CN105046377A - Method for screening optimum indexes of reservoir flood control dispatching scheme based on BP neural network - Google Patents

Abstract

The invention discloses a method for screening the optimum indexes of a reservoir flood control dispatching scheme based on a BP neural network, comprising the following steps: generating network training samples; determining the topology of the BP neural network; training and inspecting the BP neural network; identifying the importance of each index in the BP neural network; calculating the degree to which the change of each index value affects the result, and analyzing the sensitivity of each index value; and determining the criterion and threshold of index screening. A lot of network training samples are generated through an index uniform discretization method, which effectively guarantees the training precision of the BP neural network. Based on an information storage and conversion mechanism of the BP neural network, the indexes and the influence of relative change of the indexes to the result of reservoir flood control dispatching scheme assessment are quantitatively evaluated according to the relative degree of importance and relative rate of contribution. By establishing a comprehensive judgment index, index screening is converted from a subjective process of analysis and judgment into a quantitative process of analysis and calculation.

Description

Based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators

Technical field

The present invention relates to reservoir regulation for flood control technology, especially a kind of method of screening reservoir regulation for flood control scheme selection indicators.

Background technology

Reservoir regulation for flood control technology is the important non-engineering measure realizing Flood Prevention mitigation, closely related with factors such as society, economy, nature, ecologies.Comparatively general way is by the feasible scheme collection of Flood Control Dispatch model generation one group at present, the satisfactory solution of each Objective benefits of choosing comprehensively is therefrom selected to be put to decision-making by decision maker, can be summed up as Multiple Attribute Decision Problems, be also multi-scheme optimal selection problem.

To there is larger subjectivity in existing method random in the screening of index, and generally lack the identification process to index relative importance degree and sensitivity level, criterion and the corresponding Threshold of index screening are not yet effectively solved.

Therefore, the how effectively relative Link Importance of identifying index and sensitive information, filtering out from numerous and jumbled Flood Control Dispatch index system the significant index of evaluation result effect for evaluate alternatives modeling is a problem needing solution badly.

Summary of the invention

Goal of the invention a: object is to provide a kind of method based on BP neural network screening reservoir regulation for flood control scheme selection indicators, to solve the problems referred to above that prior art exists.

Technical scheme: a kind of method based on BP neural network screening reservoir regulation for flood control scheme selection indicators, comprises the steps:

Step 1. generating network training sample;

Step 2. determines the topological structure of BP neural network;

Step 3. is trained, check described BP neural network;

The significance level of each index in step 4. identification BP neural network;

Step 5. calculates the change of each desired value to the size of Influence on test result degree, analyzes the susceptibility of each desired value;

The criterion of step 6. agriculture products screening and threshold value.

Preferably, described step 1 is further:

Using desired value to be screened for each Flood Control Dispatch scheme as input, will finally be used for passing judgment on the Comprehensive Evaluation desired value of scheme quality as output; Wherein, described Comprehensive Evaluation desired value adopts and tries to achieve based on the multiple attributive decision making method of positive and negative ideal point concept; Adopt the method generating network training sample of desired value uniform discrete.

Preferably, described step 1 is further comprising the steps:

Step 11., according to reservoir Technique for Real-time Joint Operation of Flood model, generates the Flood Control Dispatch scheme collection meeting constraint condition:

{x _ij|i＝1,2,…,m；j＝1,2,…,n}，

Wherein i and j is respectively scheme sequence number and index sequence number, m and n is respectively scheme number and index number to be screened;

Step 12. determines positive ideal solution scheme { x _j ⁺| j=1,2 ..., n} and minus ideal result scheme { x _j ^-| j=1,2 ..., n};

${x_j}^{+}=\left\{\begin{array}{c}\underset{1\≤i\≤m}{\max }\left\{x_{ij}\right\}\\ \underset{1\≤i\≤m}{\min }\left\{x_{ij}\right\}\end{array}\right.,{x_j}^{-}=\left\{\begin{array}{c}\underset{1\≤i\≤m}{\min }\left\{x_{ij}\right\}\\ \underset{1\≤i\≤m}{\max }\left\{x_{ij}\right\}\end{array}\right.,$

Step 13. adopts the method for stochastic simulation to produce the input item of training sample:

