CN106127387A - A kind of platform district based on BP neutral net line loss per unit appraisal procedure - Google Patents

Abstract

The present invention provides a kind of platform district based on BP neutral net line loss per unit appraisal procedure, including: step 1: be standardized platform district electric characteristic parameter processing；Step 2: building BP neural network model, Bing Duitai district line loss per unit is normalized；Step 3: according to BP neural network model assessment platform district line loss per unit.The present invention is by filtering out the platform district electric characteristic parameter relevant to the grid structure in platform district and load, use the non-linear relation between BP neural network learning electric characteristic index based on LM algorithm optimization and platform district line loss per unit, and then obtain BP neural network model, by BP neural network model assessment platform district line loss per unit.

Description

A kind of platform district based on BP neutral net line loss per unit appraisal procedure

Technical field

The present invention relates to a kind of appraisal procedure, be specifically related to line loss per unit assessment side of a kind of platform district based on BP neutral net Method.

Background technology

China's power distribution network is in large scale, structure is complicated, has a little many, line length, wide feature.Along with socioeconomic Exhibition, the increase of power load, the problem of low-voltage distribution net wire loss is more and more prominent, and it is left that line loss electricity and line loss per unit accounting reach 50% Right.Low-voltage distribution network platform district is as the end link in power system, and line loss per unit is one of important performance assessment criteria of unit-area management.As What accurately and easily calculates platform district line loss per unit, provides according to become power supply enterprise important for the rational reducing loss measure of formulation Business.

Owing to construction and the management condition of low-voltage platform area are uneven, platform district and terminal use's huge number, account management Incomplete, circuit complex distribution is various, the collection success rate difference of electricity consumption acquisition system is relatively big, at present to platform district line loss theory meter The method great majority calculated use loss of voltage method, substitutional resistance method etc., no matter platform district calculating theoretical line loss per unit or assessing system Meter line loss per unit is required to employ substantial amounts of human and material resources and just can collect necessary and sufficient operational data, and workload is very big, Cause each power supply department to be difficult to monthly and carry out an evaluation work.

In recent years, the development of neural network theory and application calculate for line loss per unit and provide new approach.Neutral net meter Calculate line loss per unit without founding mathematical models, it is possible to use self-learning ability, Generalization Ability and the Nonlinear Processing energy that network is powerful Power carrys out non-linear relation complicated between fit line loss rate and characteristic parameter, can quickly calculate hence with neural network algorithm With assessment platform district line loss per unit.

Summary of the invention

In order to overcome above-mentioned the deficiencies in the prior art, the present invention provides a kind of platform district based on BP neutral net line loss per unit to comment Estimate method, by filtering out the platform district electric characteristic parameter relevant to the grid structure in platform district and load, use based on LM algorithm Non-linear relation between the BP neural network learning electric characteristic index and the platform district line loss per unit that optimize, and then obtain BP nerve net Network model, by BP neural network model assessment platform district line loss per unit.

In order to realize foregoing invention purpose, the present invention adopts the following technical scheme that:

The present invention provides a kind of platform district based on BP neutral net line loss per unit appraisal procedure, and described method includes:

Step 1: be standardized platform district electric characteristic parameter processing；

Step 2: building BP neural network model, Bing Duitai district line loss per unit is normalized；

Step 3: according to BP neural network model assessment platform district line loss per unit.

In described step 1, if platform district number is N, the platform district electric characteristic parameter in each district is M, NGe Tai district sample Platform district electric characteristic parameter composition platform district electric characteristic vector X, have:

$X=\left(\begin{array}{cccccc}{x}_{11}& x_{12}& \mathrm{...}& x_{1j}& \mathrm{...}& x_{1M}\\ x_{21}& x_{22}& \mathrm{...}& x_{2j}& \mathrm{...}& x_{2M}\\ .& .& & .& & .\\ .& .& & .& & .\\ .& .& & .& & .\\ x_{i1}& x_{i2}& \mathrm{...}& x_{ij}& \mathrm{...}& x_{iM}\\ .& .& & .& & .\\ .& .& & .& & .\\ .& .& & .& & .\\ x_{N1}& x_{N2}& \mathrm{...}& x_{Nj}& \mathrm{...}& x_{NM}\end{array}\right)---\left(1\right)$

