CN115169719A - Platform area state prediction model based on fuzzy clustering and BP neural network - Google Patents

Platform area state prediction model based on fuzzy clustering and BP neural network Download PDF

Info

Publication number
CN115169719A
CN115169719A CN202210866470.0A CN202210866470A CN115169719A CN 115169719 A CN115169719 A CN 115169719A CN 202210866470 A CN202210866470 A CN 202210866470A CN 115169719 A CN115169719 A CN 115169719A
Authority
CN
China
Prior art keywords
neural network
node
sample data
data
layer
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
Application number
CN202210866470.0A
Other languages
Chinese (zh)
Inventor
王记强
徐晓波
胡文超
蒋志刚
章亚辉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Anhui Mingsheng Hengzhuo Technology Co ltd
Original Assignee
Anhui Mingsheng Hengzhuo Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Anhui Mingsheng Hengzhuo Technology Co ltd filed Critical Anhui Mingsheng Hengzhuo Technology Co ltd
Priority to CN202210866470.0A priority Critical patent/CN115169719A/en
Publication of CN115169719A publication Critical patent/CN115169719A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Abstract

The invention belongs to the technical field of electric power information, and particularly relates to a platform area state prediction model based on fuzzy clustering and a BP neural network. The platform area state prediction model comprises: the device comprises a data acquisition module, a preprocessing module, a clustering module and a prediction module. The data acquisition module is used for acquiring power information of a user side, a low-voltage side and a transformer area side in the transformer area to obtain a sample data set. The preprocessing module is used for carrying out standardization processing on the sample data. And the clustering module performs clustering processing by adopting a sample data set of a fast search density peak algorithm to generate a node attribute data set. The prediction module adopts a BP neural network with a three-layer structure. The input of the prediction module is the attribute of each node, the output of the prediction module is the predicted running state of the power system of the transformer area, and the running state is divided into 'normal', 'early warning' and 'abnormal'. The invention solves the problems of large data volume of power information in the transformer area, high analysis difficulty and difficult prediction of the running state of the transformer area.

Description

Platform area state prediction model based on fuzzy clustering and BP neural network
Technical Field
The invention belongs to the technical field of electric power information, and particularly relates to a platform area state prediction model based on fuzzy clustering and a BP neural network.
Background
The electric power system is an electric energy production and consumption system which consists of links such as a power plant, a power transmission and transformation line, a power supply and distribution station, power utilization and the like. The function of the device is to convert the primary energy of the nature into electric energy through a power generation device, and then supply the electric energy to each user through power transmission, power transformation and power distribution. In order to realize coordination and control of all links, the power system also comprises an information and control system applied to all links such as power generation, power transmission, power distribution, marketing and the like, so that the production and application processes of electric energy are measured, adjusted, controlled, protected, communicated and scheduled, and a user is ensured to obtain safe and high-quality electric energy.
The development goal of modern power systems is to achieve power system automation. The fields of power system automation include automatic detection, regulation and control of production processes, automatic safety protection of systems and elements, automatic transmission of network information, automatic scheduling of system production, and automatic economic management of enterprises. The main objectives of power system automation are to ensure the quality of power supply (frequency and voltage), to ensure the safety and reliability of system operation, and to improve economic efficiency and management efficiency.
In the links of power distribution and power utilization, equipment such as the intelligent electric energy meter and the fusion terminal can realize remote collection of electric power information, and remote electric power monitoring and management can be realized based on collected electric power information big data. For example, in the existing power system, the integrated terminal and the intelligent electric energy meter can be used for realizing automatic meter reading in a platform area, and the functions of automatic settlement of power consumption fees and the like are realized by combining with an online payment system.
The large data of the electric power information can also be applied to monitoring the running state of an electric power system and the like, however, electric power equipment in a single distribution area is numerous, and the load state of an electric node is complex; how to identify and predict the operation state of the power system through complex power information is still a technical problem to be solved urgently.
Disclosure of Invention
The method aims to solve the problems that the data volume of power information in a transformer area is large, the analysis difficulty is high, and the running state of the transformer area is difficult to predict; the invention provides a platform area state prediction model based on fuzzy clustering and a BP neural network.
The invention is realized by adopting the following technical scheme:
a platform area state prediction model based on fuzzy clustering and a BP neural network is used for predicting the operation state of a power system in a platform area according to collected power information of each node in the platform area; the station area state prediction model comprises: the device comprises a data acquisition module, a preprocessing module, a clustering module and a prediction module.
The data acquisition module is used for acquiring power data and state information of each piece of electrical equipment on a user side, a low-voltage side and a station side in a station area in real time in a node division manner; and then obtaining a sample data set of each node.
The preprocessing module is used for preprocessing all the acquired sample data in the sample data sets of different nodes. The preprocessing module comprises a normalization unit and a data point checking unit. The normalization unit is used to map the non-normalized raw data into the interval of (-1, 1). The data point checking unit is used for calculating the variation rate of partial sample data in the sample data set and corresponding sample data at the previous moment, and when the variation rate is larger than a preset amplitude, the sample data at the previous moment is used for replacing the sample data at the current moment.
And the clustering module carries out clustering processing on the sample data set of each node by adopting a rapid search density peak value algorithm to determine the clustering center and the category number of each node. And then taking the clustering result of each node as the attribute of the node, and generating a node attribute data set containing all node attributes.
The prediction module adopts a BP neural network with a three-layer structure. The BP neural network comprises an input layer, a hidden layer and an output layer. The number of nodes N of the input layer is equal to the number of nodes of the node attribute dataset. The number of nodes of the output layer is 1; the number of nodes K of the hidden layer is greater than the sum of the number of nodes of the input layer and the number of nodes of the output layer and less than 2 times the number of nodes of the input layer. And a Tanh function is adopted between the input layer and the hidden layer as an activation function. And a Sigmoid function is adopted between the hidden layer and the output layer as an activation function. The input of the prediction module is the attribute of each node, the output of the prediction module is the predicted running state of the power system of the transformer area, and the running state is divided into 'normal', 'early warning' and 'abnormal'.
As a further improvement of the present invention, the electrical devices on the user side include a distributed power generation unit, a centralized charging device, a centralized energy storage device, and an intelligent electric energy meter installed at all the power users. The low-voltage side power equipment comprises a distribution box, a distribution cabinet and an independent circuit breaker. The power equipment on the transformer area side comprises a transformer, a capacitor and a transformer area environment monitoring device.
As a further improvement of the present invention, the normalization unit normalizes the data according to a preset theoretical safety threshold of each non-standardized sample data, and the normalization processing formula is as follows:
Figure BDA0003758786800000021
in the above formula, x represents the measured value of the current sample data;
Figure BDA0003758786800000022
a normalized value representing the current sample data; x is the number of max Representing the theoretical safety threshold upper limit of the current sample data; x is the number of min Representing a theoretical safety threshold lower limit of the current sample data; when some sample data does not have the lower limit of the safety threshold, x min ＝0。
As a further improvement of the present invention, the data point checking unit is used for performing error checking on non-standardized sample data. Non-standardized data refers to data collected in addition to status information.
As a further improvement of the invention, the clustering process of the sample data set by the clustering module is as follows:
firstly, acquiring a sample data set after normalization of any node, and calculating the Euclidean distance d between any two sample data in the sample data set ij The calculation formula is as follows:
Figure BDA0003758786800000031
in the above formula, a represents a sample data set, and N represents the number of data points in the sample data set a; x is the number of i And x j Representing two data points in the sample data set a that are random. dist () represents a euclidean distance computation function.
Secondly, according to Euclidean distance between the sample data and all other data points and preset truncation distance d 0 The local density rho of each data point in the sample data is calculated i (ii) a The calculation formula is as follows:
Figure BDA0003758786800000032
wherein the content of the first and second substances,
Figure BDA0003758786800000033
representing a self-defined classification function for distinguishing whether the Euclidean distance between the data point and the central point is less than the truncation distance, and meeting the following conditions:
Figure BDA0003758786800000034
then, based on the local density of each data point of the sample data set, the distance theta between the sample data and the density center is calculated i . The calculation process is as follows: judging whether the local density of the current data point is the maximum in the sample data setThe value: if yes, the current data point x is i The maximum distance from other data points in the sample dataset is taken as θ i The calculation formula is as follows: theta i ＝max j (d ij ) J = N. Otherwise, the local density in the sample data set is larger than the data point x of the current data point j With the current data point x i Is taken as the minimum distance of i (ii) a The calculation formula is as follows: theta i ＝min(d ij )，x j :ρ j ＞ρ i 。
Finally, according to rho of each sample data in the sample data set i And theta i And drawing a decision diagram. The abscissa of each sample point in the decision graph is p i Ordinate of theta i (ii) a And determining the clustering center and the category number according to the decision graph.
As a further improvement of the present invention, in the BP neural network, the expression of the Tanh activation function is as follows:
Figure BDA0003758786800000035
the expression of Sigmoid activation function is as follows:
Figure BDA0003758786800000041
as a further improvement of the invention, in the BP neural network, the transfer formula from the input layer to the hidden layer is as follows:
Figure BDA0003758786800000042
in the above formula, x i For an input value of an ith input layer, i = N, N representing the number of nodes of the input layer; h 1j For the output of the jth node of the hidden layer, j =2N +1; f. of 1 As a function of Tanh activation, ω ij Is the weight value, x, between the ith node of the input layer and the jth node of the hidden layer i Is an input value of the ith node of the input layer, a j For the hidden layer jThreshold of individual nodes.
The transfer formula from the hidden layer to the output layer is as follows;
Figure BDA0003758786800000043
in the above formula, y is the output value of the output layer, f 2 For Sigmoid activation functions, ω j The weight between the jth node of the hidden layer and the output layer, and b is the threshold of the output layer.
As a further improvement of the invention, the BP neural network constructed in the prediction module needs to be trained to update the weight and the threshold of the network model, and the training process of the BP neural network is as follows:
(1) Data set preparation: extracting power data and state information of each node in different operation states from the historical data of the platform area, and then preprocessing and normalizing the acquired data to obtain a plurality of node attribute data sets containing all node attributes and state labels corresponding to the node attribute data sets. And taking the node attribute data containing the state notes as a training set required by a model training stage.
(2) Setting initial training parameters, including: iteration number n, loss function E, target minimum loss rate E 0 And a learning rate η.
(3) Training the BP neural network by adopting a training set, updating the weight of the model in the training process, introducing the classical genetic algorithm into the training stage, and optimizing the threshold of the BP neural network through the classical genetic algorithm in each iteration process. The iterative process is as follows:
and carrying out forward propagation on the training set by using the BP neural network.
Calculating whether the value of the loss function E in each round of forward propagation is less than the target minimum loss rate E 0 (ii) a If so, finishing the training process of the network model, otherwise, entering a back propagation process.
And iii, performing a reverse propagation process by adopting a gradient descent method, and dynamically updating the weight of the output layer and the weight between the input layer and the hidden layer according to a weight updating formula from the hidden layer to the output layer and a weight updating formula between the input layer and the hidden layer.
Wherein, the weight ω between the input layer and the hidden layer ij The update formula of (2) is as follows:
Figure BDA0003758786800000051
in the above formula, ω' ij And representing the updated weight between the ith node of the input layer and the jth node of the hidden layer.
Weight omega from hidden layer to output layer j The update formula of (2) is as follows:
Figure BDA0003758786800000052
in the above formula, ω j ' denotes a weight between the jth node of the updated hidden layer and the output layer.
Meanwhile, the threshold value of the network model is adjusted to the threshold value optimized by the genetic algorithm in the iteration process.
And iv, continuously inputting the training samples into the BP neural network after the weight and the threshold are updated, and repeating the forward propagation of the next round.
V. looping steps i-iv until a predetermined number of iterations n is reached, or until the value of the loss function E converges and is less than a target minimum loss rate E 0 . And saving the weight and the threshold of the trained BP neural network.
As a further improvement of the invention, the loss function E adopts a mean square error loss function, and the expression is as follows:
Figure BDA0003758786800000053
in the above formula, T represents a true label of the training set; y represents the output of the BP neural network; m represents the sample data size of the training set; t is t i Status Note, y, representing the ith training sample i Representing the actual output of the i-th training sample as modeled.
As a further improvement of the invention, the process of optimizing the threshold value of the BP neural network by using the genetic algorithm is as follows:
s1: and converting the threshold value of the BP neural network into a corresponding chromosome individual by adopting a real number coding mode, and further randomly generating an initial population containing a plurality of chromosome individuals.
S2: and setting the iteration termination condition of the classical genetic algorithm as the training phase synchronization of the BP neural network, and performing iterative optimization on the initial population by using the classical genetic algorithm, wherein each iteration comprises the following contents.
S21: and calculating the fitness of each chromosome in the initial population by using a preset fitness function.
S22: selecting the initial population by adopting a selection operator of a classical genetic algorithm; and the selection operator adopts an elite reservation operator.
S23: performing cross operation on the initial population by adopting a cross operator of a classical genetic algorithm; the crossover operator calculates the similarity between any two individuals in the current population through a self-defined similarity function, and then performs double-point crossover on the two individuals with the lowest similarity.
S24: carrying out mutation operation on the initial population by adopting a mutation operator of a classical genetic algorithm; and the mutation operator performs single-point mutation on the chromosomes with the highest fitness and the lowest fitness according to a preset mutation probability.
S03: and in each round of population iteration process, outputting the threshold value of the chromosome characterization with the maximum fitness to the BP neural network as the threshold value of the BP neural network in the next training round.
The technical scheme provided by the invention has the following beneficial effects:
the invention designs a platform area state prediction model based on fuzzy clustering and a BP (back propagation) neural network, which is characterized in that firstly, acquired electric power data of different nodes of a platform area are normalized into standard data and then into a sample data set, then the sample data set of each node is clustered by a self-adaptive fast density peak value searching method, finally, a clustered multi-node attribute data set is input into a model based on the BP neural network, the model predicts the power supply stability state of the platform area according to a prediction sample, and a prediction result is output. The constructed BP neural network is trained by combining the collected sample data set, and the classical genetic algorithm is applied to the updating of the network model threshold in the training process, so that the iterative cycle of the training stage of the network model is effectively shortened, and the identification precision of the model is improved.
Based on the prediction model provided by the invention, massive transformer area power information can be analyzed and processed, and the sub-prediction result can be quickly obtained. The prediction model provided by the invention not only has higher prediction precision, but also improves the real-time property of model processing. And the network model provided by the invention can be automatically learned according to the recognition result in the application process, so that the prediction precision of the method can be continuously improved.
Drawings
Fig. 1 is a schematic block diagram of a platform region state prediction model based on fuzzy clustering and a BP neural network according to embodiment 1 of the present invention.
Fig. 2 is a flowchart of a step of clustering a sample data set by a clustering module in embodiment 1 of the present invention.
Fig. 3 is a network architecture diagram of the BP neural network model constructed in embodiment 1 of the present invention.
Fig. 4 is a flowchart of the procedure of the training phase of the BP neural network constructed in embodiment 1 of the present invention.
Fig. 5 is a flowchart of a procedure for optimizing the BP neural network threshold value by using a classical genetic algorithm in embodiment 1 of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
The embodiment provides a platform area state prediction model based on fuzzy clustering and a BP neural network, wherein the prediction model is used for predicting the operation state of a power system in a platform area according to collected power information of each node in the platform area; as shown in fig. 1, the platform area state prediction model includes: the device comprises a data acquisition module, a preprocessing module, a clustering module and a prediction module.
The data acquisition module is used for acquiring power data and state information of each piece of electrical equipment on a user side, a low-voltage side and a station side in a station area in real time in a node division manner; and further obtaining a sample data set of each node. In this embodiment, the electrical devices on the user side include a distributed power generation unit, a centralized charging device, a centralized energy storage device, and smart meters installed at all power users. The low-voltage side power equipment comprises a distribution box, a distribution box and an independent circuit breaker. The power equipment on the transformer area side comprises a transformer, a capacitor and a transformer area environment monitoring device.
The collected data in the sample data set may include: power monitoring data of the grid, such as: supply voltage V of power consumer corresponding node 1 Supply current I 1 Power factor of the power converter
Figure BDA0003758786800000071
The line loss rate Δ P. Device voltage V of transformer in transformer area 2 Device current I 2 Real-time load P, device temperature T, etc. Wherein the collected supply voltage V 1 Supply current I 1 Voltage V of the device 2 Device current I 2 The voltage and current of each phase A, B and C are respectively corresponded.
The power factor refers to the ratio of active power to apparent power of the alternating current circuit; is a coefficient for measuring the efficiency of the electrical equipment. The power factor is low, which indicates that the reactive power of the circuit for alternating magnetic field conversion is large, thereby reducing the utilization rate of equipment and increasing the power supply loss of a line. The power factor is related to the load properties of the grid circuit; is also an important technical index for evaluating the power system.
The line loss rate is the percentage of electrical energy lost in the grid (line loss load) to supply to the power network (supply load). The line loss rate may assess the economics of the operation of the power system. The line loss rate is related to the load power factor of the power grid, the voltage supplied to the load by the power system changes along with the change of the active power and the reactive power transmitted by the line, and when the active power transmitted by the line and the voltage at the starting end are unchanged, the more the reactive power transmitted, the greater the voltage loss of the line, and the higher the line loss rate. Therefore, when the power factor is improved, the reactive power which is absorbed by the load to the system is reduced, the voltage loss of the line is correspondingly reduced, and the line loss rate is reduced.
The equipment temperature is an index for evaluating the tolerance of the transformer equipment, heat dissipation is required in the running process of the transformer, and when the real-time load of the transformer is increased, the equipment temperature is correspondingly increased. If the environment temperature and other conditions cause the transformer to be unable to effectively dissipate heat, the equipment temperature of the transformer may rise rapidly, which may cause the transformer to be unable to operate according to the rated load, thereby affecting the normal power supply in the transformer area. Therefore, the present embodiment also needs to take the equipment temperature of the transformer into account when analyzing the grid operating state.
In addition, the collected sample data also includes state information or related data of various types of abnormalities or faults of the power grid, for example: voltage deviation, frequency deviation, three-phase voltage or three-phase current unbalance, voltage fluctuation, flicker, harmonic data and inter-harmonic data of the power grid. The harmonic data includes harmonic voltage or harmonic current content, harmonic current effective value, voltage total harmonic distortion, current total harmonic distortion, harmonic phase angle and harmonic power. The inter-harmonic data includes inter-harmonic voltage or inter-harmonic current content, and inter-harmonic current effective value. The transient data index items include event data and valid value data. Event data includes voltage sags, and short interrupts. The valid value data refers to the valid value data of the trigger record.
The preprocessing module is used for preprocessing all the acquired sample data in the sample data sets of different nodes. The preprocessing module comprises a normalization unit and a data point checking unit. The normalization unit is used to map the non-normalized raw data into the interval of (-1, 1). The data point checking unit is used for calculating the variation rate of partial sample data in the sample data set and the corresponding sample data at the previous moment, and when the variation rate is larger than a preset amplitude, the sample data at the previous moment is used for replacing the sample data at the current moment.
In this embodiment, it is considered that the dimensions and the dimension units of each item of collected data are different, and it is relatively unfavorable for the data analysis in the later period. Therefore, dimension influence among different index items is eliminated through normalization processing. In the embodiment, after normalization processing, data of different dimensions are mapped into the same interval, so that the data indexes have comparability. After the raw data are subjected to data standardization processing, all indexes are in the same order of magnitude, and the method is suitable for comprehensive comparison and evaluation.
Specifically, in this embodiment, the normalization unit normalizes the data according to a preset theoretical safety threshold of each non-standardized sample data, and the normalization processing formula is as follows:
Figure BDA0003758786800000081
in the above formula, x represents the measured value of the current sample data;
Figure BDA0003758786800000082
a normalized value representing the current sample data; x is a radical of a fluorine atom max Representing the theoretical safety threshold upper limit of the current sample data; x is the number of min Representing a theoretical safety threshold lower limit of the current sample data; wherein, when some sample data has no lower limit of safety threshold, then x min ＝0。
The data point checking unit in the present embodiment is configured to perform error checking on non-standardized sample data. Non-standardized data refers to data collected in addition to status information. The purpose of error checking is to cull collected data that is significantly erroneous. The data belong to interference data with numerical value mutation, when the data enter a feature learning stage of a network model in a later stage, the training effect of the network model may be influenced, and the overall performance of the network model is improved by eliminating the interference data in a data preprocessing stage.
And the clustering module performs clustering processing on the sample data set of each node by adopting a rapid search density peak value algorithm to determine the clustering center and the category number of each node. And then taking the clustering result of each node as the attribute of the node, and generating a node attribute data set containing all node attributes.
As shown in fig. 2, the clustering process of the clustering module on the sample data set in this embodiment is as follows:
firstly, a sample data set after normalization of any node is obtained, and the Euclidean distance d between any two sample data in the sample data set is calculated ij The calculation formula is as follows:
Figure BDA0003758786800000091
in the above formula, a represents a sample data set, and N represents the number of data points in the sample data set a; x is the number of i And x j Representing two data points in the sample data set a at random. dist () represents a euclidean distance computation function.
Secondly, according to Euclidean distance between the sample data and all other data points and preset truncation distance d 0 The local density rho of each data point in the sample data is calculated i (ii) a The calculation formula is as follows:
Figure BDA0003758786800000092
wherein the content of the first and second substances,
Figure BDA0003758786800000093
representing a self-defined classification function for distinguishing whether the Euclidean distance between the data point and the central point is less than the truncation distance, and meeting the following conditions:
Figure BDA0003758786800000094
then, based on the local density of each data point of the sample data set, the distance theta between the sample data and the density center is calculated i . The calculation process is as follows: judging whether the local density of the current data point is the maximum value in the sample data set: if yes, the current data point x is i The maximum distance from other data points in the sample dataset is taken as θ i The calculation formula is as follows: theta i ＝max j (d ij ) J = N. Otherwise, the local density in the sample data set is larger than the data point x of the current data point j And the current data point x i Is taken as the minimum distance of i (ii) a The calculation formula is as follows: theta i ＝min(d ij )，x j :ρ j ＞ρ i 。
Finally, according to rho of each sample data in the sample data set i And theta i And drawing a decision graph. The abscissa of each sample point in the decision graph is p i Ordinate of theta i (ii) a And determining the clustering center and the category number according to the decision graph.
The adaptive fast density peak searching method provided by the embodiment determines the clustering center by analyzing the densities of different data points, and the basis for realizing clustering is that the density of the clustering center is greater than the density of adjacent points around the clustering center. The method can automatically acquire the category number according to the collected data set without determining an initial clustering center, and quickly find the density peak of the data set in any shape by utilizing the characteristic that the local density of the category center is always higher than that of the nearest neighbor point, effectively distribute non-central sample points, and determine the clustering center by utilizing a decision diagram. The self-adaptive fast density peak searching method provided by the embodiment is suitable for multi-class clustering, and the clustering processing efficiency is obviously improved.
After the sample data is clustered, the category (attribute) corresponding to each node of the power system can be obtained, and in the subsequent stage, the BP neural network can accurately predict the operating state of the power system by combining the attributes of all the nodes.
The prediction module in this embodiment employs a BP neural network having a three-layer structure. The BP neural network comprises an input layer, a hidden layer and an output layer. As shown in FIG. 3, the number of nodes N of the input layer is equal to the number of nodes of the node attribute data set. The number of nodes of the output layer is 1. The number of nodes K of the hidden layer is greater than the sum of the number of nodes of the input layer and the number of nodes of the output layer and less than 2 times the number of nodes of the input layer. And a Tanh function is adopted between the input layer and the hidden layer as an activation function. And a Sigmoid function is adopted between the hidden layer and the output layer as an activation function. The input of the prediction module is the attribute of each node, the output of the prediction module is the predicted running state of the power system of the transformer area, and the running state is divided into 'normal', 'early warning' and 'abnormal'.
In the BP neural network, the expression of the Tanh activation function is as follows:
Figure BDA0003758786800000101
the expression of the Sigmoid activation function is as follows:
Figure BDA0003758786800000102
the input layer to hidden layer transfer formula is:
Figure BDA0003758786800000103
in the above formula, x i For the input value of the ith input layer, i = N, where N represents the number of nodes of the input layer; h 1j For the output of the jth node of the hidden layer, j =2N +1; f. of 1 As a function of Tanh activation, ω ij Is the weight, x, between the ith node of the input layer and the jth node of the hidden layer i Is an input value of the ith node of the input layer, a j Is the threshold of the jth node of the hidden layer.
The transfer formula from the hidden layer to the output layer is as follows;
Figure BDA0003758786800000104
in the above formula, y is the output value of the output layer, f 2 For Sigmoid activation functions, ω j The weight between the jth node of the hidden layer and the output layer, and b is the threshold of the output layer.
In this embodiment, the BP neural network constructed in the prediction module needs to be trained to update the weight and the threshold of the network model, as shown in fig. 4, the training process of the BP neural network is as follows:
(1) Data set preparation: extracting power data and state information of each node in different operation states from the historical data of the platform area, and then preprocessing and normalizing the acquired data to obtain a plurality of node attribute data sets containing all node attributes and state labels corresponding to the node attribute data sets. And taking the node attribute data containing the state notes as a training set required by the model training phase.
(2) Setting initial training parameters, including: iteration number n, loss function E, target minimum loss rate E 0 And a learning rate η. The loss function E adopts a mean square error loss function, and the expression is as follows:
Figure BDA0003758786800000111
in the above formula, T represents a real label of the training set; y represents the output of the BP neural network; m represents the sample data size of the training set; t is t i Status Note, y, representing the ith training sample i Representing the actual output of the ith training sample through the model.
(3) And training the BP neural network by adopting a training set, updating the weight of the model in the training process, introducing a classical genetic algorithm into the training stage, and optimizing the threshold of the BP neural network by the classical genetic algorithm in each iteration process. The iterative process is as follows:
and carrying out forward propagation on the training set by using the BP neural network.
Ii, calculating the positive direction of each roundWhether the value of the loss function E in the propagation process is less than the target minimum loss rate E 0 (ii) a If so, finishing the training process of the network model, otherwise, entering a back propagation process.
And iii, performing a back propagation process by adopting a gradient descent method, and dynamically updating the weight of the output layer and the weight between the input layer and the hidden layer according to a weight updating formula from the hidden layer to the output layer and a weight updating formula between the input layer and the hidden layer.
Wherein, the weight ω between the input layer and the hidden layer ij The update formula of (2) is as follows:
Figure BDA0003758786800000112
in the above formula, ω' ij And representing the updated weight between the ith node of the input layer and the jth node of the hidden layer.
Weight omega from hidden layer to output layer j The update formula of (c) is as follows:
Figure BDA0003758786800000113
in the above formula, ω' j And representing the weight between the jth node of the updated hidden layer and the output layer.
Meanwhile, the threshold value of the network model is adjusted to the threshold value optimized by the genetic algorithm in the iteration process.
And iv, continuously inputting the training samples into the BP neural network after the weight and the threshold are updated, and repeating the forward propagation of the next round.
V. looping steps i-iv until a predetermined number of iterations n is reached, or until the value of the loss function E converges and is less than the target minimum loss rate E 0 . And storing the weight and the threshold of the trained BP neural network.
In this embodiment, as shown in fig. 5, the process of optimizing the threshold value of the BP neural network by using the genetic algorithm is as follows:
s1: and converting the threshold value of the BP neural network into a corresponding chromosome individual by adopting a real number coding mode, and further randomly generating an initial population containing a plurality of chromosome individuals.
S2: and setting the iteration termination condition of the classical genetic algorithm as the training phase synchronization of the BP neural network, and performing iterative optimization on the initial population by using the classical genetic algorithm, wherein each iteration comprises the following contents.
S21: and calculating the fitness of each chromosome in the initial population by using a preset fitness function.
S22: selecting the initial population by adopting a selection operator of a classical genetic algorithm; the selection operator adopts an elite reservation operator.
S23: performing cross operation on the initial population by adopting a cross operator of a classical genetic algorithm; the crossover operator calculates the similarity between any two individuals in the current population through a self-defined similarity function, and then performs double-point crossover on the two individuals with the lowest similarity.
S24: carrying out mutation operation on the initial population by adopting a mutation operator of a classical genetic algorithm; and carrying out single-point mutation on the chromosomes with the highest fitness and the lowest fitness according to a preset mutation probability by the mutation operator.
S03: and in each round of population iteration process, outputting the threshold value of the chromosome characterization with the maximum fitness to the BP neural network as the threshold value of the BP neural network in the next training round.
In the embodiment, a BP neural network is adopted to complete the task of predicting the running state of the power system of the transformer area according to the node attribute data sets of different nodes. And the BP neural network adopts a successive approximation mode, and carries out error analysis on the obtained result and the expected result in the continuous algorithm training process, so as to modify the weight and the threshold value and obtain a model which can be output and is consistent with the expected result in one step. In particular, in order to improve the training effect of the network model, the embodiment also introduces the genetic algorithm into the threshold updating process of the network model, so as to significantly improve the threshold updating effect of the BP neural network model. The trained network model has higher response speed, so that the power supply state of the transformer area can be dynamically evaluated according to the data acquired in real time.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims (10)

1. A platform area state prediction model based on fuzzy clustering and a BP neural network is used for predicting the operation state of an electric power system in a platform area according to collected electric power information of each node in the platform area; the station area state prediction model is characterized by comprising the following steps:
the data acquisition module is used for acquiring power data and state information of each piece of electrical equipment on a user side, a low-voltage side and a station side in the station area in real time in a node division manner; further obtaining a sample data set of each node;
the preprocessing module is used for preprocessing all the acquired sample data in the sample data sets of different nodes; the preprocessing module comprises a normalization unit and a data point checking unit; the normalization unit is used for mapping the non-normalized raw data into an interval of (-1, 1); the data point checking unit is used for calculating the variation rate of partial sample data in the sample data set and corresponding sample data at the previous moment, and when the variation rate is larger than a preset amplitude, the sample data at the previous moment is used for replacing the sample data at the current moment;
the clustering module is used for clustering each node sample data set by adopting a rapid search density peak algorithm and determining the clustering center and the category number of each node; then, taking the clustering result of each node as the attribute of the node, and generating a node attribute data set containing all node attributes; and
a prediction module employing a BP neural network having a three-layer structure; the BP neural network comprises an input layer, a hidden layer and an output layer; the number N of nodes of the input layer is equal to the number of nodes of the node attribute data set; the number of nodes of the output layer is 1; the node number K of the hidden layer is more than the sum of the node number of the input layer and the node number of the output layer and less than 2 times of the node number of the input layer; the input layer and the hidden layer adopt Tanh functions as activation functions, and the hidden layer and the output layer adopt Sigmoid functions as activation functions; the input of the prediction module is the attribute of each node, the output of the prediction module is the predicted running state of the power system of the transformer area, and the running state is divided into 'normal', 'early warning' and 'abnormal'.
2. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 1, wherein: the electrical equipment at the user side comprises a distributed power generation unit, centralized charging equipment, centralized energy storage equipment and intelligent electric energy meters installed at all power users; the power equipment on the low-voltage side comprises a distribution box, a distribution box and an independent circuit breaker; the power equipment on the transformer area side comprises a transformer, a capacitor and a transformer area environment monitoring device.
3. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 1, wherein: the normalization unit is used for normalizing the data according to a preset theoretical safety threshold of each non-standardized sample data, and the normalization processing formula is as follows:
Figure FDA0003758786790000021
in the above formula, x represents the measured value of the current sample data;
Figure FDA0003758786790000022
a normalized value representing the current sample data; x is the number of max Representing the theoretical safety threshold upper limit of the current sample data; x is the number of min Representing a theoretical safety threshold lower limit of the current sample data; when some sample data does not have the lower limit of the safety threshold, x min ＝0。
4. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 1, wherein: the data point checking unit is used for carrying out error checking on non-standardized sample data; the non-standardized data refers to other data collected besides the state information.
5. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 1, wherein: the clustering process of the clustering module on the sample data set is as follows:
firstly, acquiring a sample data set after normalization of any node, and calculating the Euclidean distance d between any two sample data in the sample data set ij The calculation formula is as follows:
Figure FDA0003758786790000023
in the above formula, a represents the sample data set, and N represents the number of data points in the sample data set a; x is the number of i And x j Representing two random data points in the sample data set A; dist () represents the euclidean distance computation function;
secondly, according to Euclidean distance between the sample data and all other data points and preset truncation distance d 0 The local density rho of each data point in the sample data is calculated i (ii) a The calculation formula is as follows:
Figure FDA0003758786790000024
wherein the content of the first and second substances,
Figure FDA0003758786790000025
representing a self-defined classification function for distinguishing whether the Euclidean distance between the data point and the central point is less than the truncation distance, and satisfying the following conditions:
Figure FDA0003758786790000026
then, based on the local density of each data point of the sample data set, the distance theta between the sample data and the density center is calculated i The calculation process is as follows: judging whether the local density of the current data point is the maximum value in the sample data set: if yes, the current data point x is i The maximum distance from other data points in the sample dataset is taken as θ i The calculation formula is as follows: theta.theta. i ＝max j (d ij ) J = N; otherwise, the local density in the sample data set is larger than the data point x of the current data point j With the current data point x i Is taken as the minimum distance of theta i (ii) a The calculation formula is as follows:
θ i ＝min(d ij )，x j :ρ j ＞ρ i ；
finally, according to rho of each sample data in the sample data set i And theta i Drawing a decision graph; the abscissa of each sample point in the decision graph is p i Ordinate of theta i (ii) a And determining the clustering center and the category number according to the decision graph.
6. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 1, wherein: in the BP neural network, the expression of the Tanh activation function is as follows:
Figure FDA0003758786790000031
the expression of the Sigmoid activation function is as follows:
Figure FDA0003758786790000032
7. the platform region state prediction model based on fuzzy clustering and BP neural network of claim 6, wherein: in the BP neural network, the transfer formula from the input layer to the hidden layer is:
Figure FDA0003758786790000033
in the above formula, x i For the input value of the ith input layer, i = N, where N represents the number of nodes of the input layer; h 1j For the output of the jth node of the hidden layer, j =2N +1; f. of 1 As a function of Tanh activation, ω ij Is the weight, x, between the ith node of the input layer and the jth node of the hidden layer i Is an input value of an ith node of the input layer, a j A threshold value of the jth node of the hidden layer;
the transfer formula from the hidden layer to the output layer is as follows;
Figure FDA0003758786790000034
in the above formula, y is the output value of the output layer, f 2 For Sigmoid activation functions, ω j The weight between the jth node of the hidden layer and the output layer, and b is the threshold of the output layer.
8. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 7, wherein: the BP neural network constructed in the prediction module needs to be trained to update the weight and the threshold of the network model, and the training process of the BP neural network is as follows:
(1) Data set preparation: extracting power data and state information of each node in different operation states from the historical data of the platform area, and then preprocessing and normalizing the acquired data to obtain a plurality of node attribute data sets containing all node attributes and state labels corresponding to the node attribute data sets; taking the node attribute data containing the state notes as a training set required by a model training stage;
(2) Setting initial training parameters, including: number of iterations n, loss function E, objectiveMinimum loss rate E 0 And a learning rate η;
(3) Training the BP neural network by adopting a training set, updating the weight of the model in the training process, introducing a classical genetic algorithm into the training stage, and optimizing the threshold of the BP neural network by the classical genetic algorithm in each iteration process; the iterative process is as follows:
carrying out forward propagation on the training set by using a BP neural network;
calculating whether the value of the loss function E in each round of forward propagation is less than the target minimum loss rate E 0 (ii) a If yes, finishing the training process of the network model, otherwise, entering a back propagation process;
adopting a gradient descent method to perform a reverse propagation process, and dynamically updating the output layer weight and the weight between the input layer and the hidden layer according to a weight updating formula from the hidden layer to the output layer and a weight updating formula between the input layer and the hidden layer;
wherein, the weight ω between the input layer and the hidden layer ij The update formula of (2) is as follows:
Figure FDA0003758786790000041
in the above formula, ω' ij Representing the updated weight between the ith node of the input layer and the jth node of the hidden layer;
weight omega from hidden layer to output layer j The update formula of (2) is as follows:
Figure FDA0003758786790000042
in the above formula, ω' j Representing the weight between the jth node of the updated hidden layer and the output layer;
adjusting the threshold value of the network model to the threshold value optimized by the genetic algorithm;
continuing inputting the training sample into the BP neural network after updating the weight and the threshold, and repeating the forward propagation of the next round;
v. looping steps i-iv until a predetermined number of iterations n is reached, or until the value of the loss function E converges and is less than the target minimum loss rate E 0 (ii) a And saving the weight and the threshold of the trained BP neural network.
9. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 8, wherein: the loss function E adopts a mean square error loss function, and the expression is as follows:
Figure FDA0003758786790000051
in the above formula, T represents a real label of the training set; y represents the output of the BP neural network; m represents the sample data size of the training set; t is t i Status Note, y, representing the ith training sample i Representing the actual output of the ith training sample through the model.
10. The platform region state prediction model based on fuzzy clustering and BP neural network of claim 8, wherein: the process of optimizing the threshold value of the BP neural network by using the genetic algorithm is as follows:
s1: converting the threshold value of the BP neural network into corresponding chromosome individuals by adopting a real number coding mode, and further randomly generating an initial population containing a plurality of chromosome individuals;
s2: setting an iteration termination condition of a classical genetic algorithm as training phase synchronization of a BP neural network, and performing iterative optimization on an initial population by using the classical genetic algorithm, wherein each iteration comprises the following contents;
s21: calculating the fitness of each chromosome in the initial population by using a preset fitness function;
s22: selecting the initial population by adopting a selection operator of a classical genetic algorithm; the selection operator adopts an elite reservation operator;
s23: performing cross operation on the initial population by adopting a cross operator of a classical genetic algorithm; the crossover operator calculates the similarity between any two individuals in the current population through a self-defined similarity function, and then performs double-point crossover on the two individuals with the lowest similarity;
s24: carrying out mutation operation on the initial population by adopting a mutation operator of a classical genetic algorithm; the mutation operator performs single-point mutation on the chromosomes with highest fitness and lowest fitness according to a preset mutation probability;
s03: and in each round of population iteration process, outputting the threshold value of the chromosome characterization with the maximum fitness to the BP neural network as the threshold value of the BP neural network in the next training round.
CN202210866470.0A 2022-07-22 2022-07-22 Platform area state prediction model based on fuzzy clustering and BP neural network Pending CN115169719A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210866470.0A CN115169719A (en) 2022-07-22 2022-07-22 Platform area state prediction model based on fuzzy clustering and BP neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210866470.0A CN115169719A (en) 2022-07-22 2022-07-22 Platform area state prediction model based on fuzzy clustering and BP neural network

Publications (1)

Publication Number Publication Date
CN115169719A true CN115169719A (en) 2022-10-11

Family

ID=83496637

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210866470.0A Pending CN115169719A (en) 2022-07-22 2022-07-22 Platform area state prediction model based on fuzzy clustering and BP neural network

Country Status (1)

Country Link
CN (1) CN115169719A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117154716A (en) * 2023-09-08 2023-12-01 国网河南省电力公司 Planning method and system for accessing distributed power supply into power distribution network
CN117526443A (en) * 2023-11-07 2024-02-06 北京清电科技有限公司 Novel power system-based power distribution network optimization regulation and control method and system
CN117526443B (en) * 2023-11-07 2024-04-26 北京清电科技有限公司 Power system-based power distribution network optimization regulation and control method and system

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117154716A (en) * 2023-09-08 2023-12-01 国网河南省电力公司 Planning method and system for accessing distributed power supply into power distribution network
CN117154716B (en) * 2023-09-08 2024-04-26 国网河南省电力公司 Planning method and system for accessing distributed power supply into power distribution network
CN117526443A (en) * 2023-11-07 2024-02-06 北京清电科技有限公司 Novel power system-based power distribution network optimization regulation and control method and system
CN117526443B (en) * 2023-11-07 2024-04-26 北京清电科技有限公司 Power system-based power distribution network optimization regulation and control method and system

Similar Documents

Publication Publication Date Title
CN105117602B (en) A kind of metering device running status method for early warning
CN107516170A (en) A kind of difference self-healing control method based on probability of equipment failure and power networks risk
CN113156917A (en) Power grid equipment fault diagnosis method and system based on artificial intelligence
CN112069727B (en) Intelligent transient stability evaluation system and method with high reliability for power system
CN110826228B (en) Regional power grid operation quality limit evaluation method
CN116914751B (en) Intelligent power distribution control system
CN116167527A (en) Pure data-driven power system static safety operation risk online assessment method
Li et al. Stacking ensemble learning-based load identification considering feature fusion by cyber-physical approach
CN115118015A (en) Platform district power supply stability monitoring system based on fuse terminal
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN115204698A (en) Real-time analysis method for power supply stability of low-voltage transformer area
CN116245033B (en) Artificial intelligent driven power system analysis method and intelligent software platform
CN117394529A (en) SCADA-based auxiliary decision method and system for main distribution network loop-closing reverse power supply control conditions
CN111061152A (en) Attack recognition method based on deep neural network and intelligent energy power control device
CN110874672A (en) Quantum genetic algorithm-based FARIMA model power load prediction method
CN115169719A (en) Platform area state prediction model based on fuzzy clustering and BP neural network
Gao et al. Establishment of economic forecasting model of high-tech industry based on genetic optimization neural network
Liu et al. Analysis and prediction of power distribution network loss based on machine learning
Li et al. Residential Photovoltaic Power Forecasting Considering Division of Weather Type Index Interval
Cao et al. A hybrid electricity load prediction system based on weighted fuzzy time series and multi-objective differential evolution
Sicheng et al. Abnormal line loss data detection and correction method
CN110837932A (en) Thermal power prediction method of solar heat collection system based on DBN-GA model
Yin et al. Prediction of Distribution Network Line Loss Based on Grey Relation Analysis and XGboost
Dutta et al. Machine Learning‐Based Intelligent Power Systems
Yao et al. A Novel Aggregated Short-Term Load Forecasting Method Based on Clustering

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