CN113420510B - Energy consumption distinguishing method based on front-end sensing and learning - Google Patents
Energy consumption distinguishing method based on front-end sensing and learning Download PDFInfo
- Publication number
- CN113420510B CN113420510B CN202110770785.0A CN202110770785A CN113420510B CN 113420510 B CN113420510 B CN 113420510B CN 202110770785 A CN202110770785 A CN 202110770785A CN 113420510 B CN113420510 B CN 113420510B
- Authority
- CN
- China
- Prior art keywords
- energy consumption
- learning
- network
- data
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/06—Power analysis or power optimisation
Abstract
The invention provides an energy consumption distinguishing method based on front-end perception and learning, which deploys learning ability at a front end, improves intelligent learning ability of the front end, enables a network layer not to transmit user original data any more, and transmits mapping data after feature transformation, thereby greatly reducing data transmission data volume of a network, and greatly reducing the risk of leakage of user sensitive data. Compared with the common method, the specially designed energy consumption characteristic expression method, the unique front-end modular architecture and the deployment mode can work independently of the remote platform and can also work cooperatively with the remote platform; the energy consumption characteristic learning function of local users is achieved when the energy consumption characteristic learning device works independently, and the learning performance can be improved when the energy consumption characteristic learning device works cooperatively. Through the design of the Bayesian network, the intelligent learning system has better robustness, and after the intelligent learning system enters a working state, even if the front end and the far end lose connection due to network failure, the front end equipment can still independently complete the intelligent learning task.
Description
Technical Field
The invention relates to the field of intelligent power grids, in particular to the field of intelligent electric meter monitoring.
Background
Electric power is the mainstay of production and life in modern society. On one hand, with the improvement of living standard of substances, people also put higher requirements on electricity utilization experience; on the other hand, the power utilization scale of the user is rapidly enlarged, the power utilization load fluctuation is more severe, and a severe challenge is also provided for power grid planning, scheduling and the like. From the perspective of a user, the transparency of the electricity utilization information can help the user to know the energy consumption structure of the user, and the formation of the electricity saving consciousness is promoted; from the perspective of an electric power operation unit, monitoring and analysis of the electric load of the user is beneficial to providing accurate and lean energy consumption data service for the user, and more scientific power grid planning, power generation scheduling and dynamic electricity price policies are realized.
The ubiquitous power Internet of things is application expansion of industrial Internet of things in the power industry. The overall architecture of the ubiquitous power Internet of things sequentially comprises a perception layer, a network layer, a platform layer and an application layer. The perception layer completes the homologous collection of each unit data of 'sending, outputting, changing and using', and improves the intellectualization of each terminal through edge calculation; the network layer realizes the full coverage of each link in the current power system by using the technical means of large, cloud, object, movement, intelligence and the like; the platform layer improves the high-efficiency data processing and cloud coordination capacity by utilizing central cluster facilities such as an internet of things management center and a data center; the purpose of the application layer is to establish an intelligent comprehensive energy internet on the basis of ensuring the safe and stable operation of a power grid. The sensing layer is used as the foremost end of the ubiquitous power Internet of things, is a bridge for connecting a user and a power grid, and is core content for bearing the functions of ubiquitous power Internet of things such as calculation, communication, accurate control, remote cooperation and autonomy.
When monitoring power consumption, the electric meter needs to be modified into an intelligent electric meter, and extra hardware cost is needed. At present, a machine vision monitoring method is also available, although an electric meter is not required to be modified, the requirements on storage and bandwidth of monitoring equipment are high, and the collected images can be completely and clearly stored and transmitted, so that accurate analysis on a server is realized. Therefore, this requires a large cost, which hinders practical application of machine vision monitoring. Moreover, the user's meter image is directly transmitted over the network, which creates a risk of privacy disclosure. Meanwhile, the image acquisition is greatly influenced by the environment, so that the precision of the technology is not high, and the misjudgment rate is high. Therefore, it is necessary to provide a power consumption monitoring method with low cost, high accuracy and high safety.
