CN112883633B - Power distribution network line loss calculation method based on combined weighting method and deep learning - Google Patents
Power distribution network line loss calculation method based on combined weighting method and deep learning Download PDFInfo
- Publication number
- CN112883633B CN112883633B CN202110045657.XA CN202110045657A CN112883633B CN 112883633 B CN112883633 B CN 112883633B CN 202110045657 A CN202110045657 A CN 202110045657A CN 112883633 B CN112883633 B CN 112883633B
- Authority
- CN
- China
- Prior art keywords
- line loss
- theoretical line
- formula
- power distribution
- deep learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
-
- 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
A power distribution network line loss calculation method based on a combined weighting method and deep learning belongs to the technical field of power distribution network theoretical line loss calculation. The invention adopts a deep learning GRU network model to fit the nonlinear relation between the electrical parameters and the theoretical line loss, improves the nonlinear function approximation capability of the traditional BP algorithm, comprehensively considers subjective and objective factors aiming at the problem that the selection of the electrical parameters required by the current intelligent algorithm mostly depends on experience, provides a combined weighting method combining a mutual information theory and an analytic hierarchy process, orders the influence weights of different electrical parameters, determines the optimal input parameters, and improves the theoretical line loss calculation performance of the power distribution network.
Description
Technical Field
The invention belongs to the technical field of theoretical line loss calculation of a power distribution network, and particularly relates to a power distribution network line loss calculation method based on a combined weighting method and deep learning.
Background
The line loss is a key index for reflecting the operation level of the power grid and is an important assessment standard for reflecting the management level of a power grid company. The theoretical line loss calculation is used as an evaluation means of the line loss management level, is a powerful tool for analyzing line loss influence factors and formulating loss reduction measures by a power grid company, and aims to improve the comprehensive management level of the power distribution network and improve the economic benefit of the power grid company.
At present, a plurality of scholars propose a method for calculating theoretical line loss, which mainly comprises an equivalent algorithm based on a physical model of a power distribution network and a neural network model algorithm based on feeder data of the power distribution network. However, the former theoretical line loss calculation methods such as an equivalent resistance method, a root mean square current method and the like depend on a network structure, and the requirement on the sampling frequency of data is high. And the number of branch lines and distribution transformers of the power distribution network is large, so that the number of nodes and equivalent elements of the lines is increased, and the calculation difficulty is increased. Compared with the traditional theoretical line loss calculation method, the intelligent algorithm of the neural network model based on the feeder line data of the power distribution network can evaluate the theoretical line loss more conveniently and rapidly by fitting the nonlinear relation between the electrical parameters and the theoretical line loss. However, most of the electric parameters required by the existing method depend on experience, theoretical line loss influence factors are not fully considered, the difference of the influence degrees of different electric parameters is not considered, targeted electric parameter selection is carried out, and a targeted theoretical line loss calculation model is established.
Therefore, there is a need in the art for a new solution to solve this problem.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the power distribution network line loss calculation method based on the combination weighting method and the deep learning is provided for solving the technical problem that in the prior art, aiming at theoretical line loss calculation, pertinence analysis is carried out depending on poor fitting effect of a power distribution network topological structure and a shallow neural network.
A power distribution network line loss calculation method based on a combined weighting method and deep learning comprises the following steps which are sequentially performed,
selecting electrical parameters as an original feature set, respectively collecting sample data of each electrical parameter, and carrying out normalization processing on the electrical parameter sample data and theoretical line loss obtained by calculation of an equivalent resistance method according to a normalization formula to respectively obtain normalization values, wherein the normalization formula is as follows:
in the formula: x' i The actual value of the ith electrical parameter; x is a radical of a fluorine atom i Is a normalized value;
step two, constructing a GRU network calculation model of a deep learning gating cycle unit, and determining an activation function, the number of hidden layers and the number of neurons of the optimal GRU network calculation model;
thirdly, theoretical line loss calculated through the normalized values of the electrical parameters obtained in the first step and an equivalent resistance method is calculated, a combined assignment weight is calculated and obtained by utilizing an influence weight formula of the electrical parameters on the theoretical line loss based on a combined weighting method combining an analytic hierarchy process and a mutual information theory,
constructing different numbers of electrical parameters as input sets according to the sequence of the weights from large to small, respectively carrying out training test and verification on a theoretical line loss calculation model of the GRU network, determining the electrical parameter set with the minimum line loss calculation result error as an optimal input parameter set,
wherein, the influence weight lambda of each electrical parameter on the theoretical line loss i Is composed of
λ i ＝εW i +(1-ε)γ i
In the formula: epsilon is the specific gravity of the weight obtained by the analytic hierarchy process in the combined method, and 0.5 is taken; i is the ith electrical parameter; gamma ray i Weighting each electric parameter obtained for the mutual information theory; w is a group of i Weighting each electrical parameter obtained by an analytic hierarchy process;
step four, obtaining a final GRU network calculation model of the deep learning gating cycle unit according to the weight of each electric parameter obtained in the step three,
in practical use, an input set is formed by newly acquired line parameters according to the optimal input parameter set, the input set is used as the input of a theoretical line loss calculation model of the GRU network, the theoretical line loss calculation model of the GRU network trained in the step three is utilized to calculate and obtain the line loss value of the corresponding line,
and further determining the effectiveness of the theoretical line loss calculation method of the power distribution network based on the GRU network by taking the calculation result of the equivalent resistance method as a reference and according to the fact that the error of the calculation result is smaller than a set threshold value.