Obey equally distributed random number u for stochastic generation k _g~ U (0,1), g=1,2,, k, carries out sliding-model control to each desired value between positive and negative ideal scheme, the input item of raw k the training sample of common property, k is natural number, and its numerical value is determined according to the network training accuracy requirement of reality, and discretize formula is as follows:

x _gj＝x _j ⁺-(x _j ⁺-x _j ^-)×u _g

Positive and negative ideal scheme is also taken as two input items of training sample, then the final training sample input item generated is designated as { x _tj| t=1,2 ..., p; J=1,2 ..., n}, wherein t is sample sequence number, and p is sample size, p=k+2;

Step 14. calculates the Euclidean distance of each sample apart from positive ideal scheme and minus ideal result scheme:

$\left\{\begin{array}{c}{d_t}^{+}=\sqrt{\underset{j=1}{\overset{n}{\Σ}}{({x_j}^{+}-x_{tj})}^2}\\ {d_t}^{-}=\sqrt{\underset{j=1}{\overset{n}{\Σ}}{({x_j}^{-}-x_{tj})}^2}\end{array}\right.,$

In formula, d _t ⁺be the Euclidean distance of t sample apart from positive ideal scheme; d _t ^-be the Euclidean distance of t sample apart from ill ideal solution;

Step 15. calculates exchange premium degree coefficient c _tspan be [0,1];

Step 16. couple training sample { x _tj| t=1,2 ..., p; J=1,2 ..., the input item of n} is normalized:

${x_{tj}}^{*}=\frac{x_{tj}-x_{j\min }}{x_{j\max }-x_{j\min }},$

In formula, x _tjand x _tj ^*be respectively calculated value and the normalized value of a jth index in t sample; x _jmaxand x _jminbe respectively maximal value and the minimum value of a jth index;

{ the x that determines that training sample is expressed as _tj ^*, c _t| t=1,2 ..., p; J=1,2 ..., n}.

Preferably, described step 2 is further:

Described BP neural network comprises input layer, single hidden layer and output layer;

Wherein, input layer number is the number of index to be screened, and output layer nodes is 1, and the number of hidden nodes is determined according to Lippmann experimental formula; The transport function of input layer number is linear function, and the transport function of hidden layer and output layer adopts tanh Sigmoid function;

Connect weights and threshold according to Delta learning rules to the network between each node layer to adjust, computational grid global error E:

$E=\underset{t=1}{\overset{p}{\Σ}}{E}_t=\underset{t=1}{\overset{p}{\Σ}}\frac{{(z_t-c_t)}^2}{2},$

E _tit is single network of samples training error of t sample; z _tand c _tbe respectively network real output value and the desired output of t sample.

Preferably, described step 3 is further:

Described training sample is divided into two parts, is respectively used to network training and network checks;

The main training parameter of setting BP neural network, comprises learning rate and factor of momentum;

Initial network connects weights and threshold at interval (-1,1) interior stochastic generation;

The end condition of setting network training;

Delta learning rules are adopted to train described BP neural network.

Preferably, described step 4 is further:

The relative Link Importance of definition index, the relative importance of a jth index is designated as

$\stackrel{\‾}{R_j}=\frac{1}{N}\underset{l=1}{\overset{N}{\Σ}}{R_j}^l,$

In formula, N is the number of times of network training; R _j ^lit is the relative Link Importance of a jth index in the l time training; The relative Link Importance R of a jth index _jfor:

$R_j=\underset{s=1}{\overset{h}{\Σ}}\left(w_{js}{w}_s\right)/\underset{j=1}{\overset{n}{\Σ}}\underset{s=1}{\overset{h}{\Σ}}\left|w_{js}{w}_s\right|,$

In formula, w _jsfor the network between input layer j and hidden node s connects weights; w _sfor the network between hidden node s and output layer node connects weights; N is input layer number; H is the number of hidden nodes.

Preferably, described step 5 is further:

Calculate the change of each index to the Relative Contribution rate of net result change, the change of a jth index is designated as G to the Relative Contribution rate that network output valve z changes _j,

$G_j=\frac{{z_j}^{x\mathrm{var}}}{\underset{j=1}{\overset{n}{\Σ}}{z_j}^{x\mathrm{var}}}\×100,{z_j}^{x\mathrm{var}}=\frac{{z_j}^{\mathrm{var}}}{\stackrel{\‾}{z}}$

In formula, for network real output value, for the absolute change amount of network output valve z.

