted for each working state of the easily confused electrical appliances in the first stage, which greatly improves the algorithm's recognition effect on multi-state electrical appliances with similar impedance characteristics. It complements the advantages of the first stage of identification, takes into account both accuracy and solution speed, and provides a solution to the problem of easy confusion of multi-state electrical appliances.

[0063] In the second stage of identification, adding the electrical characteristics of the electrical appliances in different working states can show more dimensional information of the load, so as to better distinguish multi-state electrical appliances with similar impedance characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A flow chart of a non-intrusive industrial load identification method considering multi-state characteristics;

[0065] Figure 2 Typical industrial load identification results. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0067] Embodiment 1 of the present invention provides a non-intrusive industrial load identification method considering multi-state characteristics, including:

[0068] S1: Obtain voltage, current and power data of all power loads in industrial multi-type states;

[0069] Further preferably, the data quality and applicability are ensured through the collection and preprocessing of industrial load data, including collecting detailed power consumption data and normalizing preprocessing work;

[0070] Specifically, the actual industrial user load data obtained is preprocessed, including missing value filling, smoothing and data normalization. The data normalization method refers to standardizing the load data using a weighted z-score standardization method.

[0071] In various samples, the quantity level of data is different, which may bring negative effects to the model. In addition, when the traditional least square method is used to process parameter estimation, the results are often unstable. Therefore, the present invention uses weighted z-score to standardize the data. The basic idea is to give each feature or data point a weight, and then adjust the standardization process according to these weights. The formula is as follows:

[0072]

[0073] Among them, X i is the original data; μ w is the weighted average;

[0074] w i is the weight of the ith data point. The weight can be determined based on the importance of the data or other factors. For example, if a data point is considered more reliable, it can be given a higher weight;

[0075] σ w is the square root of the average of the squared deviations of all data points from the weighted mean adjusted according to their respective weights;

[0076] n is the number of data points; X novmThe data is standardized by weighted z-score.

[0077] The above-mentioned weighted Z-Score standardization method introduces a weight mechanism, making the standardization process of multi-state feature data of industrial users more flexible and controllable, and can better adapt to the characteristics and needs of multi-state data sets.

[0078] S2: Analyze the power data using a cumulative and event detection algorithm to capture the switching events of each power load, and obtain a UI trajectory diagram of each power load based on the corresponding voltage and current data;

[0079] Further preferably, a cumulative and event detection algorithm is used to capture the power load switching event and extract the UI trajectory of the load at the switching moment, including:

[0080] S201: Using the cumulative event detection algorithm to capture power load switching events:

[0081] This algorithm aims to detect cumulative events over a period of time while taking into account the timeliness and importance differences of data in industrial production processes. A sliding window will be used to track the recent industrial load data, and weights will be introduced to emphasize the importance of recent data.

[0082] Set the sliding window size W (for example, data of different states of industrial users in the last 7 hours), initialize the cumulative counter C 0 =0, set the threshold T, and trigger the event when the accumulated value exceeds this threshold.

[0083] Input the data X processed in step S1 novm And distribute the data weight within the sliding window. Using the sliding window allows the algorithm to focus on the most recent data, respond to the latest situation in a timely manner, and avoid excessive impact of historical data on the current status.

[0084] For each industrial load data point in the window, a weight is assigned according to its distance from the current time. For example, an exponential decay function o(t) = e -αt , where α is the decay coefficient and t is the distance of the time step from the rightmost side of the window.

[0085] Calculate the cumulative value C using the weighted sum t :

[0086]

[0087] Where o(i) is the weight of the i-th data point.

[0088] The cumulative value C needs to be checked later. t Whether it exceeds the threshold T. If C t>T, it is considered as a switching event trigger, and the time, load level, equipment status and other information of the theft event are recorded in detail to provide support for the subsequent extraction of UI features.

[0089] When a new data point arrives, the window is updated, the earliest state data point is removed, and the latest data point is added.

[0090] The above algorithm uses a sliding window to continuously track the recent activity data of the equipment, and gives a higher weight to the recent data through weighted accumulation, which can more accurately reflect the current state changes of the equipment. This method can quickly respond to changes in the equipment state, improve recognition accuracy, and enhance the ability to distinguish the importance of data at different time points by dynamically adjusting weights, making load identification more sensitive and reliable.