Disclosure of Invention
A front-end perception and learning-based energy consumption distinguishing method comprises the following steps:
step 1: continuously shooting power values at the ammeter side by using a camera module at the front end, and acquiring an image I; after the shot image I is processed by a preprocessing method, the shot image I is packed into a plurality of batches of original time sequence data packets B by a fixed frame number NiThe reading time sequence information of the electric meter at the user end is contained;
the pretreatment comprises the following steps: sequentially resampling and normalizing the image I, wherein the normalization control parameter is more than 0 and less than rho1< 1, and it is according to the time characteristic Td、TyCarrying out adjustment; definition of TdThe number of seconds from the zero point of the day when sampling; definition of TyThe number of days from the first month of the year;
step 2: constructing a neural network model for the time sequence data packet BiLearning is carried out; the neural network model consists of an input layer, an output layer and a hidden layer, wherein the loss function is as follows:
where x is the input to the neural network, y is the output value of the neural network,
representing the true output of the training samples, N is the number of samples. Theta1、θ2Are independent control variables;
Td、Tyand y constitutes a characteristic of the front-end energy consumption data, denoted as Fe，
And step 3: transmitting the characteristics of the front-end energy consumption data to a classification discrimination model,
defining:
wherein X represents a certain energy consumption characteristic FeThe sample of (1). Mu.s2、σ2Respectively representing the mean and the variance of all samples in the training sample set;
when c is 1, the energy consumption state in the time period corresponding to the sample is abnormal, when c is 0, the energy consumption state in the time period corresponding to the sample is normal, for a given sample set, the probability P (X) is a constant, and the probability P (c) represents that the sample with normal energy consumption and the sample with abnormal energy consumption respectively account for the proportion of all samples;
when the communication network is available, the communication network obtains a full-network training value of P (c), the currently obtained values of the energy consumption characteristic samples X and c are sent to a remote cloud platform, and the value of P (X | c) is calculated according to a formula (13) and then is returned to the front end;
when the communication network is not available, the pre-stored P (c), mu of the front end are used2、σ2Calculating the value of P (X | c) using equation (13) and equation (15);
and 4, step 4: and comparing the values of P (0| X) and P (1| X), if P (0| X) > P (1| X), judging that the current energy consumption is abnormal, and if not, judging that the energy consumption is normal.
The camera module at the front end continuously shoots the power value at the ammeter side, and the frame rate of the collected images is more than 1 frame/second.
The acquisition and pretreatment is performed at the front end.
Find FeThe steps of (1) are carried out at the front end.
The step of computing P (X | c) is implemented in the cloud platform when a network is available.
The step of calculating P (X | c) is implemented in the front end when the network is not available.
The determining step is performed at the front end.
Invention and technical effects
1. The learning ability is deployed at the front end, the intelligent learning ability of the front end is improved, the network layer does not transmit user original data any more, and only transmits mapping data after feature transformation, so that the data transmission data volume of the network is greatly reduced, and the risk of leakage of user sensitive data is greatly reduced.
2. Compared with the common method, the specially designed energy consumption characteristic expression method, the unique front-end modular architecture and the deployment mode can work independently of the remote platform and can also work cooperatively with the remote platform; the energy consumption characteristic learning function of local users is achieved when the energy consumption characteristic learning device works independently, and the learning performance can be improved when the energy consumption characteristic learning device works cooperatively.
3. Through the design of the Bayesian network, the whole system has better robustness compared with a common method, and after the system enters a working state, even if the front end and the far end lose connection due to network failure, the front-end equipment can still independently complete an intelligent learning task.
4. The intelligent data perception obtaining method based on the front-end video and the images is different from a general method based on port data collection of an intelligent electric meter, can be used for the intelligent electric meter and a non-intelligent electric meter with a numerical value display function, and is wider in application.