The electrical parameters selected in the first step comprise monthly active power supply quantity, monthly reactive power supply quantity, total line length, trunk line length, branch line length, distribution transformer capacity and line equivalent sectional area.
The calculation model formula of the GRU network in the second step is as follows:
r t ＝sigmoid(w rh h t-1 +w rx x t +b r )
z t ＝sigmoid(w zh h t-1 +w zx x t +b z )
g t ＝tanh(w gh (r t ·h t-1 )+w gx x t +b g )
h t ＝(1-z t )·h t-1 +z t ·g t
in the formula: w is a rh 、w rx To reset the gate weights; w is a zh 、w zx To update the door weight; w is a gh 、w gx To form a current memory state g t A weight of time; b is a mixture of r 、b z 、b g Is a bias matrix; sigmoid and tanh are activation functions; the operator "·" represents the inner product of the vector.
And in the second step, a sigmoid function and a tanh function are adopted as an activation function based on the GRU network theoretical line loss calculation model.
The determination process of the number of hidden layers and the number of the neurons in the step two is as follows: distributing multiple distribution lines according to a set proportion, selecting one part to train a theoretical line loss calculation model of the GRU network, testing the other part, respectively obtaining corresponding line loss values by changing the number of neurons on the basis of a single-layer hidden layer,
respectively obtaining the number of the corresponding neurons when the line loss numerical error MAPE and the RMSE are minimum by utilizing an average absolute error percentage MAPE formula and a root mean square error RMSE formula; the number of neurons of a first hidden layer is fixed, the number of layers of the hidden layer is increased layer by layer, the optimal number of neurons corresponding to each hidden layer is determined by using an average absolute error percentage MAPE (maximum amplitude distribution) formula and a Root Mean Square Error (RMSE) formula again, the number of layers of the hidden layer is continuously increased, and the number of layers of a critical hidden layer before a theoretical line loss calculation model of the GRU network is fitted and the optimal number of neurons corresponding to each hidden layer are selected as optimal GRU network parameters.
The calculation result errors in the second step and the fourth step are average absolute error percentage MAPE and root mean square error RMSE based on the theoretical line loss calculation value of the equivalent resistance method, and the formulas are respectively as follows:
in the formula:
calculated value of theoretical line loss, y, for GRU network ij Calculating a theoretical line loss value of an equivalent resistance method; and m is the number of samples in the test set.
The method for calculating the weight of each electric parameter obtained by the mutual information theory in the third step comprises the following steps:
in the theoretical line loss calculation, X is set as any electrical parameter in the original characteristic set, Y is a theoretical line loss calculation value, and mutual information I (X; Y) between the X and Y is as follows:
in the formula: p (X), p (Y) represent the marginal probability functions of the variables X, Y; p (X, Y) represents a joint probability function of the variables X, Y;
the weight γ of the influence of each electrical parameter on the theoretical line loss i Respectively as follows:
in the formula: i is the ith electrical parameter; k is the sum of theoretical linesThe number of loss-related electrical parameters; I.C. A i Is a mutual information value.
The method for calculating the weight of each electrical parameter obtained by the analytic hierarchy process in the third step comprises the following steps:
1) Building a hierarchical model
Selecting electrical parameters which influence theoretical line loss in an original characteristic set, and forming a hierarchical structure from large to small according to the relation of each electrical parameter to serve as an evaluation parameter set;
2) Establishing a decision matrix P
Comparing the relative importance of the ith electric parameter and the jth electric parameter in the evaluation parameter set, and assigning a value as a according to an importance comparison scale ij And further construct a decision matrix P, wherein a ji ＝1/a ij ，a ii ＝1；
3) Consistency check
The consistency check of the judgment matrix P is carried out by utilizing the check coefficient CR, and is used for judging whether the matrix is suitable for the hierarchical analysis,
the check coefficient CR is:
in the formula: CI and RI are consistency indexes and average random consistency indexes of the judgment matrix P respectively, CR is less than 0.1, which indicates that the judgment matrix P meets the consistency test;
4) Calculating a weight vector
After consistency check, calculating and obtaining a characteristic vector corresponding to the maximum characteristic root of the judgment matrix P, and obtaining a weight vector W which is the weight of each electrical parameter after normalization processing.
Through the design scheme, the invention can bring the following beneficial effects:
according to the method, subjective and objective factors are comprehensively considered, a combined weighting method combining a mutual information theory and an analytic hierarchy process is provided, and the influence weight of the selected electrical parameters on theoretical line loss is determined; establishing different numbers of input parameter sets according to the weight, and screening out an optimal input parameter set through calculation errors of the GRU grids when different electrical parameters are input; the theoretical line loss of the power distribution network is calculated through the trained GRU network, and the effectiveness of the method is verified by comparing the calculation results of the equivalent resistance method.
Drawings
The invention is further described in the following detailed description in conjunction with the drawings in which:
fig. 1 is a structural diagram of a GRU network in an embodiment of a power distribution network line loss calculation method based on a combined weighting method and deep learning according to the present invention.
Fig. 2 is a topological structure of a part of 10kV line a in an embodiment of a power distribution network line loss calculation method based on a combined weighting method and deep learning.
Fig. 3 is a topological structure of a part of 10kV line b in an embodiment of a power distribution network line loss calculation method based on a combined weighting method and deep learning according to the present invention.
Fig. 4 is a graph of a relationship between the number of neurons and a theoretical line loss calculation error in an embodiment of a power distribution network line loss calculation method based on a combined weighting method and deep learning of the present invention, in which a hidden layer of a GRU network is a single layer.
Fig. 5 is a graph of a relation between the number of neurons and a theoretical line loss calculation error in an embodiment of a power distribution network line loss calculation method based on a combined weighting method and deep learning of the present invention, in which a hidden layer of a GRU network is a double layer.
Fig. 6 is a comparison diagram of a GRU network calculation result and an equivalent resistance method calculation result in an embodiment of the power distribution network line loss calculation method based on a combined weighting method and deep learning.
Detailed Description
A power distribution network line loss calculation method based on a combined weighting method and deep learning comprises the following steps:
selecting 7 electrical parameters of monthly active and reactive power supply quantity, total line length, trunk line length, branch line length, distribution transformer capacity and line equivalent sectional area as an original characteristic set, and carrying out normalization processing on the 7 electrical parameters and theoretical line loss obtained by calculation of an equivalent resistance method according to a formula (1);
in the formula: x' i The actual value of the ith electrical parameter; x is a radical of a fluorine atom i Is a normalized value.
Step two, constructing a deep learning GRU network calculation model, and setting parameters such as an activation function, the number of hidden layers and the number of neurons of the GRU network calculation model, wherein the activation function adopts a sigmoid function and a tanh function, and the number of the hidden layers and the number of the neurons need to be continuously adjusted until the parameters are finally determined;
(1) Deep learning GRU network principle
A gated round robin unit (GRU) network is a typical deep learning model that has more efficient learning and nonlinear fitting capabilities than traditional machine learning methods. The structure of the GRU network is shown in FIG. 1, where x t 、r t 、z t 、g t And h t Respectively the input, reset gate, update gate, memory state and output of the hidden layer at the current time t, h t-1 Updating the gate z for the output of the hidden layer at the previous time t Controlling the memory degree of the previous hidden layer output information, resetting the gate r t The neglect degree of the hidden layer output information at the previous moment is controlled. Updating the door z t And a reset gate r t Determining whether the neuron is activated or not by an activation function, and inputting a state x into a hidden layer at the current moment t And the previous time hidden layer output h t-1 And performing operation. Reset gate r t Output and current input x t Forming a current memory state g by the operation of the activation function t Updating the door z t By controlling discard h t-1 Amount of information in and introduction g t The output h of the hidden layer at the current moment is obtained according to the amount of the medium information t And passes on to the next GRU unit. The detailed calculation formula is as follows:
r t ＝sigmoid(w rh h t-1 +w rx x t +b r ) (2)
z t ＝sigmoid(w zh h t-1 +w zx x t +b z ) (3)
g t ＝tanh(w gh (r t ·h t-1 )+w gx x t +b g ) (4)
h t ＝(1-z t )·h t-1 +z t ·g t (5)
in the formula: w is a rh 、w rx To reset the gate weight; w is a zh 、w zx To update the door weight; w is a gh 、w gx To form a current memory state g t A weight of time; b r 、b z 、b g Is a bias matrix; sigmoid and tanh are activation functions; the operator "·" represents the inner product of the vector.