[0091] S202: constructing UI trajectory graphic features by extracting normalized voltage and current data within a steady-state cycle, including:

[0092] (1) Extract the voltage and current data within one cycle after the load switches on and off and enters a stable operating state;

[0093] (2) When the resolution of the UI trajectory image is set to n×n, create an n×n zero matrix N, and map the normalized data to the zero matrix according to the following formula to draw the UI trajectory matrix.

[0094]

[0095] N(u mn ,i mn )=1

[0096] Where: u mn and i mn are the row and column indices of the zero matrix N respectively; The normalized and highly pixelated image is a 25*25 matrix, where white represents a matrix element of 0 and black represents a matrix color of 1.

[0097] Since the row index and column index u mn and i mn is a series of continuous data points, N(u mn ,i mn )=1 will be drawn as a continuous figure.

[0098] S3: Input the UI trajectory diagrams of all power loads into the pre-trained convolutional neural network to obtain the preliminary load classification results and their comprehensive performance and accuracy;

[0099] Further preferably, a convolutional neural network structure is designed to achieve the first stage industrial load classification;

[0100] In the pre-training process, a large-scale annotated power load data set (such as the PLAID data set) is first collected. These data include voltage, current, and power data under various industrial multi-type states. Then, the convolutional neural network (such as VGG16) is trained using these data, and the network parameters are optimized through back propagation and gradient descent to gradually improve the accuracy of the model in identifying the UI trajectory features of different loads. After the pre-training is completed, the UI trajectory map extracted from S2 is input into the pre-trained convolutional neural network to obtain the preliminary classification results of the load and its comprehensive performance and accuracy. For the power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, the mRMR algorithm is subsequently used to screen the optimal feature combination and feature classification is performed based on the SVM algorithm to obtain the final power load classification results.

[0101] Since convolutional neural networks can deeply mine deep feature information in images, and the VGG16 convolutional neural network has outstanding performance in image recognition and classification tasks, the VGG16 model is used in the first stage to perform preliminary classification of different types of devices using the UI trajectory curve as image input.

[0102] The original VGG16 model cannot be directly applied to the classification task of industrial load data sets. The present invention adopts the improved VGG16 model and improves its parameters to adapt to the requirements of load identification. The input image enters the fully connected layer after being transformed by the convolution layer and the pooling layer, in which the convolution layer and the fully connected layer used to calculate the image features appear alternately, and can be divided into different sections with the pooling layer as the boundary; the model has 3 fully connected layers for classification, the first two fully connected layers have 4096 neurons, and the number of vectors generated by the last fully connected layer is equal to the number of electrical appliance categories to be identified, which is 11. In order to improve the model training speed and prevent the model from overfitting, a Dropout layer is added between each fully connected layer, and the discard ratio is 0.5. The number of training times is 400, and the learning rate is 0.0001.

[0103] The convolutional neural network model structure is shown in Table 1, and the operation mode is from top to bottom.

[0104] Table 1 Convolutional neural network model structure

[0105]

[0106]

[0107] Furthermore, the present invention uses confusion matrix, F 1 Score and AccuracyA accAs an evaluation indicator of the algorithm, the numbers in the blue box on the diagonal of the confusion matrix represent the number of samples that are correctly identified, while the numbers in the non-diagonal box represent the number of samples that are misidentified.

[0108] F 1 The score can reflect the overall performance of the algorithm and is the harmonic mean of precision P and recall R. Accuracy is the ratio of the number of correctly identified samples to the total number of samples in the test set. Its calculation formula is as follows:

[0109]

[0110] Where: T p is the true positive, which means the number of samples whose true category is positive and whose predicted category is also positive;

[0111] F p is a false positive, which means the number of samples where the true class is a negative example but the predicted class is a positive example;

[0112] F N is a false negative, which means the number of samples whose true class is positive but whose predicted class is negative;

[0113] a is the number of samples classified correctly, and A is the total number of samples in the test set.

[0114] In specific implementation, after obtaining the preliminary load classification result, manual confirmation is performed to confirm whether the corresponding load real category and the preliminary load classification result are correct.

[0115] S4: For power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, the mRMR algorithm is used to screen the optimal feature combination and feature classification is performed based on the SVM algorithm to obtain the corresponding power load classification results.

[0116] Further preferably, the mRMR algorithm is used to screen the industrial load data features and the feature classification is performed based on the SVM algorithm, including:

[0117] S401: For the power load whose comprehensive performance or accuracy of the preliminary classification result does not meet the requirements, the mRMR algorithm is used to select the optimal feature combination, the steps are as follows:

[0118] Under the premise that the number of identification features is limited, selecting representative features with the largest amount of information and the smallest redundancy as the basis for load identification can effectively improve the recognition accuracy, increase the robustness of the algorithm, and speed up the calculation speed.

[0119] The mRMR feature selection algorithm can calculate the maximum correlation between features and minimize the redundancy between features, so as to find a set of features in the original feature set that has the greatest correlation with the final output result (the output result of the algorithm, the recognition result after the last algorithm execution can be taken as the final output result here) but has the smallest correlation between features, as the optimal feature combination. The specific process is as follows:

[0120] Define two random variables x and y, whose probability density functions are p(x), p(y), and p(x,y), then the mutual information is

[0121]

[0122] Where p(x,y) is the joint probability distribution of X and Y, p(x) is the marginal probability distribution of X, and p(y) is the marginal probability distribution of Y;

[0123] Calculate the maximum relevance and minimum redundancy based on the above definitions:

[0124]

[0125] Among them, x i is the value of the i-th data point in the sub-feature set S, c is the feature category variable, and S is the sub-feature set in the load feature set;

[0126] The above-mentioned category variables represent the feature categories during load identification, such as c=1 represents the first category of features, and c=2 represents the second category of features; the sub-feature set represented by S refers to the feature sets represented by the power features, current features, UI trajectory features, and harmonic features mentioned above. For example, the power features include active power and reactive power. The data of one category of features constitutes a corresponding sub-feature set;

[0127] maxD(S,c) is the maximum correlation, which means the maximum value in D(S,c) corresponding to different sub-feature sets S;

[0128] minR(S) is the minimum redundancy, which means the minimum value in R(S) corresponding to different sub-feature sets S;

[0129] In the above formula (3), S(x i ; c), S(x i ;x j ) are x i The mutual information with c, x i With x j The mutual information of is similar to that in formula (1) (2) Part, S(x i ; c), S(x i ;xj ) is calculated as:

[0130]

[0131] In the formula, p(x i ,c) is x i The joint probability distribution of p(x i ) is x i The marginal probability distribution of is, p(c) is the marginal probability distribution of c; p(x i ,x j ) is x i With x j The joint probability distribution of j ) is x j The marginal probability distribution of

[0132] Substituting formula (5) into formula (3) The calculation formula of D(S,c) is as follows:

[0133]

[0134] Substituting formula (6) into formula (4) The calculation formula of R(S) is as follows:

[0135]

[0136] Finally, calculate the overall importance score:

[0137] φ=maxD(S,c)-minR(S)

[0138] In the above mRMR algorithm, the maximum correlation maxD(S,c) measures the maximum correlation between the sub-feature set S and the target variable c by calculating the mutual information I(Fi;c) between each feature and the target variable and selecting the maximum value; the minimum redundancy min(S) measures the minimum redundancy between the features in the sub-feature set S; in the additive integration importance index maxφ(D,R)=DR (maxφ(D,R) represents the sub-feature set S selected from all the features in the load feature set that maximizes φ=DR, forming the optimal feature combination), D is replaced by maxD(S,c) and R is replaced by min(S), and the corresponding φ(D,R)=maxD(S,c)-min(S) is calculated; features with high comprehensive scores are selected to join the optimal feature combination, thereby ensuring that the selected features are highly correlated with the target variable and have low redundancy with the features in the selected feature set.

[0139] The following types of load characteristics are commonly used in load identification algorithms:

[0140] (1) Power: Active power P, reactive power Q.

[0141] (2) Current. Current RMS value I, current span I te , where the current span calculation formula is as follows; I te =I max -I min , where I max is the maximum value of the current collected, I min is the minimum value of the current in the collected data.

[0142] (3) UI trajectory feature class. Trajectory area Ar, trajectory symmetry score A asy , Y-intercept Y max , Y min .

[0143]

[0144] Where U i is the voltage value at the i-th time point; I i is the current value at the i-th time point; is the average value of the voltage; is the average value of the current; N is the total number of data points in the time window; A asy is the trajectory symmetry score, and its value range is [-1,1]. The closer the value is to 1, the more symmetrical the trajectory is.

[0145] (4) Harmonics. The total voltage distortion rate is UTHD, U3, U5, U7, U9, U11, and the total current distortion rate is ITHD, I3, I5, I7, I9, I11, etc., where U3 represents the total distortion rate of the third harmonic voltage, U5 represents the total distortion rate of the fifth harmonic voltage, U7 represents the total distortion rate of the seventh harmonic voltage, U9 represents the total distortion rate of the ninth harmonic voltage, U11 represents the total distortion rate of the eleventh harmonic voltage, I3 represents the total distortion rate of the third harmonic current, I5 represents the total distortion rate of the fifth harmonic current, I7 represents the total distortion rate of the seventh harmonic current, I9 represents the total distortion rate of the ninth harmonic current, and I11 represents the total distortion rate of the eleventh harmonic current.

[0146] The above 20 common features are used as input, and the mRMR algorithm is used to calculate their importance index φ to obtain the importance ranking of the group of features. The larger the value, the better the distinction between classes and the lower the correlation between features. The features ranked in the top i can be selected to form the optimal feature combination.

[0147] S402: For the optimal feature combination screened by the mRMR algorithm, feature classification is performed based on the SVM algorithm to obtain the corresponding power load classification result. The specific steps are as follows:

[0148] The principle of the traditional SVM algorithm is to find an optimal classification hyperplane that meets the classification requirements and maximizes the blank areas on both sides of the hyperplane.

[0149] In the present invention, since industrial load has multi-state characteristics, multi-classification needs to be implemented. In order to enable SVM to be applied to the multi-classification scenario of industrial load, it is necessary to construct a suitable multi-classifier.

[0150] The present invention constructs a classifier based on the indirect method. The indirect method handles multi-classification problems by combining multiple binary classifiers. Common methods include one-to-many method (OVR) and one-to-one method (OVO). Compared with the OVO method, the OVR method requires fewer classifiers to be constructed and has a faster training speed. In combination with the recognition content of the present invention, since the second-stage algorithm needs to consider multiple working states and model them, a one-to-many method is used to construct a classifier. For example, a total of 14 classifiers need to be constructed to recognize 14 working states, so as to improve the training speed and recognition accuracy of the algorithm, reduce training time, and improve recognition efficiency.

[0151] After extracting the feature vectors of multiple working states of industrial load equipment that were not accurately identified in the first stage according to the method of S401, they are used as the input of the SVM classifier in the second stage of identification, and the selected features have higher discrimination. Among them, the SVM classifier uses a one-to-many method to construct 14 classifier training, using a linear kernel function, C value 1, and random state 0. During training, after adjusting the relevant parameters, the feature data of all industrial loads are divided into training sets and test sets according to 7:3.

[0152] Embodiment 2 of the present invention provides a non-intrusive industrial load identification system considering multi-state characteristics, including:

[0153] A data acquisition module is used to obtain the voltage, current and power data of all power loads in various industrial states;

[0154] A data analysis module, used to analyze the power data using a cumulative and event detection algorithm, capture switching events of each power load, and obtain a UI trajectory diagram of each power load based on corresponding voltage and current data;

[0155] The preliminary classification module is used to input the UI trajectory diagrams of all power loads into the pre-trained convolutional neural network to obtain the preliminary classification results of the loads and their comprehensive performance and accuracy;

[0156] The feature classification module is used to select the optimal feature combination using the mRMR algorithm and perform feature classification based on the SVM algorithm for power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, so as to obtain the corresponding power load classification results.

[0157] Embodiment 3 of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0158] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0159] Compared with the prior art, the beneficial effects of the present invention include at least:

[0160] The present invention performs weighted Z-Score standardization on the acquired industrial load data, and by introducing a weight mechanism, makes the standardization process of multi-state characteristic data of industrial users more flexible and controllable, and can better adapt to the characteristics and requirements of multi-state data sets.

[0161] The cumulative and event detection algorithm of the present invention uses a sliding window to continuously track the recent activity data of the device, and gives a higher weight to the recent data through weighted accumulation, which can more accurately reflect the current state changes of the equipment, quickly capture the switching events of each power load, obtain the UI trajectory diagram of each power load, improve the detection accuracy, and through dynamic adjustment of weights, enhance the ability to distinguish the importance of data at different time points, making load identification more sensitive and reliable.

[0162] The present invention adopts a two-stage recognition method: in the first stage, the VGG16 convolutional neural network is used to identify the UI trajectory features. In the second stage, the mRMR algorithm is used to filter out feature combinations with high information content and low redundancy, and then the SVM algorithm is used to identify the filtered feature combinations to achieve refined load recognition.

[0163] In the second stage of identification, the mRMR feature selection algorithm is used to screen the types of load characteristics, which greatly reduces the redundancy of load characteristic information and the amount of training calculation, and can effectively improve the identification accuracy.

[0164] In the second stage of identification, feature identification is extracted for each working state of the easily confused electrical appliances in the first stage, which greatly improves the algorithm's recognition effect on multi-state electrical appliances with similar impedance characteristics. It complements the advantages of the first stage of identification, takes into account both accuracy and solution speed, and provides a solution to the problem of easy confusion of multi-state electrical appliances.

[0165] In the second stage of identification, adding the electrical characteristics of the electrical appliances in different working states can show more dimensional information of the load, so as to better distinguish multi-state electrical appliances with similar impedance characteristics.

[0166] Take the cement homogenizer in typical industrial load as an example. Figure 2The load identification result of the cement homogenizer, where the black is the actual power of the cement homogenizer and the gray is the identified load power. It can be seen from the results that the non-invasive industrial load identification method proposed in the present invention can accurately identify the switch working state and power status.

[0167] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0168] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0169] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0170] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages-such as Smalltalk, C++, etc., and conventional procedural programming languages-such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network-including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) may be personalized by utilizing the state information of a computer-readable program instruction, and the electronic circuit may execute a computer-readable program instruction, thereby realizing various aspects of the present disclosure.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A non-intrusive industrial load identification method considering multi-state characteristics, characterized in that: include: Obtain voltage, current and power data of all electrical loads in various industrial states; Analyze the power data using a cumulative and event detection algorithm to capture the switching events of each power load, and obtain a UI trajectory diagram of each power load based on the corresponding voltage and current data; The UI trajectory diagrams of all power loads are input into the pre-trained convolutional neural network to obtain the preliminary load classification results and their comprehensive performance and accuracy; For power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, the mRMR algorithm is used to screen the optimal feature combination and feature classification is performed based on the SVM algorithm to obtain the corresponding power load classification results.

2. A non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The method of obtaining the voltage, current and power data of all power loads under the industrial multi-type state also includes: performing missing value filling, smoothing and data normalization preprocessing on the obtained data; wherein the data normalization is implemented by a weighted z-score standardization method, and the formula is: Among them, X i is the original data; μ w is the weighted average; w i is the weight of the i-th data point; σ w is the square root of the average of the squares of the deviations of all data points from the weighted mean adjusted according to their respective weights; n is the number of data points; X novm The data is standardized by weighted z-score.

3. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The power data is analyzed by using the cumulative and event detection algorithm to capture the switching events of each power load, as follows: Use a sliding window to track the power data and calculate the cumulative value C of the data points in the sliding window. t : Where t is the distance between the time step and the rightmost side of the sliding window; W is the sliding window size; o(i) is the weight of the i-th data point, which is decayed by the exponential decay function o(i) = e -αi Calculation; α is the attenuation coefficient; X novm is the standardized power data; In C t >T, the switching event of the corresponding power load is captured, and T is the set threshold.

4. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The specific steps of obtaining the UI trajectory diagram of each power load according to the corresponding voltage and current data are as follows: First, the voltage and current data in a steady-state cycle after the switching moment are extracted, and then mapped to the zero matrix according to the following formula to draw the UI trajectory graph of the power load; N(u mn ,i mn )=1 Where: u m (t) and i m (t) are the voltage and current data at time t respectively; u mn and i mn are the row and column indices of the n×n zero matrix N respectively; The floor symbol.

5. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The convolutional neural network includes an input layer, two convolutional layers, a maximum pooling layer, two convolutional layers, a maximum pooling layer, three convolutional layers, a maximum pooling layer, three convolutional layers, a fully connected layer, a Dropout layer, a fully connected layer, a Dropout layer, a fully connected layer and an output layer, which are connected in sequence.

6. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The calculation formula of the comprehensive performance is: Among them, F1 is the comprehensive performance of the preliminary classification results; P and R are the precision and recall of the preliminary classification results respectively; T p is a true positive, which means the number of samples whose true category is positive and whose predicted category is also positive; F p is a false positive, indicating the number of samples where the true category is a negative example but the predicted category is a positive example; F N is a false negative, which means the number of samples whose true class is positive but whose predicted class is negative; The calculation formula of the accuracy is: Among them, a is the number of samples classified correctly; A is the total number of samples.

7. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: For the power load whose comprehensive performance or accuracy of the preliminary classification result does not meet the requirements, the mRMR algorithm is used to screen the optimal feature combination, including: For the power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, obtain their load feature sets, including power features, current features, UI trajectory features, and harmonic features; The mRMR algorithm is used to find the optimal feature combination from the load feature set. The process is as follows: The data of each feature in the load feature set constitutes a sub-feature set S, and the maximum correlation max D(S,c) and minimum redundancy min R(S) of the sub-feature set S are calculated; where D(S,c) represents the correlation of the sub-feature set S, R(S) represents the redundancy of the sub-feature set S, and c is the feature category variable; The importance comprehensive score is calculated based on the maximum relevance max D(S,c) and the minimum redundancy min R(S): φ = max D(S,c)-min R(S); Select several sub-feature sets S with the largest comprehensive importance scores to form the optimal feature combination.

8. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 7, characterized in that: The power characteristics include active power P and reactive power Q; The current characteristics include the current root mean square effective value I, the current span I te ; The UI trajectory features include trajectory area Ar, trajectory symmetry score A asy , maximum Y-axis intercept Ymax, minimum Y-axis intercept Ymin; The harmonic characteristics include the total harmonic voltage distortion rate UTHD, U3, U5, U7, U9, U11, and the total harmonic current distortion rate ITHD, I3, I5, I7, I9, I11, wherein U3, U5, U7, U9, U11 represent the total distortion rates of the 3rd, 5th, 7th, 9th, and 11th harmonic voltages, and I3, I5, I7, I9, I11 represent the total distortion rates of the 3rd, 5th, 7th, 9th, and 11th harmonic currents.

9. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 7, characterized in that: Track symmetry score A asy The calculation formula is: Where U i is the voltage value of the ith data point; I i is the current value of the ith data point; is the average value of the voltage; is the average value of the current; N is the total number of data points in the time window.

10. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 7, characterized in that: The calculation formulas for D(S,c) and R(S) are: Among them, x i is the value of the i-th data point in the sub-feature set S, c is the feature category variable, S is the sub-feature set in the load feature set; p(x i ,c) is x i The joint probability distribution of p(x i ) is x i The marginal probability distribution of is, p(c) is the marginal probability distribution of c; p(x i ,x j ) is x i With x j The joint probability distribution of j ) is x j The marginal probability distribution of ; |S| is the total number of data points in the sub-feature set S.

11. The non-intrusive industrial load identification method considering multi-state characteristics according to claim 1, characterized in that: The feature classification based on the SVM algorithm includes: inputting the optimal feature combination screened by the mRMR algorithm into a pre-trained SVM classifier for feature classification to obtain a corresponding power load classification result, wherein the SVM classifier uses a one-to-many method to construct k classifiers for pre-training, wherein k is the total number of working states, and the classifier uses a linear kernel function.

12. A non-intrusive industrial load identification system considering multi-state characteristics, using the method according to any one of claims 1 to 11, characterized in that: The system comprises: A data acquisition module is used to obtain the voltage, current and power data of all power loads in various industrial states; A data analysis module, used to analyze the power data using a cumulative and event detection algorithm, capture switching events of each power load, and obtain a UI trajectory diagram of each power load based on corresponding voltage and current data; The preliminary classification module is used to input the UI trajectory diagrams of all power loads into the pre-trained convolutional neural network to obtain the preliminary classification results of the loads and their comprehensive performance and accuracy; The feature classification module is used to select the optimal feature combination using the mRMR algorithm and perform feature classification based on the SVM algorithm for power loads whose comprehensive performance or accuracy of the preliminary classification results does not meet the requirements, so as to obtain the corresponding power load classification results.

13. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images