In addition, the energy consumption monitoring method based on front-end perception and learning described herein is to collect power energy consumption data of an electricity meter end through a facility deployed at a user end, that is, a front end, and complete learning of the energy consumption data at the front end, so as to monitor energy consumption of a user, thereby further improving service capability and service level of a power enterprise to the user; the method plays a key role in ubiquitous power Internet of things. The method comprises the steps that energy consumption data collected by a collecting module arranged at the front end of a user are utilized; forming the energy consumption data into a characteristic factor of the energy consumption of the user by utilizing a preprocessing module; the collected characteristic factors are transmitted back to a remote cloud platform through a communication module for centralized storage and post-processing; receiving a guiding model transmitted back by the cloud platform through a communication module; the learning module is used for learning the local characteristic factors and generating user energy consumption characteristics, so that the energy consumption of the user is monitored. The guiding model returned by the cloud platform is an optional item of the complete process of the method and is used for optimizing the learning performance of the learning module, and the learning module can work independently of the cloud platform.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:
FIG. 1 is a schematic diagram of the module deployment associated with the present invention.
Detailed Description
Step 1, the method for acquiring and preprocessing the front-end energy consumption data utilizes a miniature camera module (shown in fig. 1) assembled on the front-end ammeter side to continuously shoot the power value of the ammeter side at a certain shooting frame rate f, the shooting frame rate is not too low, and is usually not lower than 1 frame/second according to practical experience, so that the interval time between every two frames is
And the shot image I is fixed after being processed by a preprocessing methodAnd the frame number N is packaged into a plurality of batches of original time sequence data packets B.
Data packet BiThe reading time sequence information of the user-side electric meter is included, and i represents the ith batch of B. When the user uses electricity, the reading of the electric meter changes, and the time sequence change of the reading of the electric meter can reflect the characteristics of the energy consumption of the user because the reading of the electric meter is the measurement of the electricity consumption of the user, so the characteristics can be reflected by continuously shooting the pictures of the reading of the electric meter. The main advantage of this method is that it does not require the modification of the meter, even if a non-smart meter can be used; second, as the smart chip device (fig. 1) and the miniature camera system are reduced in cost and volume, low-cost deployment of the associated facilities is made possible.
For an image I acquired by a miniature camera module, preprocessing is carried out according to the following method:
the size of the source image matrix I is set to be w × h, and if the value of any point in the source image matrix I is known to be a gray value or a brightness value, the value of a certain pixel (x, y) in the resampling matrix R is calculated by using the following method. Order:
wherein, the symbol
Means for taking down integers, symbols
Indicating an upward integer. Order S11、S12、S21、S22Respectively representing the source image matrix in coordinates (x)1,y1)、(x1,y2)、(x2,y1)、(x2,y2) Value of (S)11、S12、S21、S22Are all known). R (x, y) is the value of the resampling matrix R at (x, y), and then R (x, y) is calculated according to the following formula:
in general, the above equation is difficult to obtain an integer, and therefore requires a rounding operation. But for the special case of sampling, i.e.
For integer cases, contract x2＝x1+1, likewise, for
For integer cases, contract y2＝y1+1. For the resampled data R (x, y), the following equation (3) is adopted for normalization to obtain a normalized image I' (x, y):
mu in the above formula1Representing the arithmetic mean, σ, of the resampling matrix R1Represents the standard deviation of the resampling matrix R; rho is more than 01< 1 is a control parameter according to the time characteristic T in step 2d、TyAnd adjusting to overcome the environmental influence and ensure that the identification is more accurate and rapid. As a set of preferred values, the following are set:
c1, when 90 < TyWhen the ratio is less than or equal to 273, if
C1.1 when 21600 < TdWhen the value is less than or equal to 64800, taking rho1＝0.45；
C1.2 when Td64800 or TdWhen the value is less than or equal to 21600, taking rho1＝0.75。
C2 otherwise, Ty> 273 or TyLess than or equal to 90, if
C2.1 when 25200 < TdWhen the value is less than or equal to 61200, taking rho1＝0.45；
C2.2 when Td61200 or TdWhen the value is less than or equal to 25200, take rho1＝0.75。
Through preliminary treatmentThe processed single-frame image size becomes 36x4, which is inferred from the definition of equation (1) as an empirical value; compared with the source image directly collected by the camera, the size is greatly reduced, so that the operation efficiency of learning in the step 2 and the step 3 is improved. According to the definition of the data packet B in the first paragraph of step 1, in combination with the single frame image size of 36x4, it is possible to obtain the original time-series data packet B for each data packet BiThe data dimension is 36 × 4 × N. Each packet covers a time length of N · T.
It will be appreciated that the above resampling is a preferred method, and may be combined with other steps in the present application, and the effect is better. Other existing methods may also be used without affecting the overall implementation of the scheme.
Step 2, the method for generating and learning the characteristics of the front-end energy consumption data obtains an original time sequence data packet B through step 1iAs an input. The user energy consumption characteristic is expressed by time-series changes of the user electricity meter reading, for example, a certain user uses more power in a specific time period of a day, and another user may use more power and consume more smoothly throughout the day. Original time series data packet BiThe energy consumption data learning method comprises information reflecting the energy consumption characteristics of the user, and the learning of the energy consumption characteristics of the user needs to be completed at the front end of the user, so that the dimensionality of the data needs to be further reduced, and the characteristics of the energy consumption data need to be further refined.
First, the user energy consumption characteristics are associated with the natural daytime time, defining TdThe number of seconds from the time of day zero at the time of sampling.
Secondly, the energy consumption characteristics of the user are related to natural seasonal time, and T is definedyThe number of the sampling time distance from the current year, namely the number of the yuan-month and a day.
Third, original time series data packet BiIncluding the power usage characteristics of a particular user. Neural network models are built herein to learn this feature.
When the neural network model is trained according to the description of step 1, the training should be carried out according to Td、TyIs normalized to the resampled image according to the definition of equation (3), where the parameter ρ is1Take a predetermined value.
The basic model of the neural network is composed of an input layer, an output layer and a hidden layer, each layer comprises a plurality of nodes called neurons, the neurons and the connection among the neurons form the neural network, and the network is determined by an excitation function, weights and the connection mode among the neurons.
Three nodes X on the leftmost side1,X21 is the input layer node, the right node y is the output layer node, h1,h2,h3For hiding layer nodes, sigma represents an excitation function and has the function of enabling the neural network to have nonlinear classification capability. The relationship between the output and the input of the neural network is defined by the following equation:
it may be a neural network with only one hidden layer, and similarly, a new hidden layer may be added to the network, and the node connection mode between layers may be freely defined.
The non-linear elements of a neural network are called the excitation function σ (x) and are used to enable the network to have the ability to classify non-linear datasets. The stimulus function is defined herein as follows:
in the above formula, x represents the input of the excitation function σ, the parameter α has the function of converging the adjustment function, and the parameter β has the function of generating a discontinuous break point at the point where x is 0, which contributes to improving the learning effect.
The loss function of the neural network training discriminant is defined as follows:
where y is the output value of the neural network,
representing the true output of the training samples, N is the number of samples. Theta1、θ2Which are independent control variables, can be selected empirically. Experiments show that the definition method is helpful for improving the accuracy of neural network discrimination.
The input data of the neural network is an original time sequence data packet BiThe dimension of (2) is 36 × 4 × N. The steps for constructing the neural network (marked as Net) are as follows:
s21, inserting the first hidden layer H1 of the network Net, and the node of the layer
The connection means is defined as follows.
H1 is based on the input data BiBy convolution kernels
The latter result. Convolution is a mathematical method of integral transformation that measures the response of a signal to a feature. The method used in this scheme is a three-dimensional convolution comprising a time dimension, as in equation (7), where p, q, r are control parameters (sliding around a central point) of a three-dimensional convolution window centered at (x, y, z),
is the weight of the corresponding position of the convolution window, b0Is a linear bias parameter. v. of0Nodes representing the input layer (being a three-dimensional matrix), i.e. BiThe superscripts x + p, y + q, and z + r respectively represent the coordinates of the node in the three dimensions of the input layer. The convolution of two dimensions except the time dimension is two-dimensional convolution; in the time dimension, the weights of the nodes with equal offset in the time window are the same.
In particular, the present invention defines the window ruler of the present layerHas a size of 3x1x9, i.e. 1. ltoreq. p.ltoreq.3, q.ltoreq.1, 1. ltoreq. r.ltoreq.9, and
s22, inserting a second hidden layer H2 of the network Net, the node of the layer
The connection means is defined as follows.
H2 is a result of finding the maximum value in a window within a range of p × q × r centered around (x, y, z) from the output data of the hidden layer H1. v. of1And representing a node of the H1 at the upper layer, and superscripts x + p, y + q and z + r respectively represent the corresponding coordinates of the node in three dimensions of the H1 layer, similar to the previous step.
In particular, the invention defines the window size of the layer as 2x2x4, i.e. 1 ≦ p, q ≦ 2, and 1 ≦ r ≦ 4.
S23, inserting a third hidden layer H3 of the network Net, and node
The connection means is defined as follows.
H3 is the output from hidden layer H2 by convolution kernel
As a result, p, q, and r are control parameters of a three-dimensional convolution window centered at (x, y, z),
is the weight of the corresponding position of the convolution window, b2Is a linear bias parameter. v. of2Representing the upper layer H2And a node, similar to the previous step, the superscripts x + p, y + q and z + r respectively represent the corresponding coordinates of the node in three dimensions of the H2 layer.
In particular, the present invention defines the window size of the layer as 3x1x3, i.e. 1 ≦ p, r ≦ 3, and q ═ 1.
S24, inserting a fourth hidden layer H4 of the network Net, wherein the node connection mode is similar to the formula (4), namely, each node of H4 and each node of H3 are connected, and the connection weights are independent. According to the window sizes of steps S21-S23, in H4, the x-dimension (corresponding to the value range of p) full-link size is: 15 in y dimension (corresponding to the value range of q), 4/2 in 2 in y dimension, and (N-9+1)/2-3+1 in z dimension (corresponding to the value range of r) in H4 node
The definition is as follows:
in the above formula, xyz is a subscript indicating H4 level nodes
All weights connected;
is a node of level H3; the expression (10) represents each H4 level node
All have connection with all nodes of H3 layer, and the connection weight is composed of
And (4) defining.
S25, and the network Net connects the output layer Y after the fourth hidden layer H4, in the form of full connection of the formula (4).
S26, defining the output layer Y as follows:
the output value is 1, which represents that the energy consumption condition in the time period corresponding to the input data is abnormal, and the output value is 0, which represents that the power consumption in the corresponding time period is normal.
During learning, a plurality of original time sequence data packets Bi are selected, the normal or abnormal energy consumption state is marked manually, and a common method is adopted to train the neural network model.
Td、TyAnd Y constitutes the characteristic of the front-end energy consumption data, denoted Fe. Wherein, Y is obtained through a neural network model, and the learning method of the neural network model is as the steps S21-S26, and the model is set when the front-end facility is out of the field. T isd、TyDerived from the front-end device clock.
And 3, the method for learning and identifying the energy consumption characteristics of the whole network is used for acquiring the energy consumption characteristic sample data of the whole network transmitted back from the remote end through the communication network on the basis of the characteristics of the front-end energy consumption data in the step 2, and further optimizing the learning of the energy consumption characteristics. When a communication network is available, step 3.1 is activated; when the communication network is not available, step 3.2 is activated and the energy consumption of the user is monitored independently at the front end according to the result of step 2. Further, the method comprises the following steps:
step 3.1When the communication network is available, the communication network obtains the energy consumption characteristic sample data of the whole network. The meaning and composition of the sample data is as follows.
Knowing that A and B are two events, the probability formula:
the probability of event a occurring when event B has occurred is called the conditional probability of event a when event B has occurred. According to the definition of the conditional probability:
equation (12) describes the relationship between the conditional probability of event a if event B occurs and the conditional probability of event B if event a occurs. In the above equation, P (a) is referred to as the prior probability of event a, and P (a | B) is referred to as the a posteriori probability of a. Bayes' theorem is the basis of the naive bayes algorithm.
Energy consumption characteristic FeComprising Td、TyAnd Y. hypothesis Td、TyAnd Y, independently of each other, defines:
wherein X represents a certain energy consumption characteristic FeThe sample of (1). Mu.s2、σ2Respectively representing the mean and variance of all samples in the training sample set. Assuming that the sample set includes K samples, from the mathematical definition of the mean and variance:
in the formula (14), i is a subscript indicating the number of the sample X.
According to the formula (12), it is possible to obtain:
c ∈ {0,1} represents a classification flag indicating whether the energy consumption state is normal or abnormal, {0,1} is defined as the definition of Y in step 2, that is, when c ═ 1, it represents that the energy consumption state in the time period corresponding to the sample is abnormal, and when c ═ 0, it represents that the energy consumption state in the time period corresponding to the sample is normal. For a given sample set, P (X) is constant, so the value of equation (15) needs only to be maximized for the molecule P (X | c) P (c). The conditional probability P (X | c) represents the probability distribution of the sample X when the sample class is c, and can be calculated according to equation (13). The probability p (c) indicates the probability that the class of a certain sample belongs to c in the whole sample set, and in this example, the samples with normal energy consumption and the samples with abnormal energy consumption respectively account for the proportion of all the samples.
When the communication network is available, the communication network obtains the full-network training value of P (c), the currently obtained values of the energy consumption characteristic samples X and c are sent to a remote cloud platform, and the value of P (X | c) is calculated according to the formula (13) and then is sent back to the front end.
The full-network training value of P (c) is calculated by the remote cloud platform according to all the data collected by the remote cloud platform, and in this case, is a vector [ P (0), P (1) ], which means:
p (0) ═ number of samples judged to be abnormal in energy consumption in the whole network/total number of samples in the whole network
P (1) ═ number of samples judged to be normal in energy consumption in the whole network/total number of samples in the whole network
And comparing the values of P (0| X) and P (1| X), if P (0| X) > P (1| X), judging that the current energy consumption is abnormal, and if not, judging that the energy consumption is normal. This classification result is also transmitted back to the remote end for reference by other front-end nodes when the communication network is available.
Step 3.2, when the communication network is not available, the definition of the formulas (11) - (15) in step 3.1 is followed, and whether the energy consumption is normal or not is judged according to the following method.
First, the value of P (c) can be pre-calculated and embedded in the front-end facility, and the pre-set value is taken when the communication network is unavailable.
Second, when the communication network is not available, μ in equation (13)2、σ2It can be pre-calculated from small-scale samples collected in advance and implanted into the front-end facility, and then P (X | c) is calculated according to equation (13).
Therefore, when the network is unavailable, P (c), P (X | c) are calculated according to the steps, the values of P (0| X) and P (1| X) are further compared, if P (0| X) > P (1| X), the current energy consumption is judged to be abnormal, and otherwise, the energy consumption is normal.
It is understood that the three steps of the present invention can be independent from each other, for example, the neural network in step 2 can also use a conventional neural network, and the implementation of step 3 is not affected. Only the neural network model of the invention can be mutually promoted and matched with the steps 1 and 3, and the effect is better.
The system comprises a plurality of functional modules for respectively implementing the steps of the method.
The test effects of the present invention are as follows
The energy consumption monitoring method based on front-end perception and learning is described, and the main purpose of monitoring the energy consumption of a user in the method is to improve the service capacity and the service level of an electric power enterprise to the user according to the electricity utilization habits of the user, for example, a competitive differentiated electricity price is formulated, and the user is reminded when the electricity utilization of the user is abnormal, and the like. It is characterized in that:
1. the system can directly complete the functions of acquisition, analysis and identification at the user side, and complete the identification and classification of the energy consumption characteristics of the user. The service ability and the service level of the power enterprise to the users can be improved according to the electricity utilization habits of the single users.
2. Compared with the common method, the unique front-end modular architecture and the deployment mode can work with the remote platform (online) or work independently (offline) from the remote platform; the off-line working has a complete learning function on the energy consumption characteristics of local users, and the on-line working can improve the learning performance.
3. The learning ability is deployed at the front end, the intelligent learning ability of the front end is improved, the network layer does not transmit user original data any more, and only transmits mapping data after feature transformation, so that the data transmission data volume of the network is greatly reduced, and the risk of leakage of user sensitive data is greatly reduced.
