CN115603446A - Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect - Google Patents
Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect Download PDFInfo
- Publication number
- CN115603446A CN115603446A CN202210930117.4A CN202210930117A CN115603446A CN 115603446 A CN115603446 A CN 115603446A CN 202210930117 A CN202210930117 A CN 202210930117A CN 115603446 A CN115603446 A CN 115603446A
- Authority
- CN
- China
- Prior art keywords
- power distribution
- distribution station
- network
- cloud
- working condition
- 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.)
- Pending
Links
Images
Classifications
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J13/00—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network
- H02J13/00002—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by monitoring
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J13/00—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network
- H02J13/00001—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by the display of information or by user interaction, e.g. supervisory control and data acquisition systems [SCADA] or graphical user interfaces [GUI]
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J13/00—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network
- H02J13/00006—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by information or instructions transport means between the monitoring, controlling or managing units and monitored, controlled or operated power network element or electrical equipment
- H02J13/00022—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by information or instructions transport means between the monitoring, controlling or managing units and monitored, controlled or operated power network element or electrical equipment using wireless data transmission
- H02J13/00024—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network characterised by information or instructions transport means between the monitoring, controlling or managing units and monitored, controlled or operated power network element or electrical equipment using wireless data transmission by means of mobile telephony
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J13/00—Circuit arrangements for providing remote indication of network conditions, e.g. an instantaneous record of the open or closed condition of each circuitbreaker in the network; Circuit arrangements for providing remote control of switching means in a power distribution network, e.g. switching in and out of current consumers by using a pulse code signal carried by the network
- H02J13/00032—Systems characterised by the controlled or operated power network elements or equipment, the power network elements or equipment not otherwise provided for
Abstract
The invention discloses a power distribution station operation monitoring system based on a convolutional neural network and cloud edge synergistic effect, which comprises a cloud center and a plurality of edge nodes, wherein the cloud center is connected with the edge nodes; the cloud center stores a state quantity acquisition module, an abnormal working condition sample generation module, a network establishment and training module and a network optimization module; the edge node is stored with an operation condition data acquisition module, a distribution result calculation module and a diagnosis result generation module; the invention provides a power distribution station operation monitoring system based on a convolutional neural network and cloud edge synergistic effect, and solves the problems that the working condition information in a single power distribution station is difficult to train a diagnostor with strong generalization capability, the influence factors considered by the current abnormal diagnosis method are too single, the computing resources of a power distribution cloud master station are difficult to establish an abnormal working condition migration diagnosis model for each power distribution station, and the like.
Description
Technical Field
The invention relates to the field of power distribution area operation monitoring, in particular to a power distribution area operation monitoring system based on the synergistic effect of a convolutional neural network and a cloud edge.
Background
With the rapid development of information technology and the internet of things, the power grid provides a power distribution internet of things as a solution for power distribution station area services. The power distribution internet of things technology adopts a 'cloud pipe side end' framework, a cloud center is responsible for deep data mining and processing of advanced services, and an edge computing terminal is responsible for data acquisition and local processing and meets the delay requirement of real-time services.
The cloud edge cooperative system can transfer part or all of the computing tasks of the original cloud computing model to the network edge equipment, so that the pressure of the bandwidth of a cloud computing center is reduced, and the processing efficiency of mass data is improved. Therefore, research on a cloud edge cooperation technology of the power distribution and utilization service is needed, so that the processing efficiency of the power distribution and utilization edge data is improved, and the advantage of the edge computing terminal close to the data end is enhanced. The distribution transformer is used as core equipment of the power distribution network, effective running state evaluation and prediction are carried out on the distribution transformer, and the distribution transformer has important significance for building a strong power distribution network and guaranteeing safe and stable running of a power system.
With the development of power distribution networks towards the direction of information-based and intelligent power distribution internet of things (D-IoT), intelligent terminals attract the close attention of a large number of manufacturers, IT enterprises, scientific research institutions and national government departments as basic elements and core equipment of the power distribution networks. From the perspective of computing resources, the edge computing hierarchy is added on the terminal side, the localized processing of sensing data is realized, the efficient cooperation of edge computing of an end layer and big data application of a cloud layer is promoted, the overall computing capacity of the power distribution network is improved, the open architecture of the intelligent terminal can support various data acquisition requirements of future power distribution services, reduce equipment transformation cost and improve operation and maintenance efficiency, and the mode change of passive management to active management of the power distribution network can be boosted.
As the integration level and the complexity of the intelligent power distribution network system are higher and higher, the problems of safe operation and maintenance guarantee of the system are more and more prominent. For repairable complex engineering systems, timely and proper equipment maintenance is one of the important means for ensuring the safety, reliability and availability of the systems.
In recent years, with the rapid development of internet of things technology, information technology and artificial intelligence, anomaly diagnosis gradually becomes a field research hotspot. The abnormity diagnosis is realized by monitoring the running state of the equipment in real time and identifying abnormity and detecting hidden equipment danger by using a big data analysis model, so that the development trend and possible abnormal modes of the equipment are judged, decision support is provided for active operation and maintenance, and the method has the remarkable advantages of reducing the operation and maintenance cost, improving the operation efficiency and maximizing the production profit. Disclosure of Invention
The invention aims to provide a power distribution station area operation monitoring system based on the synergistic effect of a convolutional neural network and cloud edges, which comprises a cloud center and a plurality of edge nodes;
the cloud center stores a state quantity acquisition module, an abnormal working condition sample generation module, a network establishment and training module and a network optimization module;
the state quantity acquisition module acquires the state quantities of the power distribution station areas of the source domain and the target domain and transmits the state quantities to the abnormal working condition sample generation module;
the XGboost importance ranking model is stored in the abnormal working condition sample generating module;
the abnormal working condition sample generation module sorts the state quantities of the power distribution station area by using an XGboost importance sorting model so as to form a source domain abnormal working condition label sample and a target domain abnormal working condition sample;
the network establishing and training module establishes a freezing layer network and a source domain full-connection classifier, performs supervised training on freezing layer network structure parameters and weight parameters of the source domain full-connection classifier by using a source domain abnormal working condition label sample, and issues the trained parameters to edge nodes;
the network optimization module receives the freezing layer network characteristic distribution and the classifier result distribution of the target domain sent by the edge node, optimizes the freezing layer network by using a gradient descent algorithm, minimizes the difference between the source domain abnormal working condition sample and the target domain abnormal working condition sample, obtains a migration diagnosis model and sends the migration diagnosis model to the edge node;
the edge node is stored with an operation condition data acquisition module, a distribution result calculation module and a diagnosis result generation module;
the operation condition data acquisition module acquires operation condition data in real time and transmits the operation condition data to the diagnosis result generation module;
the distribution result calculation module reconstructs a freezing layer network and an independent target domain full-connection classifier, inputs the abnormal working condition sample of the target domain into the reconstructed freezing layer network and target domain full-connection classifier, obtains the freezing layer network characteristic distribution and the classifier result distribution of the target domain, and uploads the freezing layer network characteristic distribution and the classifier result distribution to the cloud center;
the diagnostic result generation module receives and stores the migration diagnostic model;
and the diagnostic result generation module inputs the operation condition data into the migration diagnostic model to obtain the operation monitoring result of the power distribution area and uploads the operation monitoring result to the cloud center.
Further, the power distribution station area state quantity comprises environmental data, electrical data, statistical data and parameter data;
the environmental data comprises meteorological reference characteristics of an area where the power distribution station area is located;
the electrical data comprises measurement characteristics of the power distribution station during operation;
the parameter data comprises rated characteristics of the power distribution station area and comprehensive equipment thereof;
the statistical data is obtained by calculating the electrical data and the parametric data in real time.
Further, the source domain includes a plurality of power distribution areas similar to the target domain power distribution areas;
similar judgment criteria to the target area distribution station area include: the difference of the operation years of the target domain power distribution station and the source domain power distribution station is not more than h1 year, the equipment types and the models are the same, the difference of the total electric loads of the platform area is not more than h2%, and the difference of the power supply radius is not more than h3%. h1, h2 and h3 are constants.
Further, the edge node is a node containing state quantity of a power distribution area of the target domain; and the edge nodes and the cloud center carry out data interaction through a network.
Further, the XGBoost importance ranking model is as follows:
wherein N is the total number of trees;
is a base learner; serial number m =1, …, N; t is a node on the decision tree, Δ Gini (t) is a variation value of the kini coefficient on the t node; vim (X) j ) Represents the state quantity X j Degree of influence on abnormal conditions.
Further, the freezing layer network is a convolution neural network;
the convolution operation and activation function operation formulas of the frozen layer network are as follows:
in the formula (I), the compound is shown in the specification,
and
is the jth characteristic state quantity of the ith layer and the ith characteristic state quantity of the (l-1) th layer;
is a convolution kernel weight matrix between the ith characteristic state quantity of the l-1 layer and the jth characteristic state quantity of the l layer;
is the bias term corresponding to the jth characteristic state quantity of the ith layer; f (x) is the activation function.
Further, the maximum pooling calculation formula of the frozen layer network is as follows:
in the formula (I), the compound is shown in the specification,
is the characteristic state quantity of the l-th layer in the pooling calculation, and m × m is the local area covered by the pooling kernel; max (x) is the pooling function.
Is the characteristic state quantity of the l-1 st layer in the pooling calculation.
Further, output X of l layer of source domain full-connection classifier l As follows:
X l ＝f(∑X l-1 w l,l-1 +b l ) (4)
in the formula, X l-1 Is the output value of layer l-1; w is a l,l-1 Is a weight matrix between layer l-1 and layer l; b l Is the bias term corresponding to the l layer; f (x) is an activation functionAnd (4) counting.
Further, the objective function of the cloud center for optimizing the frozen layer network is as follows:
in the formula (I), the compound is shown in the specification,
representing source domain abnormal condition label samples;
representing a target domain abnormal condition label sample; m and n respectively represent the number of the source domain abnormal working condition label samples and the target domain abnormal working condition label samples; λ is a penalty term; m represents the number of source domain samples; k' represents the number of categories; p is a radical of formula ik Representing the predicted probability that sample i belongs to class k. Y is ik Is a symbolic function, if the true class of sample i is equal to k, then Y ik Taking 1, otherwise, taking 0; k (x) is a kernel function;
wherein the penalty term λ is as follows:
wherein p is the ratio of the current training times to the total training times.
Further, the cloud center also comprises a power distribution station operation monitoring visualization module;
the power distribution station operation monitoring visualization module is used for displaying a power distribution station operation monitoring result. And the power distribution station area operation monitoring result comprises normality and abnormality.
It is worth to be noted that the invention provides a power distribution station operation monitoring system based on a convolutional neural network and cloud-side synergistic effect aiming at the defects of the existing power distribution station abnormal diagnosis method and software and hardware resources and combining the distribution characteristics of the power distribution station, so as to solve the problems that a diagnostor with strong generalization capability is difficult to train the working condition information in a single power distribution station, the influence factor considered by the current abnormal diagnosis method is too single, and the calculation resources of a power distribution cloud master station are difficult to realize the establishment of an abnormal working condition migration diagnosis model for each power distribution station, and the like. Firstly, screening state quantities which have large influences on abnormal working conditions by using an importance sorting method based on XGboost in a cloud center to form an abnormal working condition label sample; and collecting abnormal working condition label samples of a plurality of similar power distribution station areas, and training and constructing a convolutional neural network model for diagnosing the abnormal working conditions of the source domain. Secondly, migrating the source domain diagnosis model to a single distribution station area edge node of a target domain, performing difference training on the target domain by using a migration mechanism, calculating the distribution difference between the source domain and the target domain by introducing the multi-core maximum mean difference, and constructing a target domain optimization loss function to enable the target domain and the source domain to be in self-adaptive matching, thereby effectively establishing a target domain abnormal working condition diagnosis model and exploring a feasible novel scheme for the application of an abnormal working condition diagnosis method.
The technical effect of the invention is undoubted, and the invention provides a power distribution station area operation monitoring system based on the synergistic effect of the convolutional neural network and the cloud edge, so as to solve the problems that a diagnostor with strong generalization capability is difficult to train out the working condition information in a single power distribution station area, the influence factor considered by the current abnormal diagnosis method is too single, the calculation resource of a power distribution cloud master station is difficult to realize the establishment of an abnormal working condition migration diagnosis model for each power distribution station area, and the like, and explore a feasible novel scheme for the application of the abnormal working condition diagnosis method.
Drawings
FIG. 1 is a diagram of a diagnostic architecture based on cloud-edge synergy;
FIG. 2 is a technical route flow diagram;
FIG. 3 is a state quantity importance ranking chart under heavy overload conditions;
FIG. 4 is a diagram of the results of heavy overload anomaly diagnosis for different quantities of state;
FIG. 5 is a diagram of a deep migration network diagnostic model architecture;
FIG. 6 is a flow chart of size parameter and weight issuance;
FIG. 7 is a flow chart of obtaining parameters and weights;
FIG. 8 is a graph of the results of heavy overload anomaly diagnosis for different penalty term coefficients;
FIG. 9 is a visualization of the characteristics of the diagnosis of the present invention;
FIG. 10 is a flow diagram of model triggered update;
FIG. 11 is a functional block diagram of a system;
FIG. 12 is a system architecture diagram;
FIG. 13 is a system home interface diagram.
Detailed Description
The present invention is further illustrated by the following examples, but it should not be construed that the scope of the above-described subject matter is limited to the following examples. Various substitutions and alterations can be made without departing from the technical idea of the invention and the scope of the invention is covered by the present invention according to the common technical knowledge and the conventional means in the field.
Example 1:
referring to fig. 1 to 13, the power distribution station operation monitoring system based on the convolution neural network and the cloud edge synergy includes a cloud center and a plurality of edge nodes;
the cloud center stores a state quantity acquisition module, an abnormal working condition sample generation module, a network establishment and training module, a network optimization module and a power distribution station operation monitoring visualization module;
the state quantity acquisition module acquires the state quantities of the power distribution station areas of the source domain and the target domain and transmits the state quantities to the abnormal working condition sample generation module;
the abnormal working condition sample generation module stores an XGboost importance ranking model;
the abnormal working condition sample generation module sorts the state quantities of the power distribution station area by using an XGboost importance sorting model so as to form a source domain abnormal working condition label sample and a target domain abnormal working condition sample;
the network establishing and training module establishes a freezing layer network and a source domain full-connection classifier, performs supervised training on freezing layer network structure parameters and weight parameters of the source domain full-connection classifier by using a source domain abnormal working condition label sample, and issues the trained parameters to edge nodes;
the network optimization module receives the freezing layer network characteristic distribution and the classifier result distribution of the target domain sent by the edge node, optimizes the freezing layer network by using a gradient descent algorithm, minimizes the difference between the source domain abnormal working condition sample and the target domain abnormal working condition sample, obtains a migration diagnosis model and sends the migration diagnosis model to the edge node;
the edge node is stored with an operation condition data acquisition module, a distribution result calculation module and a diagnosis result generation module;
the operation condition data acquisition module acquires operation condition data in real time and transmits the operation condition data to the diagnosis result generation module;
the distribution result calculation module reconstructs a freezing layer network and an independent target domain full-connection classifier, inputs the abnormal working condition sample of the target domain into the reconstructed freezing layer network and target domain full-connection classifier, obtains the freezing layer network characteristic distribution and the classifier result distribution of the target domain, and uploads the freezing layer network characteristic distribution and the classifier result distribution to the cloud center;
the diagnostic result generation module receives and stores the migration diagnostic model;
and the diagnostic result generation module inputs the operation condition data into the migration diagnostic model to obtain the operation monitoring result of the power distribution area and uploads the operation monitoring result to the cloud center.
The state quantity of the power distribution station area comprises environmental data, electrical data, statistical data and parameter data;
the environmental data comprises meteorological reference characteristics of an area where the power distribution station area is located;
the electrical data comprises measurement characteristics of the power distribution station during operation;
the parameter data comprises rated characteristics of the power distribution station area and comprehensive equipment thereof;
the statistical data is obtained by calculating the electrical data and the parametric data in real time.
The source domain comprises a plurality of power distribution areas similar to the target domain power distribution areas;
the judgment criteria similar to the target area distribution station area include: the difference of the operation years of the target domain power distribution station and the source domain power distribution station is not more than h1 year, the equipment types and the models are the same, the difference of the total electric loads of the platform area is not more than h2%, and the difference of the power supply radiuses is not more than h3%.
The edge node is a node containing state quantity of a power distribution area of the target domain; and the edge nodes and the cloud center perform data interaction through a network.
The XGboost importance ranking model is as follows:
wherein N is the total number of trees;
is a base learner; serial number m =1, …, N; t is a node on the decision tree, Δ Gini (t) is a variation value of the kini coefficient on the t node; vim (X) j ) Represents the state quantity X j The degree of influence on abnormal conditions.
The freezing layer network is a convolution neural network;
the convolution operation and activation function operation formulas of the frozen layer network are as follows:
in the formula (I), the compound is shown in the specification,
and
is the jth characteristic state quantity of the ith layer and the ith characteristic state quantity of the (l-1) th layer;
is a convolution kernel weight matrix between the ith characteristic state quantity of the l-1 layer and the jth characteristic state quantity of the l layer;
is the bias term corresponding to the jth characteristic state quantity of the ith layer; f (x) is the activation function.
The maximum pooling calculation for the frozen layer network is shown below:
in the formula (I), the compound is shown in the specification,
is the ith characteristic state quantity of the ith layer in the pooling calculation, and m × m is the local area covered by the pooling kernel; max (x) is the pooling function.
Is the characteristic state quantity of the l-1 st layer in the pooling calculation.
Output X of l layer of source domain full-connection classifier l As follows:
X l ＝f(∑X l-1 w l,l-1 +b l ) (4)
in the formula, X l-1 Is the output value of layer l-1; w is a l,l-1 Is a weight matrix between layer l-1 and layer l; b is a mixture of l Is the bias term corresponding to the l layer; f (x) is the activation function.
The objective function of the cloud center for optimizing the frozen layer network is as follows:
in the formula (I), the compound is shown in the specification,
respectively representing a source domain abnormal working condition label sample and a target domain abnormal working condition label sample; l is cls (Y s ) Is an anomaly diagnostic error of the source domain diagnostic model;
the multi-core maximum mean difference of the l-th layer characteristics of the source domain sample and the target domain sample; l. the 1 ，l 2 The starting layer and the terminating layer which need to compare the distribution difference in the network; λ is a penalty term; m represents the number of source domain samples; k represents the number of categories; p is a radical of ik Representing the predicted probability that sample i belongs to class k;
wherein the penalty term λ is as follows:
wherein p is the ratio of the current training times to the total training times.
The power distribution station operation monitoring visualization module is used for displaying a power distribution station operation monitoring result. And the power distribution station area operation monitoring result comprises normality and abnormality.
Example 2:
the power distribution station operation monitoring system based on the convolution neural network and the cloud edge synergistic effect comprises a cloud center and a plurality of edge nodes;
the cloud center stores a state quantity acquisition module, an abnormal working condition sample generation module, a network establishment and training module, a network optimization module and a power distribution station operation monitoring visualization module;
the state quantity acquisition module acquires the state quantities of the power distribution areas of the source domain and the target domain and transmits the state quantities to the abnormal working condition sample generation module;
the abnormal working condition sample generation module stores an XGboost importance ranking model;
the abnormal working condition sample generation module sorts the state quantities of the power distribution station area by using an XGboost importance sorting model, so that a source domain abnormal working condition label sample and a target domain abnormal working condition sample are formed;
the network establishing and training module establishes a freezing layer network and a source domain full-connection classifier, performs supervised training on freezing layer network structure parameters and weight parameters of the source domain full-connection classifier by using a source domain abnormal working condition label sample, and issues the trained parameters to edge nodes;
the network optimization module receives the freezing layer network characteristic distribution and the classifier result distribution of the target domain sent by the edge node, optimizes the freezing layer network by using a gradient descent algorithm, minimizes the difference between the source domain abnormal working condition sample and the target domain abnormal working condition sample, obtains a migration diagnosis model and sends the migration diagnosis model to the edge node;
the edge node is stored with an operation condition data acquisition module, a distribution result calculation module and a diagnosis result generation module;
the operation condition data acquisition module acquires operation condition data in real time and transmits the operation condition data to the diagnosis result generation module;
the distribution result calculation module reconstructs a freezing layer network and an independent target domain full-connection classifier, inputs the abnormal working condition sample of the target domain into the reconstructed freezing layer network and target domain full-connection classifier, obtains the freezing layer network characteristic distribution and the classifier result distribution of the target domain, and uploads the freezing layer network characteristic distribution and the classifier result distribution to the cloud center;
the diagnostic result generation module receives and stores the migration diagnostic model;
and the diagnostic result generation module inputs the operation condition data into the migration diagnostic model to obtain the operation monitoring result of the power distribution area and uploads the operation monitoring result to the cloud center.
The power distribution station operation monitoring visualization module is used for displaying a power distribution station operation monitoring result. And the monitoring result of the operation of the power distribution station area comprises normality and abnormality.
The method for using the power distribution station operation monitoring system based on the convolution neural network and cloud edge synergy comprises the following steps:
1) Acquiring the state quantities of the plurality of power distribution areas;
each abnormal working condition of the power distribution station area can be influenced by various state quantities, and the state quantity types can be divided into environmental data, electrical data, statistical data and parametric data according to state characteristics. The environmental data is a meteorological reference characteristic of an area where the power distribution station area is located and can be obtained through a temperature sensor, a humidity sensor and a timer in the power distribution station area; the electrical data is the measurement characteristics of the power distribution station during operation and can be collected in real time through an ammeter; the parameter data is rated characteristics of the power distribution area and comprehensive equipment thereof and can be acquired through equipment model query and the like; the statistical data can be obtained in real time through calculation of the electrical data and the parameter data. The updating period of the state quantity except the parameter data is short, the updating is generally carried out once in an hour, and different abnormal working conditions are influenced by different state quantities.
In order to effectively avoid negative migration caused by too large difference of characteristics, a plurality of power distribution station areas with large amount of historical data, which are similar to the environmental conditions of the power distribution station areas of the target domain and have similar equipment parameters, are selected as source domains.
The selection standard of the similar platform areas is set to be that the difference of the operation years is not more than 2 years, the types and models of equipment such as a distribution transformer, a low-voltage cable and a low-voltage branch box are the same, the total electric load difference of the platform areas is not more than 10%, and the power supply radius difference is not more than 5%.
Under the abnormal working condition in a real scene, factors such as the continuous influence with a time dimension, an accumulative effect, frequency and the like need to be considered, so that a sliding time window is set, time sequences of sampling points at the current moment and m past moments are recorded, and a certain state quantity at the current moment is represented by the sampling points corresponding to the sliding time window at the current moment. The source recording domain has n data sampling points, and a certain state quantity at the current moment consists of m +1 data sampling points and is recorded as X i ＝[x i T ,x i-1 T ,…,x i-m T ] T 。
2) The cloud center screens state quantities which have large influences on certain abnormal working conditions by using an XGboost importance sorting method based on degree labels of the abnormal working conditions of the source domain to form a source domain abnormal working condition label sample.
After abnormal working condition data of similar station areas are collected, considering that noise data with some state quantities possibly becoming useless is fitted by a model, the precision of classification results is reduced,therefore, the critical state quantity screening treatment is required to be carried out aiming at a certain abnormal working condition. An XGboost algorithm with an integration strategy is selected to establish a model, a feature Importance screening method Null import is introduced into feature engineering, the contribution of each feature on each tree in the XGboost is judged, then an average value is taken, and finally the contribution of the features is compared. The contribution is calculated by using the Gini index shown in formula (1) and a certain state quantity X j The importance index is calculated using equation (2).
Wherein: k denotes K classes, p k Representing the sample weight of class k.
The contributions made on each tree are averaged and then the contribution sizes between features are compared, some state quantity X j The importance index is calculated using equation (2).
Wherein N is the total number of trees;
is a base learner in the model; t is a node on the decision tree, and Δ Gini (t) is a change value of the kini coefficient on the t node after data is replaced. Vim (X) j ) The greater the importance of the state quantity, the more significant the effect on the abnormal condition.
3) The cloud center constructs a convolutional neural network and a source domain full-connection classifier, and the convolutional neural network structure and parameters of the cloud center and the edge nodes are the same, so that the convolutional neural network is marked as a frozen layer network.
The cloud center constructs a freezing layer network based on the source domain samples, and in the freezing layer network, the input of a convolutional neural network is an abnormal working condition sample X of the source domain s And target Domain Exception Condition sample X t N sets of samples are counted. Convolutional layer energy sampleThe learning obtains implicit performance state characteristic knowledge, and different convolution kernels and input samples are subjected to convolution operation to obtain different characterization characteristics. The convolution operation and the activation function operation are shown in equation (3):
in the formula:
is the ith characteristic state quantity of the l-1 st layer;
is a convolution kernel weight matrix between the ith characteristic state quantity of the l-1 layer and the jth characteristic state quantity of the l layer;
is the bias term corresponding to the jth characteristic state quantity of the ith layer; f (x) is the activation function.
And after the convolution operation, performing pooling layer calculation, connecting the neurons in the pooling layer with the local regions of the neurons in the previous layer, and obtaining the statistical characteristic value of the local connecting regions through calculation, wherein the size of the local regions is determined by a pooling kernel. The essence is to select some way to perform dimensionality reduction compression on the previous layer to accelerate the operation speed. The maximum pooling calculation is used herein, and the maximum value of the local connection area at a certain position is used to replace the output of the previous layer network at the position, and the maximum pooling calculation formula is shown as (4):
in the formula:
is the ith characteristic state quantity of the ith layer in the pooling calculation, m × m is the local area covered by the pooling kernel, max (x) is the pooling function, and the maximum in the coverage area of the pooling kernel is calculatedA large value.
4) The cloud center inputs training data of a source domain sample into a freezing layer network and a source domain full-connection classifier, supervised training is carried out on weight parameters of the freezing layer network and the source domain full-connection classifier by using a back propagation algorithm, the network characteristic distribution and the classifier result distribution of the source domain freezing layer are obtained, and the freezing layer network structure parameters and the weight parameters are issued to edge nodes containing a target domain sample.
After convolution and pooling operation for many times, the freezing layer network learns the characteristics implied in the signals from the input samples; and inputting the characteristics into a full connection layer, integrating the characteristics of a higher layer and diagnosing the working condition. In the fully-connected layer, the neurons between the network layers are all connected, and the calculation formula is shown as (5):
X l ＝f(∑X l-1 w l,l-1 +b l ) (5)
in the formula, X l-1 Is the output value of layer l-1; w is a l,l-1 Is a weight matrix between layer l-1 and layer l; b l Is the bias term corresponding to the l layer; f (x) is the activation function.
After extracting the size parameters and the weights of each layer of neural network, the cloud end is connected with the intelligent fusion terminal of the edge node through an MQTT protocol, and the size parameters and the weights of each layer of neural network are issued to the intelligent fusion terminal in a JSON format.
5) The method comprises the steps that an edge node reconstructs a freezing layer network and an independent target domain full-connection classifier, training data of a target domain sample are input into the reconstructed freezing layer network and the target domain full-connection classifier, and freezing layer network feature distribution and classifier result distribution of a target domain are obtained and uploaded to a cloud center;
6) The cloud center calculates the classification distribution difference, the feature distribution difference, the source domain classifier discrimination loss and the penalty coefficient of a source domain and a target domain, minimizes a target loss function by using a gradient descent algorithm, optimizes the network parameters of a freezing layer and enables the source domain data and the target domain data to be automatically matched as much as possible;
in order to migrate the model to the target domain, the model is trained by the labeled sample of the source domainAnd measuring and optimizing the characteristic probability distribution and the classification result distribution of the source domain abnormal working condition sample and the target domain abnormal working condition sample. The Multi-kernel Maximum Mean difference (MK-MMD) is an extension of the Maximum Mean Difference (MMD), and the MMD is used to determine whether the two distributions are the same, and the specific operation is to map the data in the source domain and the target domain into the regenerated kernel hilbert space, and then calculate the distance between the two domain data means. Given a source domain data distribution of s and a data sample set of
The target domain data distribution is t, the data sample set is
By using the MMD definition between the functions φ (·), s and t as shown in equation (6), the data is mapped into reproducible Hilbert space for maximum mean difference approximation measures as follows:
in the formula: e s [·]Represents the mathematical expectation of the source domain distribution, | φ | H ≦ 1 represents a series of functions within a unit sphere in the regenerative kernel Hilbert space H, | (. Cndot.) is a mapping function for a single kernel k; | | non-woven hair H Is a renewable hilbert space.
However, the existing MMD method is based on single-core transformation, and the multi-core MMD assumes that the optimal core can be linearly combined by a plurality of cores. Let H k Representing a regenerated kernel hilbert space with a unique kernel K, in which MK-MMD is represented as shown in equation (8), a mapping function phi (·) is associated with the unique kernel K, and a K-kernel function is defined as a convex combination of m kernels, as shown in equation (9).
In which distribution s is in H k Is an independent element u k (s)，β u In order to ensure that the resulting multi-core K is unique.
During model training, in order to minimize the difference of the feature probability distribution and the distribution difference of the classification result of the source domain abnormal condition sample and the target domain abnormal condition sample, a target loss function is established as formula (10).
In the formula: l is cls (Y s ) Is an anomaly diagnostic error of the source domain diagnostic model;
the multi-core maximum mean difference of the l-th layer characteristics of the source domain sample and the target domain sample; l 1 ，l 2 The starting layer and the terminating layer which need to compare the distribution difference in the network; and lambda is a penalty term which controls the abnormal working condition diagnosis recognition rate of the model on the source domain sample and the weight adjusted according to a specific target domain.
Anomaly diagnostic error L of source domain diagnostic model cls (Y s ) The cross entropy loss function frequently used in the classification problem is selected for quantization, and the cross entropy loss function of multi-classification is shown as formula (11):
in the formula: m represents the number of source domain samples; k represents the number of categories; y is ik Symbolic function(0 or 1); if the real class of the sample i is equal to k, 1 is taken, otherwise 0 is taken; p is a radical of ik The predicted probability that sample i belongs to class k.
In order to prevent the migration training overfitting and obtain a better migration fault diagnosis result, a penalty term change coefficient lambda related to the iteration number is set as shown in formula (12):
in the formula: p is the ratio of the current training times to the total training times.
According to the measurement method of the multi-kernel maximum mean difference, the whole objective loss function can be expressed as shown in formula (13):
7) Repeating 4) -6) until a given iteration number is reached, and finally outputting an optimized migration diagnosis model;
8) After the migration diagnosis model is well deployed at the edge node, the degree of the abnormal working condition is diagnosed according to the real-time collected operation working condition data. And (4) re-executing 2) to 7) until the database triggers the model to be updated.
9) And uploading the real-time diagnosis result to the cloud center by the edge node, and constructing a power distribution station operation monitoring visualization system by the cloud center according to the real-time diagnosis result.
Example 3:
the main content of a power distribution station operation monitoring system based on the synergistic effect of a convolutional neural network and a cloud edge is shown in an embodiment 1-2, and a using method of the power distribution station operation monitoring system comprises the following steps:
1) Acquiring the state quantities of the plurality of power distribution areas; table 1 lists the state quantities affecting each abnormal operating condition of the distribution substation.
TABLE 1 State variables influencing the abnormal operating conditions of the distribution area
The method comprises the steps of adopting a part of distribution transformer areas in Chongqing cities of the first year in 2017 as analysis objects, selecting 10 similar transformer areas with the difference of not more than 10% in total electrical load and the difference of not more than 5% in power supply radius in the platform area, and carrying out a migration diagnosis experiment of abnormal working conditions under cloud-edge cooperation by selecting the distribution transformer areas with the difference of not more than 2 years in operation years, the same equipment types and models of a distribution transformer, a low-voltage cable, a low-voltage branch box and the like and 43187 data in total. Wherein 8 groups of similar station area monitoring data are used as a source domain A, and 2 groups of station area data are simultaneously selected as a target domain B and a target domain C. And carrying out a migration diagnosis experiment of abnormal working conditions under cloud edge coordination. Wherein 8 groups of similar station area monitoring data are used as a source domain A, and 2 groups of station area data are simultaneously selected as a target domain B and a target domain C. In order to verify the method provided by the text, the refined diagnosis of each abnormal working condition is realized, and abnormal working condition migration diagnosis experiments shown in table 2 are performed twice on a heavy overload working condition scene, namely experiment A → B and experiment A → C. In each abnormal working condition migration diagnosis experiment, a data set on the left side of an arrow represents a source domain data set, and a data set on the right side represents a target domain data set.
TABLE 2 migration diagnostic test for abnormal conditions
2) The cloud center screens state quantities which have large influences on the heavy overload abnormal working conditions by using an XGboost importance sorting method based on the degree label of the source domain heavy overload abnormal working conditions to form a source domain heavy overload abnormal working condition label sample;
the degree of heavy overload conditions in the source domain data set is classified into three categories, including normal, abnormal and severe. The sliding time window is set to 24 hours, i.e., each state quantity consists of a sampling point at the current time combined with a sampling point at the past 23 hours. According to the feature importance calculated by the XGboost algorithm, the iteration times are 200 times, and the state quantity with the top 20 of the importance ranking is selected, as shown in FIG. 3.
Selecting the state quantity with the top importance ranking 6 as an initial training set, gradually adding the rest state quantities into the training of the migration diagnosis model, comparing the effects of the diagnosis model in the process of increasing the state quantity of the training set from 7 to 20, repeating the migration diagnosis experiments of A → B and A → C10 times each time the state quantity is increased, and the experimental result is shown in FIG. 4. In the process of adding the state quantities one by one, the diagnostic model trained by using the first 15 state quantities has the best effect, so that the first 15 state quantities are selected to carry out the heavy overload working condition migration diagnostic experiment. The state quantities of the top 15 important characteristics are load factor, zero sequence current, temperature, phase C current, bus current, phase B current, phase C voltage, season, phase A current, bus voltage, daily total active electric quantity, phase B voltage, phase B active power, distribution transformation capacity and current reverse sequence.
3) The cloud center constructs a convolutional neural network and a source domain full-connection classifier, and the convolutional neural network is recorded as a freezing layer network because the convolutional neural network structure and the parameters of the cloud center and the edge nodes are the same.
The screened key state quantities are combined into a source domain abnormal working condition sample, and a deep migration network diagnosis model structure is designed as shown in fig. 5. The network structure of the model consists of a frozen layer network f and two independent fully connected classifiers y. The freezing layer network f consists of a plurality of convolution layers and a maximum pooling layer and can directly process input abnormal working condition samples. The two fully connected classifiers are respectively composed of fully connected layers.
And (3) performing migration diagnosis of the overload abnormal degree by applying the method aiming at the target domains B and C, and arranging a 2-layer convolution pooling layer and a 1-layer full-connection layer in a freezing layer network. The fully-connected classifier of the source domain is set to be 1 layer of fully-connected layers, the size of a convolution kernel is 2 multiplied by 2, the step is 1, the pooling size is 2 multiplied by 1, and the step is 1.
4) The cloud center inputs training data of a source domain sample into a freezing layer network and a source domain full-connection classifier, supervised training is carried out on weight parameters of the freezing layer network and the source domain full-connection classifier by using a back propagation algorithm, the network characteristic distribution and the classifier result distribution of the source domain freezing layer are obtained, and the freezing layer network structure parameters and the weight parameters are issued to edge nodes containing a target domain sample;
after extracting the size parameters and the weights of each layer of neural network, the cloud end is connected with the intelligent fusion terminal of the edge node through the MQTT protocol, and the size parameters and the weights of each layer of neural network are issued to the intelligent fusion terminal in the JSON format, wherein a size parameter and weight issuing flow chart is shown in FIG. 6.
5) The edge node reconstructs a freezing layer network and an independent target domain full-connection classifier, inputs training data of a target domain sample into the reconstructed freezing layer network and the target domain full-connection classifier, acquires freezing layer network characteristic distribution and classifier result distribution of the target domain and uploads the characteristics and the classifier result distribution to the cloud center;
and (3) carrying out migration diagnosis on the overload abnormal degree by applying the method aiming at the target domains B and C, and setting a 2-layer convolution pooling layer and a 1-layer full-connection layer on the edge node reconstruction freezing layer network. The full-connected classifier of the target domain is set to be 1 full-connected layer, the size of the convolution kernel is 2 x 2, the step is 1, the pooling size is 2 x 1, and the step is 1.
The intelligent fusion terminal subscribes topic sent by cloud data, and after successful subscription, sends a data sending request containing the equipment ID and the corresponding topic in a JSON format to obtain the source domain diagnosis model parameters and weight of the cloud, and the flow chart is shown in FIG. 7.
6) The cloud center calculates the classification distribution difference, the feature distribution difference, the source domain classifier discrimination loss and the penalty coefficient of the source domain and the target domain, minimizes a target loss function by using a gradient descent algorithm, optimizes the network parameters of a freezing layer and enables the source domain and the target domain to be automatically matched as far as possible;
in order to migrate the model to the target domain, when model training is carried out through the source domain labeled samples, measurement and optimization are carried out on the source domain abnormal working condition samples, the target domain abnormal working condition samples, feature probability distribution and classification result distribution. The multinuclear Maximum Mean difference (MK-MMD) is an extension of the Maximum Mean Difference (MMD), and the MMD is used to determine whether the two distributions are the same, and specifically, the data in the source domain and the target domain are mapped into a regenerated nuclear hilbert space, and then the distance between the Mean values of the data in the two domains is calculated.
During model training, in order to minimize the difference of the feature probability distribution and the distribution difference of the classification result of the source domain abnormal condition sample and the target domain abnormal condition sample, a target loss function is established as formula (10).
In the formula: l is cls (Y s ) Is an anomaly diagnostic error of the source domain diagnostic model;
the multi-kernel maximum mean difference of the l-th layer characteristics of the source domain sample and the target domain sample; l 1 ，l 2 The starting layer and the terminating layer which need to compare the distribution difference in the network; and lambda is a penalty term which controls the abnormal working condition diagnosis recognition rate of the model on the source domain sample and the weight adjusted according to a specific target domain.
Anomaly diagnostic error L of source domain diagnostic model cls (Y s ) The cross entropy loss function frequently used in the classification problem is selected for quantization, and the cross entropy loss function of multi-classification is shown as formula (11):
in the formula: m represents the number of source domain samples; k represents the number of categories; y is ik Sign function (0 or 1); if the real class of the sample i is equal to k, 1 is taken, otherwise 0 is taken; p is a radical of ik The predicted probability that sample i belongs to class k.
In order to prevent the migration training overfitting and obtain a better migration fault diagnosis result, a penalty term change coefficient lambda related to the iteration number is set as shown in formula (12):
in the formula: p is the ratio of the current training times to the total training times.
According to the measurement method of the multi-kernel maximum mean difference, the whole objective loss function can be expressed as shown in formula (13):
the penalty term [0.2,0.5,1,10] is set to be compared with the change coefficient of the penalty term, the migration diagnosis experiments of A → B and A → C are repeated for 10 times, and the experimental result is shown in FIG. 8. Among all the penalty term coefficients, changing the penalty term coefficients achieves the best diagnostic result.
7) Repeating 3) -5) until a given iteration number is reached, and finally outputting an optimized migration diagnosis model;
the migration diagnosis precision of the system provided by the invention is higher than that of a deep convolutional neural network without a migration process, and the measurement of the difference of the characteristics and the classification distribution of the source domain sample and the target domain sample is proved to be matched with the migrated sample characteristics, so that the diagnosis result of a cross-platform interval is effectively improved; the system provided by the invention obtains a better migration diagnosis result, can effectively reduce the distribution difference of the source domain and the target domain, and obtains better domain adaptation capability and classification performance.
The learned features of the invention are visualized by using the t-SNE technology, and the feature distribution of the source domain data and the target domain data is visualized in a two-dimensional space, as shown in FIG. 9. Where "S" represents source domain data and "T" represents target domain data. 'S-1' represents a source domain normal category, 'S-2' represents a source domain abnormal category, 'S-3' represents a source domain severe category, and 'T-1' represents a target domain normal category, 'T-2' represents a target domain abnormal category, 'T-3' represents a target domain severe category.
As can be seen from FIG. 9, on the single source domain task of heavy overload anomaly diagnosis (S-1, S-2 and S-3), the strong feature learning capability of the present invention: the same abnormal degree can be clustered well, different abnormal degrees can be distinguished obviously, and samples are dispersed in 3 categories and have less category region overlapping.
8) After the migration diagnosis model is well deployed at the edge node, the degree of the abnormal working condition is diagnosed according to the real-time collected operation working condition data. And (4) re-executing 2) to 7) until the database triggers the model to be updated. After the database receives a certain amount of new feature data, the model updating mechanism is triggered, the diagnostic model is retrained, and a flowchart of the updating mechanism is shown in fig. 10.
9) And uploading the real-time diagnosis result to the cloud center by the edge node, and constructing a power distribution station operation monitoring visualization system by the cloud center according to the real-time diagnosis result. The system function module is shown in fig. 11, the system architecture diagram is shown in fig. 12, the system main interface is shown in fig. 13, and the system mainly has the following functions:
9.1 Short message login: after a user inputs a mobile phone number, a login verification code is obtained, and the verification code is input at a set time to complete login; after inputting the mobile phone number, 6 digits are randomly generated and temporarily stored in a Redis database, expiration time is set to be 60s, a short message with 6 digits is sent to the mobile phone number through the Aliyun short message service, and a user can successfully log in after inputting correct digits in 60s after receiving the short message.
9.2 Station area management: the user can manage the setting of the distribution area, including adding, modifying, deleting, inquiring and controlling the online and offline of the distribution area;
9.3 Online platform area management: a user can check detail data of an online platform area, wherein the detail data comprises latest diagnosis results of three-phase imbalance, heavy overload, line loss and low voltage, latest measurement results of temperature, humidity and the like and historical results;
9.4 User list management, the system records the existing user list, can lock the existing user and can inquire the registered user within a certain time;
9.5 User details management: the system can view details of existing users, including mobile phone numbers, registration time, certificate numbers, groups, and certificate photos stored in the Ali cloud OSS service.
9.6 User authentication management: the system administrator can authenticate the user according to the user detail material and open the related authority.
9.7 Unprocessed operation notification management: the user can add operation and maintenance notifications according to the abnormal diagnosis result of each current distribution area, notify operation and maintenance personnel to process, and simultaneously modify unprocessed operation and maintenance notifications;
9.8 Historical operation and maintenance notification management: the user may query for the operation and maintenance notifications that have been processed.
Claims (10)
1. Power distribution station district operation monitoring system based on convolution neural network and cloud limit synergism, its characterized in that: the system comprises a cloud center and a plurality of edge nodes;
the cloud center stores a state quantity acquisition module, an abnormal working condition sample generation module, a network establishment and training module and a network optimization module;
the state quantity acquisition module acquires the state quantities of the power distribution station areas of the source domain and the target domain and transmits the state quantities to the abnormal working condition sample generation module;
the abnormal working condition sample generation module stores an XGboost importance ranking model;
the abnormal working condition sample generation module sorts the state quantities of the power distribution station area by using an XGboost importance sorting model so as to form a source domain abnormal working condition label sample and a target domain abnormal working condition sample;
the network establishing and training module establishes a freezing layer network and a source domain full-connection classifier, performs supervised training on freezing layer network structure parameters and weight parameters of the source domain full-connection classifier by using a source domain abnormal working condition label sample, and issues the trained parameters to edge nodes;
the network optimization module receives the freezing layer network characteristic distribution and the classifier result distribution of the target domain sent by the edge node, optimizes the freezing layer network by using a gradient descent algorithm, minimizes the difference between the source domain abnormal working condition sample and the target domain abnormal working condition sample, obtains a migration diagnosis model and sends the migration diagnosis model to the edge node;
the edge node is stored with an operation condition data acquisition module, a distribution result calculation module and a diagnosis result generation module;
the operation condition data acquisition module acquires operation condition data in real time and transmits the operation condition data to the diagnosis result generation module;
the distribution result calculation module reconstructs a freezing layer network and an independent target domain full-connection classifier, inputs a target domain abnormal working condition sample into the reconstructed freezing layer network and the reconstructed target domain full-connection classifier, obtains freezing layer network characteristic distribution and classifier result distribution of the target domain, and uploads the freezing layer network characteristic distribution and the classifier result distribution to the cloud center;
the diagnostic result generation module receives and stores the migration diagnostic model;
and the diagnostic result generation module inputs the operation condition data into the migration diagnostic model to obtain the operation monitoring result of the power distribution area and uploads the operation monitoring result to the cloud center.
2. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the power distribution station area state quantity comprises environmental data, electrical data, statistical data and parameter data;
the environmental data comprises meteorological reference characteristics of an area where the power distribution station area is located;
the electrical data comprises measurement characteristics of the power distribution station during operation;
the parameter data comprises rated characteristics of the power distribution station area and comprehensive equipment thereof;
the statistical data is obtained by calculating the electrical data and the parametric data in real time.
3. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the source domain comprises a plurality of power distribution areas similar to the target domain power distribution areas;
similar judgment criteria to the target area distribution station area include: the difference of the operation years of the target domain power distribution station and the source domain power distribution station is not more than h1 year, the equipment types and the models are the same, the difference of the total electric loads of the platform area is not more than h2%, and the difference of the power supply radius is not more than h3%. h1, h2 and h3 are constants.
4. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the edge node is a node containing state quantity of a power distribution area of the target domain; and the edge nodes and the cloud center perform data interaction through a network.
5. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the XGboost importance ranking model is as follows:
wherein N is the total number of trees;
is a base learner; serial number m =1, …, N; t is a node on the decision tree, Δ Gini (t) is a variation value of the kini coefficient on the t node; vim (X) j ) Represents the state quantity X j Degree of influence on abnormal conditions.
6. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the freezing layer network is a convolution neural network;
the convolution operation and activation function operation formulas of the frozen layer network are as follows:
in the formula (I), the compound is shown in the specification,
and
is the jth characteristic state quantity of the ith layer and the ith characteristic state quantity of the (l-1) th layer;
is a convolution kernel weight matrix between the ith characteristic state quantity of the l-1 layer and the jth characteristic state quantity of the l layer;
is the bias term corresponding to the jth characteristic state quantity of the ith layer; f (x) is the activation function.
7. The system for monitoring operation of the power distribution station area based on the convolutional neural network and cloud edge synergy of claim 1, wherein the maximum pooling calculation formula of the freezing layer network is as follows:
in the formula (I), the compound is shown in the specification,
is the characteristic state quantity of the l-th layer in the pooling calculation, and m × m is the local area covered by the pooling kernel; max (x) is the pooling function.
Is the characteristic state quantity of the l-1 st layer in the pooling calculation.
8. The power distribution area operation monitoring system based on convolution neural network and cloud edge synergy of claim 1, characterized in that output X of ith layer of source domain full-connection classifier l As follows:
X l ＝f(∑X l-1 w l,l-1 +b l ) (4)
in the formula, X l-1 Is the output value of layer l-1; w is a l,l-1 Is between the l-1 st layer and the l-th layerA weight matrix; b l Is the bias term corresponding to the l layer; f (x) is the activation function.
9. The system for monitoring operation of a power distribution area based on the cooperation of the convolutional neural network and the cloud edge as claimed in claim 1, wherein an objective function of optimization of the freezing layer network by the cloud center is as follows:
in the formula (I), the compound is shown in the specification,
representing source domain abnormal condition label samples;
representing a target domain abnormal condition label sample; m and n respectively represent the number of the source domain abnormal working condition label samples and the target domain abnormal working condition label samples; λ is a penalty term; m represents the number of source domain samples; k' represents the number of categories; p is a radical of ik Representing the predicted probability that sample i belongs to class k. Y is ik Is a symbolic function, if the true class of sample i is equal to k, then Y ik Taking 1, otherwise, taking 0; k (x) is a kernel function;
wherein the penalty term λ is as follows:
wherein p is the ratio of the current training times to the total training times.
10. The power distribution station area operation monitoring system based on the convolution neural network and cloud edge synergy of claim 1, wherein: the cloud center also comprises a power distribution station operation monitoring visualization module;
the power distribution station operation monitoring visualization module is used for displaying a power distribution station operation monitoring result; and the power distribution station area operation monitoring result comprises normality and abnormality.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930117.4A CN115603446A (en) | 2022-08-03 | 2022-08-03 | Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930117.4A CN115603446A (en) | 2022-08-03 | 2022-08-03 | Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115603446A true CN115603446A (en) | 2023-01-13 |
Family
ID=84843184
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210930117.4A Pending CN115603446A (en) | 2022-08-03 | 2022-08-03 | Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115603446A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117519117A (en) * | 2024-01-08 | 2024-02-06 | 北京星网船电科技有限公司 | Multifunctional console control method and system |
-
2022
- 2022-08-03 CN CN202210930117.4A patent/CN115603446A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117519117A (en) * | 2024-01-08 | 2024-02-06 | 北京星网船电科技有限公司 | Multifunctional console control method and system |
| CN117519117B (en) * | 2024-01-08 | 2024-03-15 | 北京星网船电科技有限公司 | Multifunctional console control method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ke et al. | Short-term electrical load forecasting method based on stacked auto-encoding and GRU neural network | |
| CN110610280B (en) | Short-term prediction method, model, device and system for power load | |
| CN110674999A (en) | Cell load prediction method based on improved clustering and long-short term memory deep learning | |
| Dou et al. | Hybrid model for renewable energy and loads prediction based on data mining and variational mode decomposition | |
| CN113554466B (en) | Short-term electricity consumption prediction model construction method, prediction method and device | |
| CN113659565B (en) | Online prediction method for frequency situation of new energy power system | |
| Unterluggauer et al. | Short‐term load forecasting at electric vehicle charging sites using a multivariate multi‐step long short‐term memory: A case study from Finland | |
| Eseye et al. | Short-term wind power forecasting using a double-stage hierarchical hybrid GA-ANN approach | |
| CN115603446A (en) | Power distribution station area operation monitoring system based on convolution neural network and cloud edge synergistic effect | |
| CN116933318A (en) | Power consumption data privacy protection method based on federal learning | |
| CN115630320A (en) | Power distribution station area abnormal working condition diagnosis method based on deep migration network under cloud edge cooperation | |
| CN112215410B (en) | Power load prediction method based on improved deep learning | |
| Renata et al. | Modeling of electric vehicle charging energy consumption using machine learning | |
| CN115481788B (en) | Phase change energy storage system load prediction method and system | |
| Ávila et al. | Fuzzy demand forecasting in a predictive control strategy for a renewable-energy based microgrid | |
| CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
| Lai et al. | Short‐term passenger flow prediction for rail transit based on improved particle swarm optimization algorithm | |
| CN116720132A (en) | Power service identification system, method, device, medium and product | |
| CN115730526A (en) | Intelligent monitoring and predicting method and system for airport electric load | |
| CN115759343A (en) | E-LSTM-based user electric quantity prediction method and device | |
| Shayanfar et al. | Wind power prediction model based on hybrid strategy | |
| Cheng et al. | Reactive Power Load Forecasting based on K-means Clustering and Random Forest Algorithm | |
| CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
| CN109615142A (en) | A kind of wind farm wind velocity combination forecasting method based on wavelet analysis | |
| CN115660893A (en) | Transformer substation bus load prediction method based on load characteristics |
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 |