Preferably, described step 6 is further:

Calculate comprehensive distinguishing index P _j,

$P_j=\lg \frac{F_j}{\stackrel{\‾}{F_j}},F_j=\left|\stackrel{\‾}{R_j}\right|\×G_j,$

In formula, for all index F _jmean value, and G _jbe respectively relative Link Importance and the Relative Contribution rate of a jth index, and G _jvalue be positioned in [0,1] interval.If P _j> 0, the then F of this index _jvalue is greater than average level; If P _j< 0, the then F of this index _jvalue is less than average level.Work as P _jwhen≤-1, then the F of this index _jvalue with differ a more than magnitude.

Implement technical scheme of the present invention, following beneficial effect can be obtained: first, generate a large amount of training samples by the method for desired value uniform discrete, effectively ensure that the training precision of BP neural network; Secondly, based on BP neural network to the storage of information with shift to new management mechanisms, the present invention adopts relative Link Importance and Relative Contribution rate to come qualitative assessment index itself and the relative impact changed reservoir regulation for flood control evaluate alternatives result thereof; Finally, by establishing comprehensive distinguishing index, the screening of index is converted into quantitative analytical calculation process from subjective analysis deterministic process.

Accompanying drawing explanation

Fig. 1: process flow diagram of the present invention.

The topological structure schematic diagram of Fig. 2: three layers of BP neural network.

In Fig. 2, the transport function of input layer adopts linear function, and the transport function of hidden layer and output layer node adopts tanh Sigmoid function; N is input layer number, and h is the number of hidden nodes, and z is approach degree coefficient.

Embodiment

For solving the problems referred to above that prior art exists, applicant has carried out deep, creative research work to prior art.

Applicant thinks: the index of reservoir regulation for flood control evaluate alternatives be a series of from the different hierarchies of objectivies and different evaluation angles (such as: reservoir safety, downstream security, power benefit, flood water resources utilization efficiency etc.) weigh scheduling scheme quality estimate set.And evaluate alternatives is in fact adopt certain mathematical model each desired value of each scheme to be integrated into a comprehensive evaluation value, and carry out schemes ranking with preferred according to the size of this comprehensive evaluation value.

Research finds that the research of existing reservoir regulation for flood control evaluate alternatives mainly lays particular emphasis on evaluation method and improvement thereof, the screening of index exists larger subjective arbitrariness, generally lack the identification process to index relative importance degree and sensitivity level, criterion and the corresponding Threshold of index screening are not yet effectively solved.Namely existing paper or patent documentation are not furtherd investigate the process of index screening, do not recognize that index screening is significant to scheme selection result.

And applicant thinks, in the process of whole Scheme Choice, the screening of index is particularly important link, because it has direct impact to final evaluate alternatives result, this impact is embodied in two aspects:

1) each index itself is to the contribution property of there are differences of scheme evaluation result, the importance degree being embodied in different index is different, some index is very large to the effect of evaluation result, and some index is then not remarkable to the effect of evaluation result, has obvious primary and secondary feature between index;

2) the response also property of there are differences of the comprehensive evaluation value caused by relative change of each index, is embodied in comprehensive evaluation value different to the sensitivity level of each index.

Therefore, in the impact just needing distinct these two aspects early stage of evaluate alternatives modeling, to guarantee rationality and the validity of index screening process.The how effectively relative Link Importance of identifying index and sensitive information, filtering out from numerous and jumbled Flood Control Dispatch index system the significant index of evaluation result effect for evaluate alternatives modeling is the technical matters that the present invention wishes to solve.

In order to solve the problem, the present invention proposes a kind of training sample generation method based on desired value discretize; According to BP neural network to the storage of information with shift to new management mechanisms, relative Link Importance and Relative Contribution rate two indices are proposed, for qualitative assessment index itself and the impact that relatively changes reservoir regulation for flood control evaluate alternatives result thereof; Comprehensive distinguishing index is proposed, the criterion of agriculture products screening and threshold value, for decision maker carries out the analytical calculation instrument that index screening provides quantitative.

Below by embodiment, and composition graphs 1 and Fig. 2, technical scheme of the present invention is done and illustrates further.

Reservoir regulation for flood control evaluating indexesto scheme screening technique based on BP neural network mainly comprises the following steps:

Step 1, generating network training sample:

BP neural metwork training is the significant process setting up Nonlinear Mapping relation between input and output, and for reservoir regulation for flood control evaluate alternatives, its main task makes the BP neural network after training possess the ability of making decisions on one's own.Using desired value to be screened for each Flood Control Dispatch scheme input as training sample, to finally be used for passing judgment on the output of comprehensive evaluation index value as training sample of scheme quality, comprehensive evaluation index value wherein adopts and tries to achieve based on the multiple attributive decision making method of positive and negative ideal point concept, i.e. approach degree coefficient.Consider that the scheme number generated in actual Flood Control Dispatch is limited often, adopt the method for desired value uniform discrete to generate a large amount of training samples, to meet the training precision requirement of BP neural network.

Step 1 specifically comprises following sub-step:

Step 11, according to reservoir Technique for Real-time Joint Operation of Flood model, generates the Flood Control Dispatch scheme collection { x meeting constraint condition _ij| i=1,2 ..., m; J=1,2 ..., n}, wherein i and j is respectively scheme sequence number and index sequence number, m and n is respectively scheme number and index number to be screened.

Step 12, determines positive ideal scheme { x _j ⁺| j=1,2 ..., n} and ill ideal solution { x _j ^-| j=1,2 ..., n}.Positive and negative ideal scheme refers to the idealized scheme that all desired values reach optimum or the most bad simultaneously, and they have demarcated the Upper-lower Limit state of scheme quality respectively, and its computing formula is as follows:

Step 13, adopts the input item of the method generating network training sample of stochastic simulation.Obey equally distributed random number u for stochastic generation k _g~ U (0,1), g=1,2,, k, carries out sliding-model control to each desired value between positive and negative ideal scheme, the input item of raw k the training sample of common property, the value of k can be determined according to the network training accuracy requirement of reality, and discretize formula is as follows:

x _gj＝x _j ⁺-(x _j ⁺-x _j ^-)×u _g

Positive and negative ideal scheme is also taken as two input items of training sample, then the final training sample input item generated is designated as { x _tj| t=1,2 ..., p; J=1,2 ..., n}, wherein t is sample sequence number, and p is sample size, p=k+2.

Step 14, calculates the Euclidean distance of each sample apart from positive and negative ideal scheme:

$\left\{\begin{array}{c}{d_t}^{+}=\sqrt{\underset{j=1}{\overset{n}{\&Sigma;}}{({x_j}^{+}-x_{tj})}^2}\\ {d_t}^{-}=\sqrt{\underset{j=1}{\overset{n}{\&Sigma;}}{({x_j}^{-}-x_{tj})}^2}\end{array}\right.$

In formula: d _t ⁺be the Euclidean distance of t sample apart from positive ideal scheme; d _t ^-be the Euclidean distance of t sample apart from ill ideal solution.

Step 15, calculates approach degree coefficient it can be used as the comprehensive evaluation index of reflection scheme quality.

C _tspan be [0,1], reflect the relative distance of t sample and positive and negative ideal scheme.C _tlarger, show that t sample is far away apart from ill ideal solution, nearer apart from positive ideal scheme, sample is more excellent; Otherwise it is then poorer.

Step 16, in order to avoid the dimension difference between index is on the impact of Network Recognition precision, to the input item { x of training sample _tj| t=1,2 ..., p; J=1,2 ..., n} is normalized:

${x_{tj}}^{*}=\frac{x_{tj}-x_{j\min }}{x_{j\max }-x_{j\min }},$

In formula: x _tjand x _tj ^*be respectively calculated value and the normalized value of a jth index in t sample; x _jmaxand x _jminbe respectively maximal value and the minimum value of a jth index.

Input as network using the desired value after each training sample normalization, export using approach degree coefficient as network, the training sample finally determined can be expressed as { x _tj ^*, c _t| t=1,2 ..., p; J=1,2 ..., n}.

Step 2, design BP neural network topology structure:

According to Kolmogorov theorem, set up three layers of BP neural network, comprise respectively: input layer, single hidden layer and output layer.Described BP neural network input layer nodes is the number of index to be screened, and output layer nodes is 1, i.e. approach degree coefficient, and the number of hidden nodes is determined according to Lippmann experimental formula.The transport function of input layer adopts linear function, and the transport function of hidden layer and output layer node adopts tanh Sigmoid function.Connect weights and threshold according to Delta learning rules to the network between each node layer to adjust, computational grid global error E:

$E=\underset{t=1}{\overset{p}{\&Sigma;}}{E}_t=\underset{t=1}{\overset{p}{\&Sigma;}}\frac{{(z_t-c_t)}^2}{2}$

In formula: E _tit is single network of samples training error of t sample; z _tand c _tbe respectively network real output value and the desired output of t sample.If network global error E is less than the accuracy value ε preset, or when frequency of training reaches setting value, then terminate training.

Step 3, BP neural metwork training and inspection:

Data sample step 1 generated is divided into two parts, and the data sample of front 80% is used for network training, and the data sample of rear 20% is used for network checks.The main training parameter of BP neural network is set to: learning rate gets 0.8, factor of momentum gets 0.5, initial network connects weights and threshold at interval (-1,1) interior stochastic generation, the end condition of network training be network global error be less than 0.0005 or frequency of training be greater than 30000.Described BP neural metwork training adopts Delta learning rules.

Step 4, index weight identification:

The network of BP neural network to the present each layer neuron node of the storage of information and translation table is connected on weights, for a network trained, can connect weights from network the relative Link Importance information extracting each input neuron (i.e. each index), the present invention defines the relative Link Importance R of a jth index _jfor:

$R_j=\underset{s=1}{\overset{h}{\&Sigma;}}\left(w_{js}{w}_s\right)/\underset{j=1}{\overset{n}{\&Sigma;}}\underset{s=1}{\overset{h}{\&Sigma;}}\left|w_{js}{w}_s\right|$

Wherein, w _jsfor the network between input layer j and hidden node s connects weights; w _sfor the network between hidden node s and output layer node connects weights; N is input layer number; H is the number of hidden nodes.

Consider that neural network initial network when each training connects weights and threshold and determines all at random, it is all not exclusively the same with the index relative Link Importance value calculated that the network after each training connects weights.Therefore, the present invention adopts R _javerage characterize the relative Link Importance of a jth index:

$\stackrel{\&OverBar;}{R_j}=\frac{1}{N}\underset{l=1}{\overset{N}{\&Sigma;}}{R_j}^l$

Wherein, N is the number of times of network training; R _j ^lit is the relative Link Importance of a jth index in the l time training.

value larger, show that the effect of this index to Flood Control Dispatch evaluate alternatives result is more remarkable, the status in index system to be screened is also more for important.

Step 5, index sensitivity analysis:

BP neural network after training has possessed stronger mapping ability, i.e. a given input vector, and network can approach desired output accurately.The present invention adopts the BP neural network after step 3 is trained and checked to carry out sensitivity analysis to each index, the Relative Contribution rate that the change quantitatively calculating each index changes final appraisal results.

Input vector X=(x using each desired value of a certain sample as BP neural network _1,x _2,, x _j..., x _n) ^t, its network real output value is designated as for a jth desired value x _j, under the changeless condition of other index of maintenance, get x _j ^*=x _j(1+ Δ x _j), Δ x _jwhen=20%, the absolute change amount of network output valve z is calculate its relative variation:

${z_j}^{x\mathrm{var}}=\frac{{z_j}^{\mathrm{var}}}{\stackrel{\&OverBar;}{z}}$

The Relative Contribution rate G that the change of a further calculating jth index changes network output valve z _j:

$G_j=\frac{{z_j}^{x\mathrm{var}}}{\underset{j=1}{\overset{n}{\&Sigma;}}{z_j}^{x\mathrm{var}}}\&times;100\%$

Relative Contribution rate quantitatively have evaluated the change of each index to the size of scheme evaluation result influence degree, and its value is larger, shows that evaluation result is more responsive to this index.

Step 6, the criterion of agriculture products screening and threshold value:

The relative Link Importance of trying to achieve in order to integrating step 4,5 effectively and Relative Contribution rate information carry out index screening, invention defines the overall target F that simultaneously can take into account both sizes _j:

$F_j=\left|\stackrel{\&OverBar;}{R_j}\right|\&times;G_j$

In formula, and G _jbe respectively relative Link Importance and the Relative Contribution rate of a jth index. and G _jvalue be positioned in [0,1] interval, adopt product calculation definition F _jthere is good discrimination.

Understand intuitively, when and G _jwhen getting higher value simultaneously, F _jlarger; When and G _jmiddle one is comparatively large and another one is less time, F _jvalue is moderate; When and G _jwhen getting smaller value simultaneously, F _jless.According to the F of each index _jvalue size, the priority that agriculture products is deleted, obviously, F _jless index is more first considered to delete.As the F of certain index _jvalue is than all index F _javerage time more than a little magnitude, think that the relative Link Importance of this index and Relative Contribution rate are very little compared with the average level of all indexs, this index can be deleted.In order to this process of quantitative description, the present invention further defines comprehensive distinguishing index P _j:

$P_j=\lg \frac{F_j}{\stackrel{\&OverBar;}{F_j}}$

In formula, for all index F _jmean value.

Comprehensive distinguishing index P _jreflect F _jvalue with difference in magnitude.If P _j> 0, shows the F of this index _jvalue is greater than average level; If P _j< 0, shows the F of this index _jvalue is less than average level; Especially, P is worked as _jwhen≤-1, show the F of this index _jvalue with differ a more than magnitude, be markedly inferior to average level, can be deleted.Therefore, the present invention is by P _j≤-1 as the threshold value of index screening, by the P of more each index _jvalue size carries out index screening, thus changes index screening into quantitative computation process from traditional subjective analysis deterministic process.

More than describe the preferred embodiment of the present invention in detail; but the present invention is not limited to the detail in above-mentioned embodiment, within the scope of technical conceive of the present invention; can carry out multiple equivalents to technical scheme of the present invention, these equivalents all belong to protection scope of the present invention.

Claims (10)

1., based on a method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, comprise the steps:

Step 1. generating network training sample;

Step 2. determines the topological structure of BP neural network;

Step 3. is trained, check described BP neural network;

The significance level of each index in step 4. identification BP neural network;

Step 5. calculates the change of each desired value to the size of Influence on test result degree, analyzes the susceptibility of each desired value;

The criterion of step 6. agriculture products screening and threshold value.

2., as claimed in claim 1 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 1 is further:

Using desired value to be screened for each Flood Control Dispatch scheme as input, will finally be used for passing judgment on the Comprehensive Evaluation desired value of scheme quality as output; Wherein, described Comprehensive Evaluation desired value adopts and tries to achieve based on the multiple attributive decision making method of positive and negative ideal point concept;

Adopt the method generating network training sample of desired value uniform discrete.

3., as claimed in claim 2 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 1 is further comprising the steps:

Step 11., according to reservoir Technique for Real-time Joint Operation of Flood model, generates the Flood Control Dispatch scheme collection meeting constraint condition:

{x _ij|i＝1,2,…,m；j＝1,2,…,n}，

Wherein i and j is respectively scheme sequence number and index sequence number, m and n is respectively scheme number and index number to be screened;

Step 12. determines positive ideal solution scheme { x _j ⁺| j=1,2 ..., n} and minus ideal result scheme { x _j ^-| j=1,2 ..., n};

${x_j}^{+}=\left\{\begin{array}{c}\underset{1\&le;i\&le;m}{\max }\left\{x_{ij}\right\}\\ \underset{1\&le;i\&le;m}{\min }\left\{x_{ij}\right\}\end{array}\right.,{x_j}^{-}=\left\{\begin{array}{c}\underset{1\&le;i\&le;m}{\min }\left\{x_{ij}\right\}\\ \underset{1\&le;i\&le;m}{\max }\left\{x_{ij}\right\}\end{array}\right.,$

Step 13. adopts the method for stochastic simulation to produce the input item of training sample:

Obey equally distributed random number u for stochastic generation k _g~ U (0,1), g=1,2,, k, carries out sliding-model control to each desired value between positive and negative ideal scheme, the input item of raw k the training sample of common property, k is natural number, and its numerical value is determined according to the network training accuracy requirement of reality, and discretize formula is as follows:

x _gj＝x _j ⁺-(x _j ⁺-x _j ^-)×u _g

Positive and negative ideal scheme is also taken as two input items of training sample, then the final training sample input item generated is designated as { x _tj| t=1,2 ..., p; J=1,2 ..., n}, wherein t is sample sequence number, and p is sample size, p=k+2;

Step 14. calculates the Euclidean distance of each sample apart from positive ideal scheme and minus ideal result scheme:

$\left\{\begin{array}{c}{d_t}^{+}=\sqrt{\underset{j=1}{\overset{n}{\&Sigma;}}{({x_j}^{+}-x_{tj})}^2}\\ {d_t}^{-}=\sqrt{\underset{j=1}{\overset{n}{\&Sigma;}}{({x_j}^{-}-x_{tj})}^2}\end{array}\right.,$

In formula, d _t ⁺be the Euclidean distance of t sample apart from positive ideal scheme; d _t ^-be the Euclidean distance of t sample apart from ill ideal solution;

Step 15. calculates exchange premium degree coefficient c _tspan be [0,1];

Step 16. couple training sample { x _tj| t=1,2 ..., p; J=1,2 ..., the input item of n} is normalized:

${x_{tj}}^{*}=\frac{x_{tj}-x_{j\min }}{x_{j\max }-x_{j\min }},$

In formula, x _tjand x _tj ^*be respectively calculated value and the normalized value of a jth index in t sample; x _jmaxand x _jminbe respectively maximal value and the minimum value of a jth index;

{ the x that determines that training sample is expressed as _tj ^*, c _t| t=1,2 ..., p; J=1,2 ..., n}.

4., as claimed in claim 3 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 2 is further:

Described BP neural network comprises input layer, single hidden layer and output layer;

Wherein, input layer number is the number of index to be screened, and output layer nodes is 1, and the number of hidden nodes is determined according to Lippmann experimental formula; The transport function of input layer number is linear function, and the transport function of hidden layer and output layer adopts tanh Sigmoid function;

Connect weights and threshold according to Delta learning rules to the network between each node layer to adjust, computational grid global error E:

$E=\underset{t=1}{\overset{p}{\&Sigma;}}{E}_t=\underset{t=1}{\overset{p}{\&Sigma;}}\frac{{(z_t-c_t)}^2}{2},$

E _tit is single network of samples training error of t sample; z _tand c _tbe respectively network real output value and the desired output of t sample.

5., as claimed in claim 4 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 3 is further:

Described training sample is divided into two parts, is respectively used to network training and network checks;

The main training parameter of setting BP neural network, comprises learning rate and factor of momentum;

Initial network connects weights and threshold at interval (-1,1) interior stochastic generation;

The end condition of setting network training;

Delta learning rules are adopted to train described BP neural network.

6., as claimed in claim 5 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 4 is further:

The relative Link Importance of definition index, the relative importance of a jth index is designated as

$\stackrel{\&OverBar;}{R_j}=\frac{1}{N}\underset{l=1}{\overset{N}{\&Sigma;}}{R_j}^l,$

In formula, N is the number of times of network training; R _j ^lit is the relative Link Importance of a jth index in the l time training; The relative Link Importance R of a jth index _jfor:

$R_j=\underset{s=1}{\overset{h}{\&Sigma;}}\left(w_{js}{w}_s\right)/\underset{j=1}{\overset{n}{\&Sigma;}}\underset{s=1}{\overset{h}{\&Sigma;}}\left|w_{js}{w}_s\right|,$

In formula, w _jsfor the network between input layer j and hidden node s connects weights; w _sfor the network between hidden node s and output layer node connects weights; N is input layer number; H is the number of hidden nodes.

7., as claimed in claim 6 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 5 is further:

Calculate the change of each index to the Relative Contribution rate of net result change, the change of a jth index is designated as G to the Relative Contribution rate that network output valve z changes _j,

$G_j=\frac{{z_j}^{x\mathrm{var}}}{\underset{j=1}{\overset{n}{\&Sigma;}}{z_j}^{x\mathrm{var}}}\&times;100,{z_j}^{x\mathrm{var}}=\frac{{z_j}^{\mathrm{var}}}{\stackrel{\&OverBar;}{z}}$

In formula, for network real output value, for the absolute change amount of network output valve z.

8., as claimed in claim 7 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, described step 6 is further:

Calculate comprehensive distinguishing index P _j,

$P_j=\lg \frac{F_j}{\stackrel{\&OverBar;}{F_j}},F_j=\left|\stackrel{\&OverBar;}{R_j}\right|\&times;G_j,$

In formula, for all index F _jmean value, and G _jbe respectively relative Link Importance and the Relative Contribution rate of a jth index, and G _jvalue be positioned in [0,1] interval.

9., as claimed in claim 8 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, if P _j> 0, the then F of this index _jvalue is greater than average level; If P _j< 0, the then F of this index _jvalue is less than average level.

10., as claimed in claim 8 or 9 based on the method for BP neural network screening reservoir regulation for flood control scheme selection indicators, it is characterized in that, work as P _jwhen≤-1, then the F of this index _jvalue with differ a more than magnitude.