4. Compared with a common method, the method has better robustness, and when the front-end equipment enters a working state, the front-end equipment can independently complete an intelligent learning task even if the front end and the far end lose connection due to network failure.
5. The intelligent data perception obtaining method based on the front-end video and the images is different from a general method based on port data acquisition of an intelligent electric meter, can be used for the intelligent electric meter and a non-intelligent electric meter with a numerical value display function, is wider in application and high in identification accuracy.
Claims (7)
1. A front-end perception and learning-based energy consumption distinguishing method is characterized by comprising the following steps:
step 1: continuously shooting power values at the ammeter side by using a camera module at the front end, and acquiring an image I; after the shot image I is processed by a preprocessing method, the shot image I is packed into a plurality of batches of original time sequence data packets B by a fixed frame number NiThe reading time sequence information of the electric meter at the user end is contained;
the pretreatment comprises the following steps: the image I is re-sampled and normalized in sequence, wherein the normalization control parameter 0<ρ1<1, and it is according to a temporal characteristic Td、TyAdjusting; definition of TdThe number of seconds from the zero point of the day when sampling is carried out; definition of TyThe number of days from the first month of the year;
step 2: constructing a neural network model for the original time sequence data packet BiLearning is carried out; the neural network model consists of an input layer, an output layer and a hidden layer, wherein the loss function is as follows:
where y is the output value of the neural network,
representing the true output of the training samples, N being the number of samples; theta.theta.1、θ2Are independent control variables;
Td、Tyand y constitutes the characteristic of the front-end energy consumption data, denoted as Fe，
And step 3: transmitting the characteristics of the front-end energy consumption data to a classification discrimination model,
defining:
wherein X represents a certain energy consumption characteristicSign FeThe sample of (1); mu.s2、σ2Respectively representing the mean and the variance of all samples in the training sample set;
when c is 1, the energy consumption state in the time period corresponding to the sample is abnormal, when c is 0, the energy consumption state in the time period corresponding to the sample is normal, for a given sample set, the probability P (X) is a constant, and the probability P (c) represents that the sample with normal energy consumption and the sample with abnormal energy consumption respectively account for the proportion of all samples;
when the communication network is available, the communication network obtains a full-network training value of P (c), the currently obtained values of the energy consumption characteristic samples X and c are sent to a remote cloud platform, and the value of P (X | c) is calculated according to a formula (13) and then is returned to the front end;
when the communication network is not available, the pre-stored P (c), mu of the front end are used2、σ2Calculating the value of P (X | c) using equation (13) and equation (15);
and 4, step 4: and (3) comparing the values of P (0| X) and P (1| X), if P (0| X) > P (1| X), judging that the current energy consumption is abnormal, otherwise, judging that the energy consumption is normal.
2. The method of claim 1, wherein: the camera module at the front end continuously shoots the power value at the ammeter side, and the frame rate of the collected images is more than 1 frame/second.
3. The method of claim 1, wherein: the acquisition and pretreatment is performed at the front end.
4. The method of claim 1, wherein: finding FeThe steps of (1) are carried out at the front end.
5. The method of claim 1, wherein: the step of computing P (X | c) is implemented in the cloud platform when a network is available.
6. The method of claim 1, wherein: the step of calculating P (X | c) is implemented in the front end when the network is not available.
7. The method of any one of claims 1-6, wherein: the determining step is performed at the front end.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110770785.0A CN113420510B (en) | 2021-07-07 | 2021-07-07 | Energy consumption distinguishing method based on front-end sensing and learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110770785.0A CN113420510B (en) | 2021-07-07 | 2021-07-07 | Energy consumption distinguishing method based on front-end sensing and learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113420510A CN113420510A (en) | 2021-09-21 |
| CN113420510B true CN113420510B (en) | 2022-06-17 |
Family
ID=77720564
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110770785.0A Active CN113420510B (en) | 2021-07-07 | 2021-07-07 | Energy consumption distinguishing method based on front-end sensing and learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113420510B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015158198A1 (en) * | 2014-04-17 | 2015-10-22 | 北京泰乐德信息技术有限公司 | Fault recognition method and system based on neural network self-learning |
| CN107291664A (en) * | 2017-05-14 | 2017-10-24 | 浙江志杰电力科技有限公司 | A kind of ammeter energy consumption dynamic evaluation system and method |
| CN111680831A (en) * | 2020-05-25 | 2020-09-18 | 广东电网有限责任公司广州供电局 | Substation power utilization analysis system and method |
| WO2021133253A1 (en) * | 2019-12-25 | 2021-07-01 | Envision Digital International Pte. Ltd. | Method and apparatus for predicting energy consumption, and device and storage medium thereof |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2779527A1 (en) * | 2013-03-15 | 2014-09-17 | Yetu AG | System and method for analysing the energy consumption of electrical consumers in a network |
| CN108124487B (en) * | 2017-12-22 | 2023-04-04 | 达闼机器人股份有限公司 | Cloud meter reading method and device |
-
2021
- 2021-07-07 CN CN202110770785.0A patent/CN113420510B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015158198A1 (en) * | 2014-04-17 | 2015-10-22 | 北京泰乐德信息技术有限公司 | Fault recognition method and system based on neural network self-learning |
| CN107291664A (en) * | 2017-05-14 | 2017-10-24 | 浙江志杰电力科技有限公司 | A kind of ammeter energy consumption dynamic evaluation system and method |
| WO2021133253A1 (en) * | 2019-12-25 | 2021-07-01 | Envision Digital International Pte. Ltd. | Method and apparatus for predicting energy consumption, and device and storage medium thereof |
| CN111680831A (en) * | 2020-05-25 | 2020-09-18 | 广东电网有限责任公司广州供电局 | Substation power utilization analysis system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113420510A (en) | 2021-09-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Fekri et al. | Distributed load forecasting using smart meter data: Federated learning with Recurrent Neural Networks | |
| Fiot et al. | Electricity demand forecasting by multi-task learning | |
| Jawad et al. | Machine learning based cost effective electricity load forecasting model using correlated meteorological parameters | |
| CN110503531A (en) | The dynamic social activity scene recommended method of timing perception | |
| CN109670675B (en) | Method and device for evaluating running state of charging pile | |
| CN110069706A (en) | Method, end side equipment, cloud side apparatus and the end cloud cooperative system of data processing | |
| Tiwari et al. | Machine learning‐based model for prediction of power consumption in smart grid‐smart way towards smart city | |
| CN110796307A (en) | Distributed load prediction method and system for comprehensive energy system | |
| US20220237917A1 (en) | Video comparison method and apparatus, computer device, and storage medium | |
| CN105005852A (en) | Image analysis based intelligent monitoring system for dormitory environment | |
| Kumar et al. | Cloud-based electricity consumption analysis using neural network | |
| CN115775085A (en) | Smart city management method and system based on digital twin | |
| Sridharan et al. | Deep learning-based ensemble model for classification of photovoltaic module visual faults | |
| CN113420510B (en) | Energy consumption distinguishing method based on front-end sensing and learning | |
| Zhu et al. | Multiscale deep network based multistep prediction of high‐dimensional time series from power transmission systems | |
| CN117009785A (en) | Security monitoring method, device, server and system | |
| CN112418485A (en) | Household load prediction method and system based on load characteristics and power consumption behavior mode | |
| CN116680641A (en) | User electricity consumption abnormality detection method based on machine learning algorithm | |
| CN110751055A (en) | Intelligent manufacturing system | |
| CN110045691A (en) | A kind of multitasking fault monitoring method of multi-source heterogeneous big data | |
| Copiaco et al. | Exploring deep time-series imaging for anomaly detection of building energy consumption | |
| CN115907228A (en) | Short-term power load prediction analysis method based on PSO-LSSVM | |
| CN110210755B (en) | User demand response capability assessment method based on K _ means clustering algorithm | |
| Brander et al. | Improving Data-Scarce Image Classification Through Multimodal Synthetic Data Pretraining | |
| Zhou et al. | Mathematical model of yield forecast based on long and short-term memory image neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |