Electric power data identification method and system

Abstract

The invention provides a power data identification method and system, and the method comprises the steps: obtaining the time sequence power utilization data of a target enterprise in a plurality of continuous analysis periods, converting the obtained data into a basic power utilization feature vector, constructing a three-dimensional tensor based on the basic power utilization feature vector of a historical period, and carrying out the Tuck decomposition, the method comprises the following steps of: forming a power utilization association network, calculating a structure potential time sequence of each enterprise node, fitting the structure potential time sequence by adopting a nonlinear autoregression model, defining an absolute value of a prediction residual error as evolution activity of the node, weighting a current period feature vector according to the evolution activity in an identification stage, and identifying the current period feature vector. And reconstructing a power utilization association network and calculating node structure potential, and when the average evolution activity of internal nodes of a certain enterprise cluster exceeds a first threshold value, the average structure potential of the cluster is monotonically increased in the latest N periods, and the total amplification exceeds a second threshold value, identifying the enterprise set corresponding to the cluster as an advanced enterprise unit.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data processing, and particularly relates to a power data identification method and system. BACKGROUND

[0002] Advanced technology enterprises represent the contemporary advanced productive forces born from technological revolutionary breakthroughs, innovative allocation of production factors, and deep transformation and upgrading of industries. Traditional high-growth enterprise identification methods mainly rely on macroeconomic indicators, enterprise financial reports, industry research analysis, or patent application data. Data such as financial reports and annual reports have a long publication cycle and are lagging, making it difficult to detect early signals of industrial development. Qualitative analysis methods based on expert interviews or industry reports are highly subjective and have limited coverage, making it difficult to conduct large-scale evaluations. With the development of energy Internet of Things and smart grid technology, high-frequency, high-dimensional enterprise electricity data provides new possibilities for real-time and objective insight into economic operation. As a "barometer" of enterprise production and operation activities, power data has the unique advantages of strong real-time, objectivity, and wide coverage. Some studies have attempted to use power data for economic analysis, such as predicting macroeconomic sentiment by analyzing total electricity consumption changes or assessing business performance by using individual enterprise electricity patterns. Existing research mostly ignores the relevance between enterprise electricity behaviors, making it difficult to construct a correlation model that reflects the coordination relationship of the industrial chain from the perspective of network evolution, and lacking a useful evaluation system to comprehensively represent the structural importance and evolutionary activity of enterprises or enterprise clusters in the industrial network, thus making it difficult to accurately identify advanced enterprise units in the embryonic or rapid growth stage. SUMMARY

[0003] The application proposes a power data identification method, comprising the following steps: Obtain the time series electricity data of a plurality of target enterprises in a plurality of consecutive analysis periods, and convert the time series electricity data of each enterprise in each period into a basic electricity feature vector containing power entropy, harmonic distortion rate, and instantaneous frequency deviation; For each historical analysis period, construct a three-dimensional tensor based on the basic electricity feature vector and perform Tucker decomposition, construct an electricity correlation network and calculate the structural potential of each enterprise node in the network, thereby obtaining the structural potential time series of each enterprise node; Based on the structural potential time series of each enterprise node, a nonlinear autoregressive model is used for fitting, and the absolute value of the prediction residual of the model for the structural potential of the most recent historical period is taken as the evolutionary activity of the node; the basic electricity feature vector of the current analysis period is weighted according to the evolutionary activity, the same Tucker decomposition method as the historical period is used to construct the current electricity correlation network, and the structural potential of each enterprise node in the current network is calculated; In the current electricity correlation network, when the average evolution activity of the internal nodes of a certain enterprise cluster exceeds a first threshold value, and the average structure potential of the cluster is monotonically increasing in the last N cycles including the current cycle and the total increase exceeds a second threshold value, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit.

[0004] Optionally, the converting of the time-series electricity data of each enterprise in each cycle into the basic electricity feature vector including power entropy, harmonic distortion rate and instantaneous frequency deviation degree comprises: performing fast Fourier transform on the current and voltage time-series data in a cycle to obtain a power spectrum density function P(f); normalizing the power spectrum density into a probability distribution p(f) to calculate the power entropy; extracting the fundamental component amplitude and the amplitude of each harmonic component , i=2,3,...,k, to calculate the harmonic distortion rate ; extracting the instantaneous frequency sequence using a third-order generalized integrator phase-locked loop, and calculating the standard deviation of the instantaneous frequency sequence in a cycle as the instantaneous frequency deviation degree.

[0005] Optionally, the constructing of a three-dimensional tensor based on the basic electricity feature vector for each historical analysis cycle and Tucker decomposition, and the constructing of the electricity correlation network, comprises: forming a three-dimensional tensor of K=3 basic electricity features of M enterprises in L consecutive historical analysis cycles ; performing Tucker decomposition on the tensor to obtain a core tensor G and an enterprise factor matrix A; constructing an adjacency matrix of the electricity correlation network based on the enterprise factor matrix , wherein the connection weight between enterprise u and enterprise v is obtained by calculating the cosine similarity of the corresponding row vector and the row vector in the matrix .

[0006] Optionally, the calculating of the structure potential of each enterprise node in the network comprises: defining the structure potential of the enterprise node as the PageRank value in the electricity correlation network; calculating the PageRank value of each node using an iterative algorithm, and for the weighted undirected network constructed by the cosine similarity, the iterative formula is:

[0007] ​wherein PR(u) is the PageRank value of node u, d is the damping coefficient, N is the total number of enterprise nodes, N(u) is the neighbor node set of node u, is the weight of the edge connecting node u and v, is the sum of the weights of all edges connected to node v.

[0008] Optionally, the structure potential time series of each enterprise node is fitted using a nonlinear autoregressive model, comprising: a nonlinear autoregressive neural network model is used, the structure of which is: an input layer containing the structure potential data of the previous 4 historical periods as input, a hidden layer containing 10 neurons, and an output layer; the activation function of the hidden layer uses the hyperbolic tangent function, and the activation function of the output layer uses the linear function; the model is trained until the mean square error is less than .

[0009] Optionally, the base electricity consumption feature vector of the current analysis period is weighted according to the evolution activity, comprising: for enterprise node i, the evolution activity is , and the base electricity consumption feature vector of the current analysis period is ; the evolution activity is taken as a scalar multiplier to weight the feature vector to obtain the weighted feature vector .

[0010] Optionally, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit, comprising: a community discovery algorithm is used to divide the current electricity consumption correlation network to obtain a plurality of enterprise clusters; for any cluster C, the average evolution activity of all internal nodes is calculated , if exceeds the first threshold value; and the average structure potential time series of the cluster C in the last 4 periods N=4 containing the current period is calculated , and it is verified that the sequence is strictly monotonically increasing; and the total increase of the average structure potential relative to the previous 4 periods of the current period is calculated, if , the cluster is identified as an advanced enterprise unit.

[0011] In addition, the present application also relates to a power data recognition system, comprising the following modules: a transformation module, configured to obtain time-series power consumption data of a plurality of target enterprises in a plurality of consecutive analysis periods, and transform time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation; a construction module, configured to, for each historical analysis period, construct a three-dimensional tensor based on the basic power consumption feature vector, perform Tucker decomposition, construct a power consumption correlation network and calculate structural potential of each enterprise node in the network, so as to obtain a structural potential time series of each enterprise node; a calculation module, configured to, based on the structural potential time series of each enterprise node, perform fitting by using a nonlinear autoregressive model, and take an absolute value of a prediction residual of the model for a structural potential of a most recent historical period as an evolution activity of the node; weight the basic power consumption feature vector of a current analysis period according to the evolution activity, construct a current power consumption correlation network by using a same Tucker decomposition method as that of a historical period, and calculate structural potential of each enterprise node in the current network; a recognition module, configured to, in the current power consumption correlation network, when an average evolution activity of internal nodes of a certain enterprise cluster exceeds a first threshold value, and an average structural potential of the cluster monotonically increases in the most recent N periods including the current period and a total increase amplitude exceeds a second threshold value, identify an enterprise set corresponding to the cluster as an advanced enterprise unit.

[0012] Preferably, the transformation of the time-series power consumption data of each enterprise in each period into the basic power consumption feature vector comprising the power entropy, the harmonic distortion rate and the instantaneous frequency deviation comprises: performing fast Fourier transform on current and voltage time-series data in the period to obtain a power spectrum density function P(f); normalizing the power spectrum density into a probability distribution p(f) to calculate the power entropy; extracting fundamental component amplitude based on the Fourier transform result and each harmonic component amplitude , i = 2, 3,..., k, calculating the harmonic distortion rate ; extracting an instantaneous frequency sequence by using a third-order generalized integrator phase-locked loop, and calculating a standard deviation of the instantaneous frequency sequence in one period as the instantaneous frequency deviation.

[0013] Preferably, the construction of the three-dimensional tensor based on the basic power consumption feature vector for each historical analysis period and the Tucker decomposition and the construction of the power consumption correlation network comprise: forming an M×L×K three-dimensional tensor by using K=3 basic power consumption features of M enterprises in L consecutive historical analysis periods ; performing Tucker decomposition on the tensor to obtain a core tensor G and an enterprise factor matrix A of . An adjacency matrix W of the electricity consumption correlation network is constructed based on the enterprise factor matrix A, where the connection weight between enterprise u and enterprise v is The cosine similarity between the corresponding row vector and the row vector in the matrix A is calculated.

[0014] Preferably, the structural potential of each enterprise node in the calculation network comprises: The structural potential of the enterprise node is defined as the PageRank value in the electricity consumption correlation network; The PageRank value of each node is calculated by using an iterative algorithm, and for the weighted undirected network constructed by the cosine similarity, the iterative formula is:

[0015] Where PR(u) is the PageRank value of node u, d is the damping coefficient, N is the total number of enterprise nodes, N(u) is the neighbor node set of node u, is the weight of the edge connecting node u and v, is the sum of the weights of all edges connected to node v.

[0016] Preferably, based on the time series of the structural potential of each enterprise node, a nonlinear autoregressive model is used for fitting, comprising: A nonlinear autoregressive neural network model is used, and the structure of the nonlinear autoregressive neural network model is: the input layer contains the structural potential data of the previous 4 historical periods as input, a hidden layer containing 10 neurons, and an output layer; The activation function of the hidden layer uses the hyperbolic tangent function, and the activation function of the output layer uses the linear function; The model is trained until the mean square error is less than 10 -5 .

[0017] Preferably, the basic electricity consumption feature vector of the current analysis period is weighted according to the evolutionary activity, comprising: For enterprise node i, the evolutionary activity is , and the basic electricity consumption feature vector of the current analysis period is ; The evolutionary activity is taken as a scalar multiplier to weight the feature vector to obtain the weighted feature vector .

[0018] Preferably, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit, comprising: The community discovery algorithm is used to divide the current power consumption correlation network, and a plurality of enterprise clusters are obtained. For any cluster C, the average evolution activity of all nodes inside is calculated If Exceeds the first threshold value; And, the average structural potential time sequence of the cluster C in the last 4 periods N=4 containing the current period is calculated Verify that the sequence is strictly monotonically increasing; And, the total increase of the average structural potential of the current period relative to 4 periods ago is calculated, if The cluster is identified as an advanced enterprise unit.

[0019] The present application uses high-frequency, multi-dimensional power data, not only analyzes power consumption, but also analyzes power quality data such as power entropy and harmonic distortion rate, which can reflect the advancement of enterprise production equipment and process stability. By constructing an enterprise power consumption correlation network, the isolated enterprise analysis is promoted to the industrial ecosystem level, the present application proposes two core evaluation indexes of structural potential and evolution activity, the former represents the importance of the enterprise in the network, and the latter detects the mutability and growth potential of the enterprise behavior through the nonlinear model prediction residual, which can identify the enterprise cluster advanced enterprise unit that has a sustained rise in influence and is in a rapid growth period. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the first embodiment is shown in Figure 1. Figure 2 The basic power consumption feature vector diagram is shown in Figure 2. Figure 3 The power consumption correlation network diagram is shown in Figure 3. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0022] In a first embodiment, the present application proposes a power data recognition method, such as Figure 1 , comprising the following steps: S1, obtaining the time-series power data of a plurality of target enterprises in a plurality of consecutive analysis periods, and converting the time-series power data of each enterprise in each period into a basic power feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation degree; Collect 12 consecutive months of power data, and define each month as an analysis period. Divide the numerical range of the active power sequence of each enterprise in a single analysis period into multiple intervals, and count the frequency of power values falling into each interval. Based on the information entropy formula, the power entropy is calculated, which represents the complexity and stability of the enterprise power mode. Collect voltage and current waveform data, calculate the amplitude of each harmonic component by Fourier transform, and obtain the harmonic distortion rate according to the calculation formula of the total harmonic distortion rate. This value reflects the use of nonlinear equipment and the advancement of production process of the enterprise. Record the instantaneous frequency of the power grid, calculate the root mean square error of the instantaneous frequency relative to the standard industrial frequency (e.g. 50Hz) in an analysis period, and obtain the instantaneous frequency deviation degree. This value reflects the impact of the start and stop of the enterprise's high-power equipment on the power grid. Each enterprise in each analysis period forms a three-dimensional basic power feature vector with the three indicators, as shown in Figure 2 .

[0023] S2, for each historical analysis period, construct a three-dimensional tensor based on the basic power feature vector and perform Tucker decomposition, construct a power consumption correlation network and calculate the structural potential of each enterprise node in the network, thereby obtaining the structural potential time series of each enterprise node; For any historical analysis period, calculate the feature similarity between all pairs of enterprises. For example, the similarity between enterprise i and enterprise j in terms of power entropy can be represented by the reciprocal of the difference between their power entropy values. Taking the enterprise, the enterprise and the power consumption feature as three dimensions, a three-dimensional tensor of N×N×3 is constructed, where N is the total number of enterprises, 3 is the number of feature dimensions, and each element in the tensor represents the correlation strength between two enterprises in a certain feature. Perform Tucker decomposition on the tensor to obtain a core tensor and three factor matrices. Each row of the factor matrix corresponding to the enterprise dimension can be regarded as the coordinate vector of the enterprise in the potential industrial correlation space. By calculating the inner product of any two enterprise coordinate vectors, an N×N power consumption correlation network adjacency matrix can be constructed, and the value in the matrix represents the correlation tightness between enterprises, as shown in Figure 3 The importance score of each enterprise node in the network is calculated using the web page ranking algorithm or the feature vector centrality algorithm, and the score is taken as the structural potential of the enterprise in the period. Repeat the above process for all historical analysis periods to obtain a structural potential time series for each enterprise.

[0024] S3, based on the structure potential time series of each enterprise node, a nonlinear autoregressive model is fitted, and the absolute value of the prediction residual of the model for the recent historical period structure potential is taken as the evolution activity of the node; according to the evolution activity, the basic electricity consumption feature vector of the current analysis period is weighted, the same Tucker decomposition method as the historical period is used to construct the current electricity consumption correlation network, and the structure potential of each enterprise node in the current network is calculated; For each enterprise, take the structure potential time series of the past 11 historical periods as training data. A multilayer perceptron containing an input layer, a hidden layer and an output layer is selected as the implementation method of the nonlinear autoregressive model, wherein the input of the model is the structure potential of the previous several periods, for example, the structure potential of the previous 3 periods, and the output is the prediction of the structure potential of the next period. The model is trained using historical data until the model converges. Using the trained model, input the structure potential data of the 9th, 10th and 11th periods to predict the structure potential of the 12th period, i.e. the most recent historical period. The difference between the predicted value and the actual structure potential of the 12th period is obtained. The absolute value of the residual is defined as the evolution activity of the enterprise node. The size of the obtained value reflects the degree of change or unpredictability of the enterprise's recent position in the industrial network. Enter the current analysis period, i.e. the 13th month, and calculate the three-dimensional basic electricity consumption feature vector of each enterprise in the current period according to the above method. Multiply the evolution activity of each enterprise calculated above by the basic electricity consumption feature vector of the enterprise in the current period to obtain the weighted feature vector. The higher the evolution activity of the enterprise, the longer the feature vector, indicating that the influence of the enterprise will be amplified when constructing the current network. Based on the weighted feature vectors of all enterprises, repeat the whole process of tensor construction, Tucker decomposition, correlation network construction and structure potential calculation to obtain the latest structure potential of each enterprise node in the current analysis period.

[0025] S4, in the current electricity consumption correlation network, when the average evolution activity of the internal nodes of a certain enterprise cluster exceeds the first threshold value, and the average structure potential of the cluster is monotonically increasing in the last N periods including the current period and the total increase exceeds the second threshold value, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit.

[0026] A community detection algorithm, such as Louvain algorithm, is run on the current electricity consumption correlation network to divide the closely correlated enterprise nodes into different enterprise clusters. For each identified cluster, the arithmetic mean of the evolution activity of all enterprise nodes inside is calculated. If the mean is greater than a preset first threshold, such as the level of the top 10% of the evolution activity of all enterprises, the cluster passes the activity test. For the cluster that passes the activity test, the average structure potential of the last N periods, such as the last 4 periods, is calculated, that is, the average of the structure potential of all members of the cluster in each period. It is checked whether the average structure potential sequence of length 4 is strictly monotonically increasing, and whether the total increase of the last period relative to the first period exceeds a preset second threshold, such as 30%. When an enterprise cluster meets both the high average evolution activity and the continuous growth of the average structure potential, all enterprises in the cluster collectively constitute an identified advanced enterprise unit.

[0027] In an optional embodiment, the transformation of the time series electricity consumption data of each enterprise in each period into a basic electricity consumption feature vector containing power entropy, harmonic distortion rate and instantaneous frequency deviation degree comprises: Performing fast Fourier transform on the current and voltage time series data in a period to obtain a power spectral density function P(f); normalizing the power spectral density to a probability distribution p(f) to calculate the power entropy; Based on the Fourier transform result, extracting the fundamental component amplitude And the amplitude of each harmonic component , i = 2, 3,..., k, calculating the harmonic distortion rate ; Using a third-order generalized integrator phase-locked loop to extract the instantaneous frequency sequence, and calculating the standard deviation of the instantaneous frequency sequence in a period as the instantaneous frequency deviation degree.

[0028] Taking the data of an enterprise in an analysis period of one hour as an example, the current and voltage time series data sampled at a frequency of 5 kHz in the period are collected. Fast Fourier transform is performed on the two sets of time series data to obtain a power spectral density function representing how the signal power is distributed with frequency. The power spectral density function is normalized to make the power spectral density function a probability distribution, such as calculating that the power proportion at different frequency points is 0.2, 0.15, etc. Further, the information entropy formula The power entropy value is calculated, and the obtained value is assumed to be 2.1. Using the Fourier transform result, the fundamental wave, i.e., the voltage amplitude of the 50 Hz frequency, such as 220 volts, is extracted, and the voltage amplitudes of the second, third, and up to the fiftieth harmonics, such as 5 volts, 3 volts, etc., are extracted. According to the calculation formula of the harmonic distortion rate, the square sum of the amplitudes of each harmonic is taken, and then divided by the fundamental wave amplitude to obtain a rate value, such as 0.04. The voltage signal is processed by a third-order generalized integrator phase-locked loop technology to obtain an instantaneous frequency sequence fluctuating around 50 Hz, and the standard deviation of the sequence is calculated, such as 0.02 Hz. The basic power consumption feature vector of the enterprise in this period is composed of a vector containing three numerical values, i.e., 2.1, 0.04, and 0.02.

[0029] In an optional embodiment, for each historical analysis period, a three-dimensional tensor is constructed based on the basic power consumption feature vector, and Tucker decomposition is performed, and a power consumption correlation network is constructed, including: K=3 basic power consumption features of M enterprises in L consecutive historical analysis periods form a three-dimensional tensor of MxLxK ; The tensor is subjected to Tucker decomposition to obtain a core tensor G and an enterprise factor matrix A of ; The adjacency matrix W of the power consumption correlation network is constructed based on the enterprise factor matrix A, and the connection weight between enterprise u and enterprise v is The cosine similarity between the corresponding row vectors and the row vector in the matrix A is calculated to obtain

[0030] Assume that the power consumption of 100 enterprises in the past 12 months is analyzed, and each enterprise has a three-dimensional feature vector composed of power entropy, harmonic distortion rate, and instantaneous frequency deviation degree in each month. A data tensor with dimensions of 100 enterprise numbers x 12 cycle numbers x 3 feature numbers is formed. The tensor records all power consumption feature information of all enterprises at all time points. The three-dimensional tensor is subjected to Tucker decomposition, and the rank of the enterprise dimension after decomposition is set to 10, the rank of the time dimension is set to 4, and the rank of the feature dimension is set to 2. After decomposition, a core tensor and three factor matrices are obtained, in the form of where A is the 100x10 matrix of enterprise factors. Each row of the matrix is a ten-dimensional latent feature representation of a corresponding enterprise, capturing the enterprise's relationship with other enterprises, time, and electricity characteristics. To construct the electricity correlation network among enterprises, the cosine similarity of any two rows of the enterprise factor matrix A is calculated. For example, the cosine similarity of the fifth and eighth rows is calculated to be 0.85, so the connection weight between enterprise five and enterprise eight in the network's adjacency matrix is 0.85. This is done for all pairs of enterprises, resulting in a 100x100 weighted adjacency matrix that represents the electricity correlation network among enterprises.

[0031] In an alternative embodiment, the structural potential of each enterprise node in the computing network is calculated, comprising: defining the structural potential of each enterprise node as the PageRank value in the electricity correlation network; calculating the PageRank value of each node using an iterative algorithm, for a weighted undirected network constructed from the cosine similarity, the iterative formula is:

[0032] where PR(u) is the PageRank value of node u, d is the damping factor, N is the total number of enterprise nodes, N(u) is the set of neighbor nodes of node u, is the weight of the edge connecting nodes u and v, is the sum of the weights of all edges connected to node v.

[0033] In an electricity correlation network containing 100 enterprise nodes, each enterprise node is assigned an initial ranking value, preferably 1% of the total number of nodes. The damping factor is set to 0.85. In the first iteration calculation, for enterprise node A, it needs to combine all its directly connected neighbor nodes, such as B and C. It needs to know the current PageRank value of B and C, which is 1% in the first iteration. It needs to calculate the weighted influence of the flow from B and C to A. Through the edge weight connecting A and B, such as 0.9, divided by the sum of the weights of all edges connected to B, such as 3.5, multiplied by the PageRank value of B, we get. The same calculation is done for C. Sum the weighted influence of all neighbor nodes flowing to A, multiply by the damping factor 0.85, and add a base value of 0.15 / 100 to get the new PageRank value of A after the first iteration. The process is performed simultaneously for all 100 nodes in the network, and the entire iteration process is repeated until the PageRank values of all nodes tend to be stable and no longer change significantly. The stable value obtained is the structural potential score of each enterprise node.

[0034] In an alternative embodiment, the structural potential time series of each enterprise node is fitted with a nonlinear autoregressive model, comprising: The nonlinear autoregressive neural network model has the structure of an input layer containing the structural potential data of the previous four historical periods as input, a hidden layer containing ten neurons, and an output layer; The activation function of the hidden layer uses the hyperbolic tangent function, and the activation function of the output layer uses the linear function; The model is trained until the mean square error is less than 10 -5 .

[0035] Take a certain enterprise as an example, obtain its structural potential score time series in the past 24 consecutive analysis periods, such as 0.021, 0.023, etc. In order to train a nonlinear autoregressive model to predict future trends, the sequence is converted into a set of input-output pairs. The input of the first training sample is the structural potential score of the first to fourth periods, and the output is the score of the fifth period; the input of the second sample is the score of the second to fifth period, and the output is the score of the sixth period, and so on, a total of 20 training samples are generated.

[0036] The training sample is input into a predetermined neural network, which has four input nodes receiving the structural potential data of the previous four periods, a hidden layer containing ten neurons, and an output node for predicting the structural potential of the next period. During training, the weights inside the network are constantly adjusted using optimization algorithms, with the goal of minimizing the mean square error between the network's predicted output and the actual output. When the error value is successfully reduced to less than 10 -5 , the training stops. After training is complete, the model can be used to predict the structural potential of the next period.

[0037] In an alternative embodiment, the base electricity consumption feature vector of the current analysis period is weighted according to the evolutionary activity, comprising: For enterprise node i, the evolutionary activity is , and the base electricity consumption feature vector of the current analysis period is ; The evolutionary activity is used as a scalar multiplier to weight the feature vector , obtaining the weighted feature vector .

[0038] Specifically, for enterprise A, in the current analysis cycle, the structural potential data of the past four cycles is input into the nonlinear autoregressive model trained in the previous stage to predict the structural potential of the current cycle. Then the predicted value is compared with the actual calculated structural potential of the current cycle, and the difference between the two is defined as the evolution activity of the enterprise. Assuming that the evolution activity of enterprise A is calculated to be 1.3. Normally, the basic power consumption feature vector of enterprise A in the current cycle is calculated, assuming that the power entropy is 2.0, the harmonic distortion rate is 0.05, and the instantaneous frequency deviation is 0.015, the vector composed of 2.0, 0.05, and 0.015. Multiply the scalar value of the evolution activity, i.e. 1.3, by each component of the three-dimensional feature vector. After calculation, the new weighted feature vector is 2.6, 0.065, and 0.0195. The new vector not only reflects the current basic power consumption state of enterprise A, but also integrates the change trend of the structural potential, making the feature more time-effective.

[0039] In an optional embodiment, the step of identifying the enterprise set corresponding to the cluster as an advanced enterprise unit comprises: dividing the current power consumption correlation network by using a community discovery algorithm to obtain a plurality of enterprise clusters; For any cluster C, calculating the average evolution activity of all internal nodes If exceeds the first threshold value 1.5; And, calculating the average structural potential time series of the cluster C in the last 4 cycles N=4 including the current cycle Verify that the sequence is strictly monotonically increasing; And, calculate the total increase of the average structural potential of the current cycle relative to four cycles ago, if The cluster is identified as an advanced enterprise unit.

[0040] Specifically, the community discovery algorithm is used to process the power consumption correlation network in the current cycle, and the enterprise nodes in the network are automatically divided into several closely connected clusters. One cluster containing 10 enterprises is selected for analysis. The evolution activity of each of the 10 enterprises is calculated, and the average value is obtained. If the calculated average evolution activity is 1.6, which is higher than the preset first threshold value 1.5, the first condition is met.

[0041] The performance of the cluster in the last four periods is traced back. The structural potential average of the 10 enterprises in each period is calculated to obtain a time series with a length of four, for example, the values are 0.030, 0.035, 0.042, and 0.048 respectively. Since the latter values are all greater than the former, the sequence is strictly monotonically increasing, which meets the second condition. The overall growth rate of the structural potential in the four periods is calculated, and the result is 0.6. The obtained growth rate is greater than the set second threshold value 0.5, so the third condition is also met. Since the cluster passes the test of all three conditions, the set of the 10 enterprises is identified as an advanced enterprise unit.

[0042] In a second embodiment, a power data identification system is provided, comprising the following modules: A transformation module is configured to obtain time-series power consumption data of a plurality of target enterprises in a plurality of consecutive analysis periods, and transform the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate, and instantaneous frequency deviation degree; A construction module is configured to, for each historical analysis period, construct a three-dimensional tensor based on the basic power consumption feature vector and perform Tucker decomposition, construct a power consumption correlation network, and calculate the structural potential of each enterprise node in the network, thereby obtaining a structural potential time series of each enterprise node; A calculation module is configured to, based on the structural potential time series of each enterprise node, use a nonlinear autoregressive model for fitting, and take the absolute value of the prediction residual of the model for the structural potential of the most recent historical period as the evolution activity of the node; weight the basic power consumption feature vector of the current analysis period according to the evolution activity, construct the current power consumption correlation network using the same Tucker decomposition method as the historical period, and calculate the structural potential of each enterprise node in the current network; An identification module is configured to, in the current power consumption correlation network, when the average evolution activity of the internal nodes of a certain enterprise cluster exceeds the first threshold value, and the average structural potential of the cluster is monotonically increasing in the last N periods including the current period and the total increase exceeds the second threshold value, identify the enterprise set corresponding to the cluster as an advanced enterprise unit.

[0043] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocksFigure 1 means for performing the function specified in the block or blocks.

[0044] The above merely provides an example of the present application, but does not serve to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A power data recognition method, characterized by, The method comprises the following steps: obtaining time-series power consumption data of a plurality of target enterprises in a plurality of consecutive analysis periods, and converting the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation; for each historical analysis period, constructing a three-dimensional tensor based on the basic power consumption feature vector and performing Tucker decomposition, constructing a power consumption correlation network and calculating the structural potential of each enterprise node in the network, thereby obtaining a structural potential time series of each enterprise node; based on the structural potential time series of each enterprise node, fitting a nonlinear autoregressive model, and taking the absolute value of the prediction residual of the model for the structural potential of the most recent historical period as the evolutionary activity of the node; weighting the basic power consumption feature vector of the current analysis period according to the evolutionary activity, constructing a current power consumption correlation network using the same Tucker decomposition method as the historical period, and calculating the structural potential of each enterprise node in the current network; in the current power consumption correlation network, when the average evolutionary activity of the internal nodes of a certain enterprise cluster exceeds a first threshold value, and the average structural potential of the cluster is monotonically increasing in the most recent N periods including the current period and the total amplitude exceeds a second threshold value, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit.

2. The method of claim 1, wherein, The conversion of the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation comprises: performing fast Fourier transform on the current and voltage time-series data in the period to obtain a power spectrum density function P(f); normalizing the power spectrum density to a probability distribution p(f) to calculate the power entropy; Based on the Fourier transform result, the fundamental component amplitude and each harmonic component amplitude are extracted, i=2,3,...,k, the harmonic distortion rate is calculated ; extracting the instantaneous frequency sequence using a third-order generalized integrator phase-locked loop, and calculating the standard deviation of the instantaneous frequency sequence in one period as the instantaneous frequency deviation.

3. The method of claim 1, wherein, The construction of a three-dimensional tensor based on the basic power consumption feature vector for each historical analysis period and the Tucker decomposition and the construction of a power consumption correlation network comprise: K=3, M=3, L=3 ; on the tensor Tucker decomposition to obtain the core tensor G and the business factor matrix A; An adjacency matrix W of the electricity consumption correlation network is constructed based on the enterprise factor matrix A, wherein the connection weight between enterprise u and enterprise v is The cosine similarity between the corresponding row vectors of matrix A and the row vector is calculated to obtain.

4. The method according to claim 1 or 3, characterized in that, The calculation of the structural potential of each enterprise node in the network comprises: defining the structural potential of an enterprise node as the PageRank value in the power consumption correlation network; using an iterative algorithm to calculate the PageRank value of each node, and for a weighted undirected network constructed by cosine similarity, the iteration formula is: where PR(u) is the PageRank value of node u, d is the damping factor, N is the total number of enterprise nodes, N(u) is the neighbor node set of node u, is the weight of the edge connecting node u and v, is the sum of the weights of all edges connected to node v.

5. The method of claim 1, wherein, The fitting of a nonlinear autoregressive model based on the structural potential time series of each enterprise node comprises: using a nonlinear autoregressive neural network model, the structure of which is: the input layer contains the structural potential data of the previous 4 historical periods as input, a hidden layer containing 10 neurons, and an output layer; the activation function of the hidden layer uses the hyperbolic tangent function, and the activation function of the output layer uses the linear function; Train the model until the mean squared error is less than 10 -5 .

6. The method of claim 1, wherein, The weighting of the basic power consumption feature vector of the current analysis period according to the evolutionary activity comprises: For enterprise node i, the evolution activity is The basic electricity feature vector of the current analysis cycle is ; Evolution activity As a scalar multiplier, the eigenvector is weighted to obtain a weighted eigenvector .

7. The method of claim 1, wherein, The identification of the enterprise set corresponding to the cluster as an advanced enterprise unit comprises: using a community discovery algorithm to divide the current power consumption correlation network to obtain a plurality of enterprise clusters; For any one cluster C, calculate the average evolutionary activity of all nodes inside , if exceeds a first threshold value; And, compute the average structural potential time series of the cluster C over the last 4 periods N=4 including the current period , verify that the series is strictly monotonically increasing; And, the average structure potential of the current period is calculated relative to the total increase of 4 periods ago, if it satisfies The cluster is identified as an advanced enterprise unit.

8. An electric power data recognition system characterized by, comprise the following modules: The conversion module is configured to acquire time-series power consumption data of a plurality of target enterprises in a plurality of continuous analysis periods, and convert the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation; The construction module is configured to, for each historical analysis period, construct a three-dimensional tensor based on the basic power consumption feature vector and perform Tucker decomposition, construct a power consumption correlation network and calculate the structural potential of each enterprise node in the network, so as to obtain a structural potential time series of each enterprise node; The calculation module is configured to, based on the structural potential time series of each enterprise node, fit a nonlinear autoregressive model, and take the absolute value of the prediction residual of the model for the structural potential of the most recent historical period as the evolution activity of the node; weight the basic power consumption feature vector of the current analysis period according to the evolution activity, construct a current power consumption correlation network by using the same Tucker decomposition method as the historical period, and calculate the structural potential of each enterprise node in the current network; The identification module is configured to, in the current power consumption correlation network, when the average evolution activity of the internal nodes of a certain enterprise cluster exceeds a first threshold value, and the average structural potential of the cluster monotonically increases in the most recent N periods including the current period and the total increase amplitude exceeds a second threshold value, identify the enterprise set corresponding to the cluster as an advanced enterprise unit.

9. The system of claim 8, wherein, The conversion module is configured to acquire time-series power consumption data of a plurality of target enterprises in a plurality of continuous analysis periods, and convert the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation; The conversion module is configured to acquire time-series power consumption data of a plurality of target enterprises in a plurality of continuous analysis periods, and convert the time-series power consumption data of each enterprise in each period into a basic power consumption feature vector comprising power entropy, harmonic distortion rate and instantaneous frequency deviation; Based on the Fourier transform result, the fundamental component amplitude and each harmonic component amplitude are extracted, i = 2, 3,..., k, the harmonic distortion rate is calculated ; The construction module is configured to, for each historical analysis period, construct a three-dimensional tensor based on the basic power consumption feature vector and perform Tucker decomposition, construct a power consumption correlation network and calculate the structural potential of each enterprise node in the network, so as to obtain a structural potential time series of each enterprise node; 10. The system of claim 8, wherein, ​ K=3, M=3, L=3 ; on the tensor Tucker decomposition to obtain the core tensor G and the business factor matrix A; An adjacency matrix W of the electricity consumption correlation network is constructed based on the enterprise factor matrix A, wherein the connection weight between enterprise u and enterprise v is The cosine similarity between the corresponding row vectors of matrix A and the row vector is calculated to obtain.

Images