Electricity stealing user intelligent identification method based on electricity utilization information acquisition

Abstract

The invention discloses an electricity stealing user intelligent identification method based on electricity consumption information acquisition. The method comprises the following steps: S1, importing user electricity consumption data; s2, rejecting users whose missing number is greater than or equal to 725; s3, data preprocessing; s4, dividing the power consumption data set according to a K-fold cross validation mode; s5, defining a search space, and performing architecture search through an NAS algorithm; s6, performing model evaluation on a model obtained by NAS algorithm search; s7, repeating the step S4, the step S5 and the step S6, completing K-fold cross validation, and taking average as final performance evaluation; neural architecture searching, automatic network structure searching and hyper-parameter searching are adopted to optimize the neural network performance, electricity stealing identification in a complex environment is dealt with, effective feature representation and model structures suitable for electricity stealing data are automatically explored, more complex and effective neural network structures are automatically explored, the method is more suitable for specific tasks, and the method is suitable for large-scale popularization and application. And continuous automatic updating is realized, so that the accuracy, robustness and adaptability of electricity larceny identification are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of identifying electricity stealing users in power grids, and in particular relates to an intelligent identification method for electricity stealing users based on electricity consumption information collection. Background Art

[0002] In the current power industry, anti-electricity theft is still a difficult problem. How to more accurately detect electricity theft users and prevent the occurrence of electricity theft has become a new problem in the informatization of the power grid. With the development of technology, it is becoming more and more difficult to identify electricity theft users, and the challenges it brings to us are becoming more and more daunting.

[0003] With the rapid development of new power systems and the promotion and coverage of electricity consumption information collection systems, it has become possible to remotely read user electricity consumption records, laying a data foundation for power big data and providing fertile ground for development. At present, the use of big data analysis methods to identify electricity theft users has received widespread attention from scholars at home and abroad, and the massive amount of electricity consumption information has laid a solid foundation for the development of power big data technology, making it possible to use big data technology to intelligently identify and detect electricity theft users.

[0004] Big data technology using neural network structure is suitable for intelligent identification of electricity theft users, but the traditional neural network structure design mainly relies on manual experience and professional knowledge, which requires a lot of time and resources to adjust the network structure and is prone to fall into local optimality. In order to solve the above problems, it is necessary to develop an intelligent identification method for electricity theft users based on electricity consumption information collection. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide an intelligent identification method for electricity theft users based on electricity consumption information collection with higher accuracy, robustness and adaptability.

[0006] The object of the present invention is achieved by: a method for intelligently identifying electricity stealing users based on electricity consumption information collection, comprising the following steps:

[0007] Step S1, importing user power consumption data;

[0008] Step S2, remove users with missing number ≥ 725;

[0009] Step S3, data preprocessing;

[0010] Step S4, dividing the electricity consumption data set by K-fold cross validation;

[0011] Step S5, defining the search space and performing architecture search using the NAS algorithm;

[0012] Step S6, performing model evaluation on the model searched by the NAS algorithm;

[0013] Step S7, repeat steps S4, S5, and S6 to complete K-fold cross validation, and take the average as the final performance evaluation.

[0014] Preferably, in step S3, data preprocessing specifically includes first filling in missing data and then normalizing the data.

[0015] Preferably, when filling the missing data, for the time period with more than 1 missing day, 0 is filled in, and the corresponding missing time point position is recorded and input into the model for training as the missing position information; for the time period with 1 missing day, linear interpolation is used to fill the data, and the interpolation function F is as follows:

[0016]

[0017] In the formula, point (x1, y1) and point (x2, y2) are the left and right points adjacent to the missing time point x respectively; the filling is completed from the middle to both sides.

[0018] Preferably, the data normalization uses Max-min standardization to convert the user's power consumption per unit time into a standard value in the range of [0,1]. The Max-min standardization formula is as follows:

[0019]

[0020] In the formula, X m is the original data vector, X m ′ is the data vector after Max-min normalization.

[0021] Preferably, in step S4, the electricity consumption data set is divided into a training set and a test set, wherein the training set accounts for 50%, 75%, or 80%.

[0022] Preferably, in step S5, the search space consists of ConvBlock, XceptionBlock and ResNetBlock, Auto-Keras searches for all neural network structures and their combinations in the defined search space, and adjusts hyperparameters through Keras-Tuner, where the maximum number of trials is set to 200.

[0023] Preferably, in step S5, the data feature extraction model is automatically generated and optimized using NAS, and the NAS framework used is Auto-Keras, which uses Bayesian optimization to guide the operation selection of the architecture search based on network morphism.

[0024] Preferably, the Bayesian optimization is specifically as follows: first input the prediction function f, the parameter space range X to be searched, the acquisition function S and the type of the model M used to fit the distribution, then initialize a random data set D, the random data set D does not contain any information about the prediction function f, use the random data set D to fit the model M, then use the model M to estimate the distribution of y, and finally when the model M converges (that is, when the optimal parameters are found), stop the iteration and output the optimal parameters.

[0025] Due to the adoption of the above technical scheme, the beneficial effects of the present invention are as follows: the present invention adopts neural architecture search, which can optimize the performance of neural networks by automatically searching network structures and hyperparameters to cope with electricity theft identification in complex environments. At the same time, neural architecture search can also automatically explore effective feature representations and model structures suitable for electricity theft data, automatically explore more complex and effective neural network structures, make them more suitable for specific tasks, and can be continuously and automatically updated, thereby improving the accuracy, robustness and adaptability of intelligent identification of electricity theft users. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of the method for intelligently identifying electricity stealing users of the present invention.

[0027] Figure 2 It is a block diagram of the search space composition of the present invention.

[0028] Figure 3 It is a flowchart of the Bayesian optimization process of the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0030] Neural Architecture Search (NAS) is a method for automatically designing neural network structures. It automatically searches and optimizes network structures to achieve better performance on given tasks. Traditional neural network structure design often relies on manual experience and expertise, takes a lot of time and resources to adjust the network structure, and is prone to falling into local optimality. NAS can automatically explore more complex and effective neural network structures, making neural network models more suitable for specific tasks.

[0031] like Figure 1 As shown, the present invention provides an intelligent identification method for electricity theft users based on electricity consumption information collection, comprising the following steps:

[0032] Step S1, importing user power consumption data;

[0033] Step S2, remove users with missing number ≥ 725;

[0034] Step S3, data preprocessing;

[0035] Step S4, dividing the electricity consumption data set by K-fold cross validation;

[0036] Step S5, defining the search space and performing architecture search using the NAS algorithm;

[0037] Step S6, performing model evaluation on the model searched by the NAS algorithm;

[0038] Step S7, repeat steps S4, S5, and S6 to complete K-fold cross validation, and take the average as the final performance evaluation.

[0039] The core goal of neural architecture search is to optimize the performance of the network by automatically searching the network topology and hyperparameters. The NAS process can be seen as an optimization problem of finding the optimal network structure in a large search space, where the search space includes different network layers, widths, connection methods, and various modules (such as convolution, loop, etc.).

[0040] In neural architecture search, existing common and classic search algorithms can be used, including evolutionary algorithms, genetic algorithms, reinforcement learning, Bayesian optimization, etc. These algorithms can evaluate the performance of network structures based on different evaluation criteria, such as accuracy on the validation set, model complexity and other indicators. The search algorithm automatically selects and adjusts the network structure according to these criteria, and gradually optimizes the network performance. These search spaces are often very large and exceed the ability of manual design, so intelligent search algorithms are needed to efficiently find better network structures. In the process of neural architecture search, technologies such as strategy parameter sharing and model weight sharing can be used to accelerate the search and reduce the computational cost of the search.

[0041] When the backend system collects user data, data collection may fail in a certain period of time. The specific reasons include power failure of the meter, interference with the carrier signal, interruption of the transmission network, etc. Failure to collect data will cause discontinuity of data at a certain time, that is, data loss, so data preprocessing is required.

[0042] Data preprocessing first performs missing data filling and then data normalization.

[0043] When filling missing data, the time period with more than 1 missing day is filled with 0, and the corresponding missing time point position is recorded and input into the model for training as the missing position information; for the time period with 1 missing day, linear interpolation is used to fill the data, and the interpolation function F is as follows:

[0044]

[0045] In the formula, point (x1, y1) and point (x2, y2) are the left and right points adjacent to the missing time point x respectively; the filling is completed from the middle to both sides.

[0046] In order to better understand the common features and electricity theft features among different users, the data needs to be matrixed as follows:

[0047]

[0048] In the formula, Xm is the electricity consumption data vector of user m, n is the number of days, x mn is the daily electricity consumption of user m on the nth day.

[0049] There are large differences in electricity consumption among users, and the model mainly estimates the user's suspicion of electricity theft through the changing trend of electricity consumption. Therefore, it is necessary to normalize the main sentence of electricity consumption to eliminate the impact of electricity consumption differences.

[0050] Data normalization uses Max-min standardization, which converts the user's unit time power consumption into a standard value in the range of [0,1]. The Max-min standardization formula is as follows:

[0051]

[0052] In the formula, X m is the original data vector, X m ′ is the data vector after Max-min normalization.

[0053] Preferably, in step S4, the electricity consumption data set is divided into a training set and a test set, wherein the training set accounts for 50%, 75%, or 80%.

[0054] like Figure 2 As shown, the search space consists of ConvBlock, XceptionBlock and ResNetBlock. Auto-Keras searches for all neural network structures and their combinations in the defined search space, and adjusts the hyperparameters through Keras-Tuner, where the maximum number of trials is set to 200. If the computer obtains the best model in 200 trials, the search is stopped. The present invention uses an early stopping mechanism to stop the search training. When the loss of the development set does not decrease for 10 consecutive epochs, the current training is stopped.

[0055] The data feature extraction model is automatically generated and optimized using NAS, and the NAS framework used is Auto-Keras. Auto-Keras uses Bayesian optimization to guide the operation selection of architecture search based on network morphism. The main purpose of using Bayesian optimization is to reduce the amount of calculation required for architecture search and improve model performance. The final result of the algorithm is the optimized parameters and model structure of the feature extraction model, and the optimal model is then used to extract features.

[0056] For a function f: x→R, we need to find X inside x∈X * =argmin x∈X f(x), where x is a set of structural parameters and X is the structural parameter space. Since the function f is not necessarily convex and the calculation takes a lot of resources, Bayesian optimization is used to deal with this problem.

[0057] like Figure 3 As shown in Figure 1, Bayesian optimization can use a relatively simple optimization method based on a sequence model, where: f is a prediction function or model, that is, a set of structural parameters is set for the model, and an output value is obtained by calculating the model; X is the parameter space; D represents a data set consisting of multiple (x, y); x is a set of parameters; y represents the result corresponding to the set of parameters; S is the acquisition function, which is used to determine X. * =argmin x∈X The selection of x in f(x); M is a model obtained by fitting the data set D, and commonly used ones include Gaussian model, random forest, etc. The present invention adopts Gaussian distribution model, and INITSAMPLE is the initialization D.

[0058] First, you need to input the prediction function f, the range X of the parameter space to be searched, the acquisition function S, and the type of model M used to fit the distribution. It is crucial to design appropriate acquisition and fitting functions. After designing the parameters, initialize a random data set that does not contain any information about the function f. Then, use the data set D to fit the model M, and then use M to estimate the distribution of y. The first iteration does not have much information about f, so the fitted model M does not conform to the actual distribution, but as the information about f (x i ,y i ) are continuously added to D, and the model M's estimate of the distribution becomes more and more accurate. Finally, when the model converges (that is, when the optimal parameters are found), the iteration stops and the optimal parameters are output.

[0059] The core issue of Bayesian optimization is which set of parameters to evaluate; because the cost of evaluating each point is very high, it is crucial to choose the next set of parameters to be evaluated. The present invention makes this decision through the acquisition function. Therefore, the most important thing in Bayesian optimization is to design the acquisition function, which plays a vital role in the performance of the algorithm.

[0060] Assume f′minf, where f′ represents the currently known minimum value of f, and define the ulity function as:

[0061]

[0062] The above u(x) can be understood as a reward function. If f(x) is not greater than f', there is a reward, otherwise there is no reward. The acquisition function can be defined as follows:

[0063]

[0064] Just need to issue a PI The maximum value of (x) can be used to find the x that meets the requirements based on the Gaussian distribution model.

[0065] The following are several comparative tests provided by the present invention and analysis of their results.

[0066] The following table shows the test results of the NAS model trained on missing location data, normal electricity consumption data, and their combination (using multi-fold cross validation, where the training set accounts for 50%, 75%, and 80% respectively.)

[0067]

[0068] It can be seen that when the training set accounts for 75%, the model trained on missing location information (missing location model) achieves better performance, with AUC and F1 of 0.926 and 0.605 respectively. The AUC and F1 of the model trained on electricity consumption data are 0.877 and 0.486 (75%) respectively. When the combination of normal electricity consumption data and missing value location information is used as the training set, the AUC and F1 reach 0.928 and 0.603 (75%) respectively. It can be seen that adding electricity consumption data does not improve the performance of the model. The experimental results show that the location information of missing values ​​is highly correlated with electricity theft.

[0069] The following table shows the results of the NAS model and the manually designed model (using multiple cross-validation, where the training set accounts for 50%, 75% and 80% respectively).

[0070]

[0071] It can be seen that when the training set is 75%, the AUC and F1 of the NAS search model are 0.877 and 0.486 respectively, and the AUC and F1 of the manual model are 0.857 and 0.481 respectively. The NAS model is better than the manually designed model. However, when the training set is 80%, the AUC and F1 of the NAS model are 0.874 and 0.452 respectively, and the AUC and F1 of the manually designed model are 0.877 and 0.490 respectively. The performance of the manually designed model is better than the NAS model. It can be seen that there is no obvious difference between the performance of the NAS model and the manually designed model.

[0072] The comparative test results show that NAS search technology can achieve performance comparable to that of manually designed models in a very short time without the need for professional neural network construction. Moreover, power companies can update the model at any time to respond to changes in electricity consumption behavior and electricity theft patterns.

[0073] 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, those skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for intelligently identifying electricity theft users based on electricity usage information collection, characterized in that: The steps include: Step S1, importing user power consumption data; Step S2, remove users with missing number ≥ 725; Step S3, data preprocessing; Step S4, dividing the electricity consumption data set by K-fold cross validation; Step S5, defining the search space and performing architecture search using the NAS algorithm; Step S6, performing model evaluation on the model searched by the NAS algorithm; Step S7, repeat steps S4, S5, and S6 to complete K-fold cross validation, and take the average as the final performance evaluation.

2. The method for intelligently identifying electricity stealing users based on electricity consumption information collection according to claim 1 is characterized in that: In step S3, data preprocessing specifically includes first filling in missing data and then normalizing the data.

3. The method for intelligently identifying electricity stealing users based on electricity consumption information collection according to claim 2 is characterized in that: When filling missing data, the missing days are filled with 0 for time periods greater than 1, and the corresponding missing time point positions are recorded and input into the model for training as missing position information; for time periods with 1 missing day, linear interpolation is used to fill data, and the interpolation function F is as follows: In the formula, point (x1, y1) and point (x2, y2) are the left and right points adjacent to the missing time point x respectively; the filling is completed from the middle to both sides.

4. The method for intelligently identifying electricity stealing users based on electricity usage information collection according to claim 2 is characterized in that: The data normalization adopts Max-min standardization to convert the user's unit time power consumption into a standard value in the range of [0,1]. The Max-min standardization formula is as follows: Where, X m is the original data vector, X m ′ is the data vector after Max-min normalization.

5. The method for intelligently identifying electricity stealers based on electricity usage information collection according to claim 1 is characterized in that: In step S4, the electricity consumption data set is divided into a training set and a test set, wherein the training set accounts for 50%, 75%, or 80%.

6. The method for intelligently identifying electricity stealers based on electricity usage information collection according to claim 1 is characterized in that: In step S5, the search space consists of ConvBlock, XceptionBlock and ResNetBlock, Auto-Keras searches for all neural network structures and their combinations in the defined search space, and adjusts hyperparameters through Keras-Tuner, where the maximum number of trials is set to 200.

7. The method for intelligently identifying electricity stealers based on electricity usage information collection according to claim 1 is characterized in that: In step S5, the data feature extraction model is automatically generated and optimized using NAS, and the NAS framework used is Auto-Keras. Auto-Keras uses Bayesian optimization to guide the operation selection of the architecture search based on network morphism.

8. The method for intelligently identifying electricity theft users based on electricity usage information collection according to claim 7 is characterized in that: The Bayesian optimization is specifically as follows: first input the prediction function f, the parameter space range X to be searched, the acquisition function S and the type of the model M used to fit the distribution, then initialize a random data set D, the random data set D does not contain any information about the prediction function f, use the random data set D to fit the model M, then use the model M to estimate the distribution of y, and finally when the model M converges (that is, when the optimal parameters are found), stop the iteration and output the optimal parameters.

Images