Wherein, x_ijFor the i-th row, the jth column element of platform district electric characteristic vector X, i=1,2 ..., N, j=1,2 ..., M；

It is standardized platform district electric characteristic parameter processing, has:

$Z_{ij}=\frac{x_{ij}-{\stackrel{\‾}{x}}_j}{\sqrt{s_{ij}}}---\left(2\right)$

${\stackrel{\‾}{x}}_j=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{x}_{ij}---\left(3\right)$

$s_{ij}=\frac{1}{N-1}\underset{i=1}{\overset{N}{\Σ}}{(x_{ij}-{\stackrel{\‾}{x}}_j)}^2---\left(4\right)$

Wherein, Z_ijFor x_ijAmount after standardization,For x_ijMeansigma methods, s_ijFor x_ijVariance.

Described step 2 comprises the following steps:

Step 2-1: structure BP neural network model:

BP neural network model includes input layer, hidden layer and output layer, and transmission function f (a) between each layer uses Logsig function, has:

$f\left(a\right)=\frac{1}{1+e^{-a}}---\left(5\right)$

Wherein, a is the independent variable of transmission function f (a) between each layer, 0 ＜ f (a) ＜ 1；

Step 2-2: platform district line loss per unit is normalized:

If platform district line loss per unit is d, uses formula (6) that platform district line loss per unit d is normalized, have:

$d_i^{\′}=\mathrm{\α}\frac{d_i-d_{min}+\mathrm{\β}}{d_{\max }-d_{min}+\mathrm{\β}}---\left(6\right)$

Wherein, d_i' for the value after the platform district line loss per unit normalization in i-th platform district, d_iFor the platform district line loss per unit in i-th platform district, d_minFor the minima of all district's line losses per unit, d_maxFor the maximum of all district's line losses per unit, α, β are constant, and 0.9 ＜ α ＜ 1,0 ＜ β ＜ 0.1.

Described step 3 comprises the following steps:

Step 3-1: utilize BP neural network model to Z_ijAnd d_i' carry out learning training；

Step 3-2: platform district electric characteristic parameter is substituted in BP neural network model, calculate platform district line loss per unit d.

In described step 3-1, for arbitrary district, owing to the platform district electric characteristic parameter in each district is M, thus defeated Enter layer and contain M BP neuron, then set input layer input vector as Z_r=(Z₁,Z₂,…,Z_m,…,Z_M)^T, the output of hidden layer Vector is Y_r=(y₁,y₂,...,y_p,...,y_P)^T, the output vector of output layer is o_r=(o₁,o₂,...,o_l,...o_L)^T, it is desirable to Output vector is d_r=(d₁,d₂,...,d_l,...,d_L)^T, wherein, T represents transposition, Z_mFor the m-th BP neuron of input layer, y_p For pth the BP neuron of hidden layer, o_lFor output layer the l BP neuron, d_lBeing the l desired throughput, P is hidden BP neuron number containing layer, L is the BP neuron number of output layer, m=1,2 ..., M, p=1,2 ..., P, l=1, 2,...,L；Input layer is respectively w to weights and the threshold value of hidden layer_mpAnd b_mp, weights and the threshold value of hidden layer to output layer are divided Wei w_plAnd b_pl；

Then, utilize BP neural network model to Z_ijAnd d_i' carry out learning training forward-propagating process as follows:

$y_p=f\left(\underset{m=1}{\overset{M}{\Σ}}\right(w_{mp}{Z}_m+b_{mp}\left)\right)---\left(7\right)$

$o_l=f\left(\underset{p=1}{\overset{P}{\Σ}}\right(w_{pl}{y}_p+b_{pl}\left)\right)---\left(8\right)$

Output error e is expressed as:

$\begin{array}{c}e=\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-o_l)}^2\\ =\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{pl}{y}_p+b_{pl}\right)\left)\right)}^2\\ =\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{pl}f\left(\underset{m=1}{\overset{M}{\Σ}}\left(w_{mp}{Z}_m+b_{mp}\right)\right)+b_{pl}\right)\left)\right)}^2\end{array}---\left(9\right)$

Utilize BP neural network model to Z_ijAnd d_i' carry out learning training back-propagation process as follows:

In (n+1)th iterative process, e is launched by Taylor's formula, obtains formula (10):

E (w (n+1))=e (w (n))+g^T(n)Δw(n)+0.5Δw^T(n)A(n)Δw(n) (10)

Wherein, w (n) is the weights during nth iteration, and e (w (n)) is output error during nth iteration；w(n + 1) being the weights during the (n+1)th generation, e (w (n+1)) is output error in (n+1)th iterative process；G (n) is gradient vector, T represents transposition；Δ w (n) is (n+1)th and the variable quantity of weights during n-th generation, i.e. Δ w (n)=w (n+1)-w (n), works as Δ W (n)=-A^-1N, time () g (n), e (w (n+1)) obtains minima；A (n) is Hessian matrix；

LM algorithm is used to be expressed as by Hessian matrix A (n):

A (n)=J^TJ (11)

Wherein, J is Jacobian matrix；

Gradient vector g (n) is expressed as:

G (n)=J^Te (12)

W (n+1) following formula correction:

W (n+1)=w (n)-[J^TJ+μI]^-1J^Te (13)

Wherein, I is unit vector, and μ is constant；

In like manner, threshold value b (n+1) the following formula correction during the (n+1)th generation:

B (n+1)=b (n)-[J^TJ+μI]^-1J^Te (14)

Wherein, the threshold value during b (n) was the n-th generation.

Compared with immediate prior art, the technical scheme that the present invention provides has the advantages that

1) self-learning ability, Generalization Ability and the Nonlinear Processing ability that the present invention utilizes network powerful carrys out fit line loss rate And non-linear relation complicated between characteristic parameter, it is not necessary to set up complicated mathematical model；

2) platform district based on the BP neutral net line loss per unit appraisal procedure that the present invention proposes has Fast Convergent, high accuracy etc. Advantage.

3) platform district based on the BP neutral net line loss per unit appraisal procedure that the present invention proposes can assess the platform do not surveyed The line loss per unit in district, provides technical support for platform district line loss analyzing.

Accompanying drawing explanation

Fig. 1 is platform district based on BP neutral net line loss per unit appraisal procedure flow chart in the embodiment of the present invention；

Fig. 2 is BP Artificial Neural Network Structures schematic diagram in the embodiment of the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is described in further detail.

Such as Fig. 1, the present invention provides a kind of platform district based on BP neutral net line loss per unit appraisal procedure, and described method includes:

Step 1: be standardized platform district electric characteristic parameter processing；

Before being standardized platform district electric characteristic parameter processing, first determining table district electric characteristic parameter；Platform district is electric Characteristic parameter includes reflecting the parameter of grid structure and the parameter relevant to load；The parameter of reflection grid structure includes power supply half Footpath, low-voltage circuit total length；The parameter relevant to load includes load factor, load nature of electricity consumed and ratio.

Choose Ge Tai district, somewhere 10, the platform district electric characteristic parameter such as table 1 in each district:

Table 1

Platform district	Radius of electricity supply R/ rice	Low-voltage circuit total length D/ rice	Load nature of electricity consumed and ratio EC/%	Load factor L/%
					Platform district 1	135	270	85.83	48
Platform district 2	140	290	95.07	39.42708
					Platform district 3	126	300	92.76	29.60417
Platform district 4	132	924	100.00	27.11694
					Platform district 5	145	1015	100.00	3.347414
Platform district 6	118	826	100.00	4.546424
					Platform district 7	109	763	72.09	4.356545
Platform district 8	122	854	85.75	6.297043
					Platform district 9	136	952	85.26	10.95262
Platform district 10	144	1008	88.67	7.236223

Electric characteristic index is the input as neural network algorithm, namely independent variable.Each parameter has different lists Position and magnitude, only distinguish the size of value data for neural network algorithm, can not reflect the unit of data.In order to Preferably apply above-mentioned algorithm, need to eliminate each parameter Jian Bu commensurate and the impact of magnitude logarithm value, prevent and " count greatly and eat Decimal " phenomenon.And the standardization of data is exactly by data bi-directional scaling, it is allowed to fall into a little specific interval removal number According to unit limit, be translated into nondimensional pure values.

If platform district number is N, the platform district electric characteristic parameter in each district is 4, and the platform district of NGe Tai district sample is the most special Levy parameter composition platform district electric characteristic vector X, have:

$X=\left(\begin{array}{cccc}{x}_{11}& x_{12}& x_{13}& x_{14}\\ x_{21}& x_{22}& x_{23}& x_{24}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ x_{N1}& x_{N2}& x_{N3}& x_{N4}\end{array}\right)---\left(1\right)$

Use x_ijExpression i-th row of platform district electric characteristic vector X, jth column element, i=1,2 ..., N, j=1,2 ..., 4；

It is standardized platform district electric characteristic parameter processing, has:

$Z_{ij}=\frac{x_{ij}-{\stackrel{\‾}{x}}_j}{\sqrt{s_{ij}}}---\left(2\right)$

${\stackrel{\‾}{x}}_j=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{x}_{ij}---\left(3\right)$

$s_{ij}=\frac{1}{N-1}\underset{i=1}{\overset{N}{\Σ}}{(x_{ij}-{\stackrel{\‾}{x}}_j)}^2---\left(4\right)$

Wherein, Z_ijFor x_ijAmount after standardization,For x_ijMeansigma methods, s_ijFor x_ijVariance；

It is standardized processing to the electric characteristic parameter in 10 Ge Tai districts, result such as table 2:

Table 2

Platform district	Radius of electricity supply R/ rice	Low-voltage circuit total length D/ rice	Load nature of electricity consumed and ratio EC/%	Load factor L/%
					Platform district 1	-0.78757	-0.703	-0.68032	0.989197
Platform district 2	-0.75876	-0.68834	0.513898	0.662709
					Platform district 3	-0.83943	-0.68101	0.215602	0.288615
Platform district 4	-0.80486	-0.22361	1.151219	0.193892
					Platform district 5	-0.72996	-0.1569	1.151219	-0.71134
Platform district 6	-0.88552	-0.29544	1.151219	-0.66568
					Platform district 7	-0.93738	-0.34162	-2.45674	-0.67291
Platform district 8	-0.86248	-0.27492	-0.69093	-0.59901
					Platform district 9	-0.78181	-0.20308	-0.75435	-0.42171
Platform district 10	-0.73572	-0.16203	-0.31359	-0.56324

Step 2: building BP neural network model, Bing Duitai district line loss per unit is normalized；Specifically have:

Step 2-1: structure BP neural network model:

BP neural network model includes input layer, hidden layer and output layer, and transmission function f (a) between each layer uses Logsig function, has:

$f\left(a\right)=\frac{1}{1+e^{-a}}---\left(5\right)$

Wherein, a is the independent variable of transmission function f (a) between each layer, 0 ＜ f (a) ＜ 1；

Step 2-2: platform district line loss per unit is normalized:

If platform district line loss per unit is d, uses formula (6) that platform district line loss per unit d is normalized, have:

$d_i^{\′}=\mathrm{\α}\frac{d_i-d_{min}+\mathrm{\β}}{d_{\max }-d_{min}+\mathrm{\β}}---\left(6\right)$

Wherein, d_i' for the value after the platform district line loss per unit normalization in i-th platform district, d_iFor the platform district line loss per unit in i-th platform district, d_minFor the minima of all district's line losses per unit, d_maxFor the maximum of all district's line losses per unit, α, β are constant, and 0.9 ＜ α ＜ 1,0 ＜ β ＜ 0.1, takes α=0.99, β=0.01 here；

Step 3: according to BP neural network model assessment platform district line loss per unit, specifically have:

Step 3-1: utilize BP neural network model to Z_ijAnd d_i' carry out learning training；

The learning training process of BP algorithm is made up of the forward-propagating of signal and two parts of back propagation of error.Forward Propagation refers to that inputting sample inputs from input layer, successively processes through each hidden layer and is transmitted to output layer.If output layer output result It is not reaching to expected value, then forwards the back propagation of error to.Error back propagation is with some form by hidden by output error Containing layer successively anti-pass, each neuron weights and threshold value are adjusted.The process that weights and threshold value constantly adjust is exactly network Learning training process, until error reaches desired extent or reaches the study number of times set.

In described step 3-1, such as Fig. 2, for arbitrary district, owing to the platform district electric characteristic parameter in each district is 4, So input layer contains 4 BP neurons, output layer contains 1 BP neuron, then sets input layer input vector as Z_r=(Z₁, Z₂,Z₃,Z₄)^T, the output vector of hidden layer is Y_r=(y₁,y₂,...,y_p,...,y_P)^T, the output vector of output layer is o_r=o₁, Desired output vector is d_r=d₁, wherein, T represents transposition, y_pFor pth the BP neuron of hidden layer, P is the BP god of hidden layer Through unit number, p=1,2 ..., P；Input layer is respectively w to weights and the threshold value of hidden layer_mpAnd b_mp, hidden layer is to output layer Weights and threshold value be respectively w_p1And b_p1；

Then, utilize BP neural network model to Z_ijAnd d_i' carry out learning training forward-propagating process as follows:

$y_p=f\left(\underset{m=1}{\overset{4}{\Σ}}\right(w_{mp}{Z}_m+b_{mp}\left)\right)---\left(7\right)$

$o_1=f\left(\underset{p=1}{\overset{P}{\Σ}}\right(w_{p1}{y}_p+b_{p1}\left)\right)---\left(8\right)$

Output error e is expressed as:

$\begin{array}{c}e=\frac{1}{2}{(d_1-o_1)}^2\\ =\frac{1}{2}{(d_1-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{p1}{y}_p+b_{p1}\right)\left)\right)}^2\\ =\frac{1}{2}{(d_1-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{p1}f\left(\underset{m=1}{\overset{4}{\Σ}}\left(w_{mp}{Z}_m+b_{mp}\right)\right)+b_{p1}\right)\left)\right)}^2\end{array}---\left(9\right)$

Utilize BP neural network model to Z_ijAnd d_i' carry out learning training back-propagation process as follows:

In (n+1)th iterative process, e is launched by Taylor's formula, obtains formula (10):

E (w (n+1))=e (w (n))+g^T(n)Δw(n)+0.5Δw^T(n)A(n)Δw(n) (10)

Wherein, w (n) is the weights during nth iteration, and e (w (n)) is output error during nth iteration；w(n + 1) being the weights during the (n+1)th generation, e (w (n+1)) is output error in (n+1)th iterative process；G (n) is gradient vector, T represents transposition；Δ w (n) is (n+1)th and the variable quantity of weights during n-th generation, i.e. Δ w (n)=w (n+1)-w (n), works as Δ W (n)=-A^-1N, time () g (n), e (w (n+1)) obtains minima；A (n) is Hessian matrix；

LM algorithm is used to be expressed as by Hessian matrix A (n):

A (n)=J^TJ (11)

Wherein, J is Jacobian matrix；

Gradient vector g (n) is expressed as:

G (n)=J^Te (12)

W (n+1) following formula correction:

W (n+1)=w (n)-[J^TJ+μI]^-1J^Te (13)

Wherein, I is unit vector, and μ is constant；

In like manner, threshold value b (n+1) the following formula correction during the (n+1)th generation:

B (n+1)=b (n)-[J^TJ+μI]^-1J^Te (14)

Wherein, the threshold value during b (n) was the n-th generation.

Step 3-2: platform district electric characteristic parameter is substituted in BP neural network model, calculate platform district line loss per unit d.

10 Ge Tai districts are calculated, actual value, estimated value, absolute error and the relative error of platform district line loss per unit such as table 3:

Table 3

Platform district	Practical line loss rate	Line loss per unit estimated value	Absolute error	Relative error
					Platform district 1	3.06	3.061377	0.00	0.05%
Platform district 2	4.42	4.328309	0.09	2.07%
					Platform district 3	5.22	5.249034	0.03	0.56%
Platform district 4	4.00	3.990267	0.01	0.31%
					Platform district 5	4.00	4.089978	0.09	2.18%
Platform district 6	4.20	4.266087	0.07	1.69%
					Platform district 7	4.10	4.103187	0.00	0.02%
Platform district 8	4.00	4.035961	0.04	0.91%
					Platform district 9	3.99	4.154729	0.16	4.09%
Platform district 10	4.10	4.08458	0.01	0.27%

Finally should be noted that: above example only in order to illustrate that technical scheme is not intended to limit, institute The those of ordinary skill in genus field still the detailed description of the invention of the present invention can be modified with reference to above-described embodiment or Equivalent, these are without departing from any amendment of spirit and scope of the invention or equivalent, all await the reply in application this Within bright claims.

Claims (5)

1. platform district based on a BP neutral net line loss per unit appraisal procedure, it is characterised in that: described method includes:

Step 1: be standardized platform district electric characteristic parameter processing；

Step 2: building BP neural network model, Bing Duitai district line loss per unit is normalized；

Step 3: according to BP neural network model assessment platform district line loss per unit.

Platform district based on BP neutral net the most according to claim 1 line loss per unit appraisal procedure, it is characterised in that: described step In rapid 1, if platform district number is N, the platform district electric characteristic parameter in each district is M, the platform district electric characteristic of NGe Tai district sample Parameter composition platform district electric characteristic vector X, has:

$X=\left(\begin{array}{cccccc}{x}_{11}& x_{12}& \mathrm{...}& x_{1j}& \mathrm{...}& x_{1M}\\ x_{21}& x_{22}& \mathrm{...}& x_{2j}& \mathrm{...}& x_{2M}\\ .& .& & .& & .\\ .& .& & .& & .\\ .& .& & .& & .\\ x_{i1}& x_{i2}& \mathrm{...}& x_{ij}& \mathrm{...}& x_{iM}\\ .& .& & .& & .\\ .& .& & .& & .\\ .& .& & .& & .\\ x_{N1}& x_{N2}& \mathrm{...}& x_{Nj}& \mathrm{...}& x_{NM}\end{array}\right)---\left(1\right)$

Wherein, x_ijFor the i-th row, the jth column element of platform district electric characteristic vector X, i=1,2 ..., N, j=1,2 ..., M；

It is standardized platform district electric characteristic parameter processing, has:

$Z_{ij}=\frac{x_{ij}-{\stackrel{\‾}{x}}_j}{\sqrt{s_{ij}}}---\left(2\right)$

${\stackrel{\‾}{x}}_j=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{x}_{ij}---\left(3\right)$

$s_{ij}=\frac{1}{N-1}\underset{i=1}{\overset{N}{\Σ}}{(x_{ij}-{\stackrel{\‾}{x}}_j)}^2---\left(4\right)$

Wherein, Z_ijFor x_ijAmount after standardization,For x_ijMeansigma methods, s_ijFor x_ijVariance.

Platform district based on BP neutral net the most according to claim 2 line loss per unit appraisal procedure, it is characterised in that: described step Rapid 2 comprise the following steps:

Step 2-1: structure BP neural network model:

BP neural network model includes input layer, hidden layer and output layer, and transmission function f (a) between each layer uses logsig Function, has:

$f\left(a\right)=\frac{1}{1+e^{-a}}---\left(5\right)$

Wherein, a is the independent variable of transmission function f (a) between each layer, 0 ＜ f (a) ＜ 1；

Step 2-2: platform district line loss per unit is normalized:

If platform district line loss per unit is d, uses formula (6) that platform district line loss per unit d is normalized, have:

$d_i^{\′}=\mathrm{\α}\frac{d_i-d_{min}+\mathrm{\β}}{d_{\max }-d_{min}+\mathrm{\β}}---\left(6\right)$

Wherein, d '_iFor the value after the platform district line loss per unit normalization in i-th platform district, d_iFor the platform district line loss per unit in i-th platform district, d_min For the minima of all district's line losses per unit, d_maxFor the maximum of all district's line losses per unit, α, β are constant, and 0.9 ＜ α ＜ 1,0 ＜ β ＜ 0.1.

Platform district based on BP neutral net the most according to claim 3 line loss per unit appraisal procedure, it is characterised in that: described step Rapid 3 comprise the following steps:

Step 3-1: utilize BP neural network model to Z_ijWith d '_iCarry out learning training；

Step 3-2: platform district electric characteristic parameter is substituted in BP neural network model, calculate platform district line loss per unit d.

Platform district based on BP neutral net the most according to claim 4 line loss per unit appraisal procedure, it is characterised in that: described step In rapid 3-1, for arbitrary district, owing to the platform district electric characteristic parameter in each district is M, so input layer contains M BP Neuron, then sets input layer input vector as Z_r=(Z₁,Z₂,...,Z_m,...,Z_M)^T, the output vector of hidden layer is Y_r= (y₁,y₂,...,y_p,…,y_P)^T, the output vector of output layer is o_r=(o₁,o₂,…,o_l,…o_L)^T, it is desirable to output vector is d_r =(d₁,d₂,...,d_l,…,d_L)^T, wherein, T represents transposition, Z_mFor the m-th BP neuron of input layer, y_pFor hidden layer P BP neuron, o_lFor output layer the l BP neuron, d_lBeing the l desired throughput, P is that the BP of hidden layer is neural Unit's number, L is the BP neuron number of output layer, m=1,2 ..., M, p=1,2 ..., P, l=1,2 ..., L；Input layer Weights and threshold value to hidden layer are respectively w_mpAnd b_mp, weights and the threshold value of hidden layer to output layer are respectively w_plAnd b_pl；

Then, utilize BP neural network model to Z_ijWith d '_iThe forward-propagating process carrying out learning training is as follows:

$y_p=f\left(\underset{m=1}{\overset{M}{\Σ}}\right(w_{mp}{Z}_m+b_{mp}\left)\right)---\left(7\right)$

$o_l=f\left(\underset{p=1}{\overset{P}{\Σ}}\right(w_{pl}{y}_p+b_{pl}\left)\right)---\left(8\right)$

Output error e is expressed as:

$\begin{array}{c}e=\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-o_l)}^2\\ =\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{pl}{y}_p+b_{pl}\right)\left)\right)}^2\\ =\frac{1}{2}\underset{l=1}{\overset{L}{\Σ}}{(d_l-f(\underset{p=1}{\overset{P}{\Σ}}\left(w_{pl}f\left(\underset{m=1}{\overset{M}{\Σ}}\left(w_{mp}{Z}_m+b_{mp}\right)\right)+b_{pl}\right)\left)\right)}^2\end{array}---\left(9\right)$

Utilize BP neural network model to Z_ijWith d '_iThe back-propagation process carrying out learning training is as follows:

In (n+1)th iterative process, e is launched by Taylor's formula, obtains formula (10):

E (w (n+1))=e (w (n))+g^T(n)Δw(n)+0.5Δw^T(n)A(n)Δw(n) (10)

Wherein, w (n) is the weights during nth iteration, and e (w (n)) is output error during nth iteration；w(n+1) Being the weights during the (n+1)th generation, e (w (n+1)) is output error in (n+1)th iterative process；G (n) is gradient vector, T table Show transposition；Δ w (n) is (n+1)th and the variable quantity of weights during n-th generation, i.e. Δ w (n)=w (n+1)-w (n), as Δ w (n)=-A^-1N, time () g (n), e (w (n+1)) obtains minima；A (n) is Hessian matrix；

LM algorithm is used to be expressed as by Hessian matrix A (n):

A (n)=J^TJ (11)

Wherein, J is Jacobian matrix；

Gradient vector g (n) is expressed as:

G (n)=J^Te (12)

W (n+1) following formula correction:

W (n+1)=w (n)-[J^TJ+μI]^-1J^Te (13)

Wherein, I is unit vector, and μ is constant；

In like manner, threshold value b (n+1) the following formula correction during the (n+1)th generation:

B (n+1)=b (n)-[J^TJ+μI]^-1J^Te (14)

Wherein, the threshold value during b (n) was the n-th generation.