(2) Determining GRU network parameters
The activation function based on the GRU network theoretical line loss calculation model adopts a sigmoid function and a tanh function, and the number of hidden layers and the number of neurons need to be determined by testing the GRU network theoretical line loss calculation error. Training and testing a GRU network according to a certain proportion by a power distribution line, and searching the number of neurons corresponding to the minimum test results MAPE and RMSE by changing the number of the neurons on the basis of a single-layer hidden layer; the number of neurons in the first hidden layer is fixed, the number of layers of the hidden layer is increased layer by layer, and the optimal number of neurons corresponding to each hidden layer is determined, so that the optimal GRU network parameters are determined.
Thirdly, calculating and obtaining combined assignment weight by using an influence weight formula of the electric parameters on theoretical line loss based on a combined weighting method combining an analytic hierarchy process and a mutual information theory,
and constructing different numbers of input parameter sets according to the sequence of the weights from large to small, and respectively training and testing the theoretical line loss calculation model of the GRU network. The electrical parameter set with the minimum line loss calculation result Error is determined as the optimal input parameter set by comparing the average absolute Error Percentage (MAPE) and the Root Mean Square Error (RMSE) of the calculation results by formula (6) and formula (7). The calculation result errors are average absolute error percentage MAPE and root mean square error RMSE based on theoretical line loss calculation values of an equivalent resistance method, and calculation result error formulas are respectively as follows:
in the formula:
calculated value of theoretical line loss, y, for GRU network ij Calculating a theoretical line loss value of an equivalent resistance method; m is the number of samples in the test set.
The optimal input parameter set determining method comprises the following steps:
(1) Original input parameter set for theoretical line loss calculation
According to the easy acquireability of the electrical parameters, the correlation among indexes and the contribution degree to the line loss influence, 7 electrical parameters of monthly active power supply quantity, reactive power supply quantity, total line length, trunk line length, branch line length, distribution transformer capacity and line equivalent sectional area are selected as input original feature sets.
(2) Theory of mutual information
In the theoretical line loss calculation, let X be the above 7 electrical parameters, and Y be the theoretical line loss calculation, and the mutual information I (X; Y) between the two is:
in the formula: p (X), p (Y) represent the marginal probability functions of the variables X, Y; p (X, Y) represents the joint probability function of the variables X, Y.
The weight γ of the influence of the different electrical parameters on the theoretical line loss i Comprises the following steps:
in the formula: i is the ith electrical parameter; k is the number of electrical parameters related to theoretical line loss; I.C. A i Is a mutual information value.
(3) Analytic hierarchy process
The method comprises the following steps of taking theoretical line loss as an evaluation target, obtaining the influence weight of each electrical parameter on the theoretical line loss through a chromatographic analysis method, and calculating the weight through the following steps:
1) Building a hierarchical model
And determining electrical parameters influencing theoretical line loss, and forming a hierarchical structure according to the relationship of each parameter so as to determine an evaluation parameter set.
2) Establishing a decision matrix P
Comparing the relative importance of the ith electrical parameter to the jth electrical parameter, assigning an importance value of a to the importance comparison scale shown in Table 1 ij Further construct a decision matrix P, where a ji ＝1/a ij ，a ii ＝1；
TABLE 1 importance comparison Scale
3) Consistency check
Due to the complexity of objective objects and the difference in different experts' recognitions, in order to examine whether the judgment matrix is suitable for hierarchical analysis, the consistency of the judgment matrix P needs to be checked by using a check coefficient CR. The test factor CR is:
in the formula: CI and RI are consistency indexes and average random consistency indexes of the judgment matrix P, respectively. When CR is less than 0.1, the judgment matrix P is satisfied with the consistency check.
4) Calculating a weight vector
After consistency check, calculating a characteristic vector corresponding to the maximum characteristic root of the judgment matrix P, and obtaining a weight vector W which is the weight of each electrical parameter after normalization processing.
(4) Combined empowerment method
The mutual information theory has objectivity, but when the weight of the electric parameters and the theoretical line loss influence is calculated, the electric parameters are processed in an equal weight mode and are not consistent with the ohm law and the line loss law; the analytic hierarchy process has subjectivity, and a judgment matrix P can be constructed according to expert experience to determine the weight of each electrical parameter.
Based on the characteristics of the two methods, a combined weighting method is provided, the two methods are effectively combined, and the influence weight lambda of each electrical parameter on the theoretical line loss is determined i ：
λ i ＝εW i +(1-ε)γ i (11)
In the formula: epsilon is the proportion of the weight obtained by the analytic hierarchy process in the combined process, and 0.5 is taken; i is the ith electrical parameter; gamma ray i Weighting each electric parameter obtained for the mutual information theory; w is a group of i And weighting each electric parameter obtained by the analytic hierarchy process.
And step four, obtaining a final deep learning gated circulation unit GRU network calculation model according to the weight of each electric parameter obtained in the step three, forming an input set according to an optimal input parameter set for newly collected line parameters during practical use, using the input set as the input of the final deep learning gated circulation unit GRU network calculation model, calculating and obtaining line loss values of corresponding lines by using the final deep learning gated circulation unit GRU network calculation model trained in the step three, and determining the effectiveness of the theoretical line loss calculation method of the power distribution network based on the GRU network according to a line loss calculation result error formula by taking the calculation result of an equivalent resistance method as reference.
Example (b):
taking a 10kV power distribution network in a certain northeast region as an example, according to the structure and the electricity consumption properties of the lines, historical data of 25 lines are selected as samples for building a theoretical line loss calculation model of the GRU network, and the topological structures of part of the lines are shown in fig. 2 and fig. 3.
(1) Determining GRU network parameters
In order to determine the number of hidden layers and the number of neurons in each layer of the GRU network, the 25 lines are trained and tested according to the proportion of 4. As can be seen from fig. 4, when the hidden layer is a single layer and the number of neurons is 10, the line loss calculation results have the minimum MAPE and RMSE, which are 16.678 and 0.551 respectively; based on the number of the neurons in the first layer as 10, a hidden layer is added. As can be seen from fig. 5, when the number of neurons in the hidden layer of the second layer is 10, the calculated MAPE and RMSE are minimum, 16.132 and 0.493, respectively. When the number of hidden layers is continuously increased, due to the increase of network parameters, the GRU network model tends to be over-fitted, and the calculation error becomes large. Therefore, the number of hidden layers of the theoretical line loss calculation model of the power distribution network is finally determined to be 2, and the number of neurons in each layer is 10.
(2) Determining optimal input parameters
In order to determine the optimal input parameter set of the theoretical line loss calculation model of the GRU network, the influence weights of 7 electrical parameters of monthly active and reactive power supply quantity, total line length, trunk line length, branch line length, distribution transformer capacity and equivalent line sectional area on the theoretical line loss are calculated according to a formula (12) and are sorted, and the results are shown in Table 2.
TABLE 2 electric parameters ranked by weight
According to the calculation results in table 2, different numbers of input parameter sets are constructed in the order of the weights from large to small, and the sample data trains and tests the GRU network calculation model according to the proportion of 4. Table 3 shows the test set calculation errors with the equivalent resistance method calculation results as reference: the result is optimal when the monthly active and reactive power supply quantity, the total line length and the capacity of the distribution transformer are used as the input of the theoretical line loss calculation model of the distribution network.
TABLE 3 calculation error of GRU network under different input parameters
In order to test the application effect of the theoretical line loss calculation model of the GRU network, the monthly active power supply quantity and the monthly reactive power supply quantity of the 25 lines are changed, and 4 parameters of the monthly active power supply quantity, the monthly reactive power supply quantity, the total line length and the capacity of a distribution transformer of the lines are input into the GRU network, so that the obtained output is the calculated line loss value of the GRU network. Meanwhile, the theoretical line loss of the 25 lines is calculated by adopting an equivalent resistance method. Comparing the calculation results of the two methods in fig. 6, it can be seen that the calculation result of the GRU network is very close to the calculation result of the equivalent resistance method, which indicates that the proposed method can replace the equivalent resistance method to perform theoretical line loss calculation.
In order to embody the practicability of the method, the calculation result of the theoretical line loss of the GRU network is compared with the calculation results of an equivalent resistance method and a BP algorithm, and the comparison result is shown in a table 4.
TABLE 4 theoretical line loss calculation results
Table 5 shows the calculation errors of the two methods of the GRU network and the conventional BP algorithm, where MAPE and RMSE of the theoretical line loss calculation result of the GRU network are respectively 6.295 and 0.218, and compared with the conventional BP algorithm, the GRU network has a better calculation effect by virtue of a strong nonlinear fitting capability.
TABLE 5 MAPE and RMSE of the results
In summary, aiming at the problem that the selection of the electrical parameters required by the current intelligent algorithm mostly depends on experience, the traditional mutual information theory and the analytic hierarchy process are combined, the weight of the electrical parameters relative to the theoretical line loss influence is calculated, and the optimal input parameter set of the calculation model is determined according to the weight, so that the effectiveness of the method is ensured, and the number of the required electrical parameters is reduced. In addition, compared with the traditional BP algorithm, the GRU deep learning model improves the nonlinear function approximation capability by means of a multilayer network structure, and has higher calculation precision.
Claims (8)
1. A power distribution network line loss calculation method based on a combined weighting method and deep learning is characterized by comprising the following steps: comprises the following steps which are sequentially carried out,
selecting electrical parameters as an original feature set, respectively collecting sample data of each electrical parameter, and carrying out normalization processing on the electrical parameter sample data and theoretical line loss obtained by calculation of an equivalent resistance method according to a normalization formula to respectively obtain normalization values, wherein the normalization formula is as follows:
in the formula: x' i The actual value of the ith electrical parameter; x is the number of i Is a normalized value;
step two, constructing a GRU network calculation model of a deep learning gating cycle unit, and determining an activation function, the number of hidden layers and the number of neurons of the optimal GRU network calculation model;
thirdly, theoretical line loss calculated through the normalized values of the electrical parameters obtained in the first step and an equivalent resistance method is calculated, a combined assignment weight is calculated and obtained by utilizing an influence weight formula of the electrical parameters on the theoretical line loss based on a combined weighting method combining an analytic hierarchy process and a mutual information theory,
constructing different numbers of electrical parameters as input sets according to the sequence of the weights from big to small, respectively carrying out training test and verification on a theoretical line loss calculation model of the GRU network, determining the electrical parameter set with the minimum error of a line loss calculation result as an optimal input parameter set,
wherein, the influence weight lambda of each electrical parameter on the theoretical line loss i Is composed of
λ i ＝εW i +(1-ε)γ i
In the formula: epsilon is the proportion of the weight obtained by the analytic hierarchy process in the combined process, and 0.5 is taken; i is the ith electrical parameter; gamma ray i Weighting each electric parameter obtained for the mutual information theory; w i Weighting each electrical parameter obtained by an analytic hierarchy process;
step four, obtaining a final deep learning gating cycle unit GRU network calculation model according to the weight of each electric parameter obtained in the step three,
in practice, an input set is formed by newly acquired line parameters according to the optimal input parameter set, the input set is used as the input of a theoretical line loss calculation model of the GRU network, the theoretical line loss calculation model of the GRU network trained in the step three is utilized to calculate and obtain the line loss value of the corresponding line,
and taking the calculation result of the equivalent resistance method as a reference, and further determining the effectiveness of the theoretical line loss calculation method of the power distribution network based on the GRU network according to the fact that the error of the calculation result is smaller than a set threshold value.
2. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 1, wherein the method comprises the following steps: the electrical parameters selected in the step one comprise monthly active power supply quantity, monthly reactive power supply quantity, total line length, trunk line length, branch line length, distribution transformer capacity and line equivalent sectional area.
3. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 1, wherein the method comprises the following steps: the calculation model formula of the GRU network in the second step is as follows:
r t ＝sigmoid(w rh h t-1 +w rx x t +b r )
z t ＝sigmoid(w zh h t-1 +w zx x t +b z )
g t ＝tanh(w gh (r t ·h t-1 )+w gx x t +b g )
h t ＝(1-z t )·h t-1 +z t ·g t
in the formula: w is a rh 、w rx To reset the gate weights; w is a zh 、w zx To update the door weight; w is a gh 、w gx To form a current memory state g t A weight of time; b r 、b z 、b g Is a bias matrix; sigmoid and tanh are activation functions; the operator "·" represents the inner product of the vector.
4. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 1, wherein: and in the second step, a sigmoid function and a tanh function are adopted as an activation function based on the GRU network theoretical line loss calculation model.
5. The power distribution network line loss calculation method based on the combined weighting method and the deep learning as claimed in claim 4, wherein: the determination process of the number of hidden layers and the number of the neurons in the step two is as follows: distributing multiple distribution lines according to a set proportion, selecting one part to train a theoretical line loss calculation model of the GRU network, testing the other part, respectively obtaining corresponding line loss values by changing the number of neurons on the basis of a single-layer hidden layer,
respectively obtaining the number of the corresponding neurons when the line loss numerical error MAPE and the RMSE are minimum by utilizing an average absolute error percentage MAPE formula and a root mean square error RMSE formula; the number of neurons of a first hidden layer is fixed, the number of layers of the hidden layer is increased layer by layer, the optimal number of neurons corresponding to each hidden layer is determined by using an average absolute error percentage MAPE (maximum amplitude distribution) formula and a Root Mean Square Error (RMSE) formula again, the number of layers of the hidden layer is continuously increased, and the number of layers of a critical hidden layer before a theoretical line loss calculation model of the GRU network is fitted and the optimal number of neurons corresponding to each hidden layer are selected as optimal GRU network parameters.
6. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 5, wherein: the calculation result errors in the second step and the fourth step are average absolute error percentage MAPE and root mean square error RMSE based on the theoretical line loss calculation value of the equivalent resistance method, and the formulas are respectively as follows:
in the formula:
calculated value of theoretical line loss, y, for GRU network ij Calculating a theoretical line loss value of an equivalent resistance method; m is the number of samples in the test set.
7. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 1, wherein: the method for calculating the weight of each electric parameter obtained by the mutual information theory in the third step comprises the following steps:
in the theoretical line loss calculation, X is set as any electrical parameter in the original characteristic set, Y is a theoretical line loss calculation value, and mutual information I (X; Y) between the X and Y is as follows:
in the formula: p (X), p (Y) represent the edge probability functions of the variables X, Y; p (X, Y) represents the joint probability function of the variables X, Y;
the weight γ of the influence of each electrical parameter on the theoretical line loss i Respectively as follows:
in the formula: i is the ith electrical parameter; k is the number of electrical parameters related to theoretical line loss; i is i Is a mutual information value.
8. The method for calculating the line loss of the power distribution network based on the combined weighting method and the deep learning as claimed in claim 1, wherein: the method for calculating the weight of each electrical parameter obtained by the analytic hierarchy process in the third step comprises the following steps:
1) Building a hierarchical model
Selecting electrical parameters which influence theoretical line loss in an original characteristic set, and forming a hierarchical structure from large to small according to the relation of each electrical parameter to serve as an evaluation parameter set;
2) Establishing a decision matrix P
Comparing the relative importance of the ith electric parameter and the jth electric parameter in the evaluation parameter set, and assigning a value as a according to an importance comparison scale ij And further construct a decision matrix P, wherein a ji ＝1/a ij ，a ii ＝1；
3) Consistency check
The consistency check of the judgment matrix P is carried out by utilizing the check coefficient CR, and is used for judging whether the matrix is suitable for the hierarchical analysis or not,
the check coefficient CR is:
in the formula: CI and RI are consistency indexes and average random consistency indexes of the judgment matrix P respectively, CR is less than 0.1, which indicates that the judgment matrix P meets the consistency test;
4) Calculating a weight vector
After consistency check, calculating and obtaining a feature vector corresponding to the maximum feature root of the judgment matrix P, and obtaining a weight vector W, namely the weight of each electrical parameter, after normalization processing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110045657.XA CN112883633B (en) | 2021-01-14 | 2021-01-14 | Power distribution network line loss calculation method based on combined weighting method and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110045657.XA CN112883633B (en) | 2021-01-14 | 2021-01-14 | Power distribution network line loss calculation method based on combined weighting method and deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112883633A CN112883633A (en) | 2021-06-01 |
| CN112883633B true CN112883633B (en) | 2022-10-11 |
Family
ID=76045817
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110045657.XA Active CN112883633B (en) | 2021-01-14 | 2021-01-14 | Power distribution network line loss calculation method based on combined weighting method and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112883633B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109614997A (en) * | 2018-11-29 | 2019-04-12 | 武汉大学 | A kind of stealing Risk Forecast Method and device based on deep learning |
| CN110334875A (en) * | 2019-07-10 | 2019-10-15 | 东北电力大学 | Meter and the wind power combined probability prediction technique of evaluation index conflict |
| CN110598854A (en) * | 2019-09-20 | 2019-12-20 | 国网福建省电力有限公司 | GRU model-based transformer area line loss rate prediction method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10573295B2 (en) * | 2017-10-27 | 2020-02-25 | Salesforce.Com, Inc. | End-to-end speech recognition with policy learning |
-
2021
- 2021-01-14 CN CN202110045657.XA patent/CN112883633B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109614997A (en) * | 2018-11-29 | 2019-04-12 | 武汉大学 | A kind of stealing Risk Forecast Method and device based on deep learning |
| CN110334875A (en) * | 2019-07-10 | 2019-10-15 | 东北电力大学 | Meter and the wind power combined probability prediction technique of evaluation index conflict |
| CN110598854A (en) * | 2019-09-20 | 2019-12-20 | 国网福建省电力有限公司 | GRU model-based transformer area line loss rate prediction method |
Non-Patent Citations (5)
| Title |
|---|
| DC Voltage Utilization Improvement to Enlarge Power Balance Constraint Range for Photovoltaic Cascaded Inverter;JUAN LI等;《IEEE Access》;20170228;第1-12页 * |
| 基于供电分区的中高压配电网综合评价策略;田春筝等;《电力系统保护与控制》;20181101(第21期);158-165 * |
| 基于门控循环神经网络的线损电量预测方法;李云冰等;《电力设备管理》;20200225(第02期);134-136 * |
| 改进组合赋权法的配电网隐患评估;姜媛媛等;《科学技术与工程》;20200808(第22期);193-198 * |
| 配电网线损计算与分析中存在问题解决对策;杨冬锋等;《吉林电力》;20121225(第06期);30-31、40 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112883633A (en) | 2021-06-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108038300B (en) | Optical fiber state evaluation method based on improved membership function combined with neural network | |
| CN107544253B (en) | Large missile equipment retirement safety control method based on improved fuzzy entropy weight method | |
| CN110991786B (en) | 10kV static load model parameter identification method based on similar daily load curve | |
| CN111950918B (en) | Market risk assessment method based on power transaction data | |
| CN108108875A (en) | A kind of distribution Running State fuzzy synthetic appraisement method based on cooperative game method and trapezoidal cloud model | |
| CN107274067B (en) | Distribution transformer overload risk assessment method | |
| CN107067341B (en) | RBFNN power distribution automation system state operation evaluation method based on multi-level entropy weight | |
| CN105956779A (en) | Power transformer running state evaluation method and power transformer running state evaluation device | |
| CN109377068B (en) | Helicopter performance analysis and evaluation method | |
| CN106656357B (en) | Power frequency communication channel state evaluation system and method | |
| CN109376933A (en) | Lithium ion battery negative material energy density prediction technique neural network based | |
| CN109684689A (en) | A kind of insulating gas comprehensive performance evaluation method of multi-parameter fusion | |
| CN109165807A (en) | A kind of energy quality comprehensive assessment method based on probabilistic neural network | |
| CN110705859A (en) | PCA-self-organizing neural network-based method for evaluating running state of medium and low voltage distribution network | |
| CN113688458B (en) | Foundation scheme optimization method based on analytic hierarchy process | |
| Yang et al. | Consistency improvement method of pairwise matrix based on consistency ratio decreasing rate and attribute weighting method considered decision makers’ levels in analytic hierarchy process: application to hip joint prosthesis material selection | |
| CN112990627A (en) | Electric energy quality evaluation method | |
| CN117370766A (en) | Satellite mission planning scheme evaluation method based on deep learning | |
| CN112883633B (en) | Power distribution network line loss calculation method based on combined weighting method and deep learning | |
| CN113591322A (en) | Low-voltage transformer area line loss rate prediction method based on extreme gradient lifting decision tree | |
| CN111932081B (en) | Method and system for evaluating running state of power information system | |
| CN116797096A (en) | Fuzzy comprehensive evaluation method for toughness level of supply chain based on AHP-entropy weight method | |
| CN114021905A (en) | Credit risk evaluation method for small and medium-sized enterprises | |
| CN113608136A (en) | Method for predicting health state of multi-scale lithium ion battery | |
| Zhao et al. | Optimizing radial basis probabilistic neural networks using recursive orthogonal least squares algorithms combined with micro-genetic algorithms |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |