CN106156487A - A kind of mining under reservoir water level computational methods considering the jacking impact of many rivers - Google Patents

Abstract

The present invention provides kind of the mining under reservoir water level computational methods considering the jacking impact of many rivers, comprises the following steps: tributary flow and the reservoir storage outflow of downstream jacking is exported as model as mode input predictor, mining under reservoir water level, calculates the level of tail water；Consider the tentative calculation of early stage impact, find the result that error is minimum；Parameter tentative calculation: arrange different hidden neuron number, maximum iteration time, training speed and training precision numerical value, carries out tentative calculation repeatedly, the result that error identifying is minimum；Data export: show that the level of tail water number list of optimum, training period are received assorted efficiency factor, probative term and receive assorted efficiency factor, probative term maximum error, probative term mean error.The inventive method can reflect that when the non-monotonic complex relationship of tributary flow and jacking water level, especially jacking amount are big, calculating error is less, can be used for producing reality.

Description

A kind of mining under reservoir water level computational methods considering the jacking impact of many rivers

Technical field

The present invention relates to prediction management field, Reservoir region, a kind of mining under reservoir considering the jacking impact of many rivers Water level computational methods, especially can be used for jacking impact irregular in the case of.

Background technology

At present, mining under reservoir water level-storage outflow has obvious regularity under native state, but when downstream has relatively During big branch afflux, jacking can be had to affect mining under reservoir, make section stage discharge relation more sophisticated, add mining under reservoir The forecast difficulty of water level.Traditional method is to analyze tributary, the downstream water jacking amount to the level of tail water to estimate the level of tail water, with Lower abbreviation " jacking method ", the precision of " jacking method " is the highest, and especially jacking amount is big when, error is bigger.Must in consideration of it, have Find a kind of mining under reservoir water level computational methods considering the jacking impact of many rivers.

Summary of the invention

The technical problem to be solved is to provide a kind of mining under reservoir water-level gauge considering the jacking impact of many rivers Calculation method, it is possible to solve the deficiency of " jacking method ", the level of tail water can be calculated fast and accurately.

For solving above-mentioned technical problem, the technical solution adopted in the present invention is: a kind of consider the impact of many rivers jackings Mining under reservoir water level computational methods, comprise the following steps:

One, using the tributary flow of downstream jacking and reservoir storage outflow as mode input predictor, mining under reservoir water Position exports as model, utilizes the neural computing level of tail water of error backpropagation algorithm；

Two, the tentative calculation of early stage impact is considered: choose all factors of front a period as mode input predictor, together Time choose the tributary flow of downstream jacking of current time and reservoir storage outflow as mode input predictor, carry out repeatedly Tentative calculation, finds the result that error is minimum；

Three, parameter tentative calculation: different hidden neuron number, maximum iteration time, training speed and training precision is set Numerical value, carries out tentative calculation repeatedly, the result that error identifying is minimum；

Four, data output: show that the level of tail water number list of optimum, training period receive assorted efficiency factor, probative term Receive assorted efficiency factor, probative term maximum error, probative term mean error.

Preferably in scheme, tentative calculation process is:

1) forward-propagating input information, in forward-propagating input information process, the data of input are passed through to swash from input layer Encourage function successively to propagate to hidden layer, output layer；The internodal connection of each two all represents one for by this connection signal Weighted value, referred to as weight；

2) reversely feedforward calculates error, if the output valve of output layer is relatively big with differing of desired value, result of calculation does not reaches To expection, then error is pushed away to input layer is the most counter by output layer, adjust connection weight between each layer simultaneously, re-start forward Computing.By the continuous adjustment of weight, network is made to develop towards the direction of energy desired output, when output result reaches desired pre- When surveying precision, network calculations stops.

Preferably in scheme, concrete operation process is: set, and in 3 layers of error backpropagation algorithm network, input information is

X=(x₁, x₂..., x_i..., x_n)^T, i=1,2 ..., n,

Wherein x_iFor inputting the i-th element of information；

Output layer output information is O=(o₁, o₂..., o_k..., o_l)^T, k=1,2 ..., l,

Wherein o_kKth element for output layer output information；

Desired output information is d=(d₁, d₂..., d_k..., d_l)^T, k=1,2 ..., l,

Wherein d_kKth element for desired output information；

Weight square V between input layer and hidden layer represents, V=(V₁, V₂..., V_j..., V_m)^T,

J=1,2 ..., m,

Wherein column vector V_jFor the weight vector that hidden layer jth row neuron is corresponding；

Weight matrix W between hidden layer and output layer represents,

W=(W₁, W₂..., W_k... W_l)^T, k=1,2 ..., l,

Wherein column vector W_kFor the weight vector that output layer kth row neuron is corresponding；

When exporting result and being inconsistent with expectation, there is calculating error E:

$E=\frac{1}{2}{(d-O)}^2---\left(1\right)$

After expansion:

$E=\frac{1}{2}\underset{k=1}{\overset{l}{\mathrm{\Σ}}}{\{d_k-f\[\underset{j=0}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_{jk}f\left(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_{ij}{x}_i\right)\]\}}^2---\left(2\right)$

In formula, f is each layer neuron excitation function；

ω_jkFor the weight between hidden layer neuron and output layer neuron；

v_ijFor the weight between input layer and hidden layer neuron；

Adjusting weights purpose is to make output result constantly approach expected results, uses error gradient descent algorithm to adjust power Weight:

${\mathrm{\Δ\ω}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\ω}}_{jk}}---\left(3\right)$

${\mathrm{\Δ\υ}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\υ}}_{jk}}$

△ ω in formula_jkFor the weight adjustment amount between hidden layer neuron and output layer neuron；

△v_ikFor the weight adjustment amount between input layer and hidden layer neuron；

η ∈ (0,1) is proportional control factor.

In step 3, setting hidden neuron number as 10, maximum iteration time is 10000 times, and training speed is 0.1, When training precision is 0.0000004, thus setup algorithm error is minimum.

A kind of mining under reservoir water level computational methods considering the jacking impact of many rivers that the present invention provides, more than using Step, have the advantage that

1, " jacking method " is the most rough estimation, and it is monotonically increasing that its result reflects jacking water level with tributary flow Rule, is not exclusively consistent with practical situation, and error is relatively big, and the inventive method can reflect tributary flow and jacking water level Calculate error when non-monotonic complex relationship, especially jacking amount are big less, can be used for producing reality.

2, in jacking method, artificial amount of calculation is relatively big, needs to make substantial amounts of data analysis in advance, and different people calculating can Different results can be caused, and the result precision that the inventive method obtains is high, concordance is preferable.

3) the inventive method is easy to use, favorable expandability.The result of the readily available global optimum of calculating process of the present invention. Frequency of training is less, and the learning efficiency is high, and convergence rate is very fast.

The application of the inventive method, can be effectively improved hydrologic forecast automatization, become more meticulous, level of standardization.

Accompanying drawing explanation

The invention will be further described with embodiment below in conjunction with the accompanying drawings:

Fig. 1 is the neutral net schematic diagram of the error backpropagation algorithm that the present invention uses.

Fig. 2 is the schematic flow sheet of inventive algorithm.

Fig. 3 is the calculating error contrast schematic diagram of inventive algorithm and " jacking method ".

Detailed description of the invention

Below by embodiment, in conjunction with chart, technical scheme is further elaborated with.

Certain hydrometric station is positioned at some hydropower station downstream, has certain tributary 1 to import at the 2km of survey station downstream, has certain tributary 2 to converge at 30km Enter, there is backwater effect in certain tributary 1 to hydrometric station when high water and certain tributary 2 high water, utilize error backpropagation algorithm to calculate water Literary composition station water level.

Step 1, data input: using pre-as mode input to certain tributary 1 flow, certain tributary 2 flow, reservoir storage outflow Surveying the factor, hydrometric station water level exports as model, utilizes error backpropagation algorithm to calculate hydrometric station water level.

Step 2, it is considered to the tentative calculation of early stage impact: consider early stage tributary 1 flow, tributary 2 flow, reservoir storage outflow and water The hydrometric station water level of present period is had a certain impact by literary composition station water level, through tentative calculation repeatedly, choose front 3 periods all because of Son is minimum as calculating error during mode input predictor, chooses early stage tributary 1 flow of current time, tributary 2 is flowed simultaneously Amount, reservoir storage outflow are as mode input predictor.

Step 3, parameter tentative calculation: through tentative calculation repeatedly, set hidden neuron number as 10, maximum iteration time is 10000 times, training speed is 0.1, when training precision is 0.0000004, calculates error minimum.Error backpropagation algorithm is neural Network training speed is the biggest, and weight changes the most greatly, restrains the fastest；But training speed is excessive, can cause the vibration of system, therefore, Training speed, under being not resulted in vibration premise, is the bigger the better.This algorithm picks suitably trains speed to overcome BP neutral net Convergence is slow, the shortcoming of training time length.

1) forward-propagating input information.In forward-propagating input information process, the data of input are passed through to swash from input layer Encourage function successively to propagate to hidden layer, output layer.The internodal connection of each two all represents one for by this connection signal Weighted value, referred to as weight, this is equivalent to the memory of artificial neural network.

2) reversely feedforward calculates error.If the output valve of output layer is relatively big with differing of desired value, result of calculation does not reaches To expection, then error is pushed away to input layer is the most counter by output layer, adjust connection weight between each layer simultaneously, re-start forward Computing.By the continuous adjustment of weight, network is made to develop towards the direction of energy desired output, when output result reaches desired pre- When surveying precision, network calculations stops.

It is assumed that in 3 layers of BP network, input information is X=(x₁, x₂..., x_i..., x_n)^T, i=1,2 ..., n, wherein x_i For inputting the i-th element of information；Output layer output information is O=(o₁, o₂..., o_k..., o_l)^T, k=1,2 ..., l, wherein o_kKth element for output layer output information；Desired output information is d=(d₁, d₂..., d_k..., d_l)^T, k=1,2 ..., L, wherein d_kKth element for desired output information；Weight square V between input layer and hidden layer represents, V=(V₁, V₂..., V_j..., V_m)^T, j=1,2 ..., m, wherein column vector V_jFor the weight vector that hidden layer jth row neuron is corresponding；Hidden layer With the weight matrix W between output layer represents, W=(W₁, W₂..., W_k... W_l)^T, k=1,2 ..., l, wherein column vector W_kFor The weight vector that output layer kth row neuron is corresponding.

When exporting result and being inconsistent with expectation, there is calculating error E:

$E=\frac{1}{2}{(d-O)}^2---\left(1\right)$

After expansion:

$E=\frac{1}{2}\underset{k=1}{\overset{l}{\mathrm{\Σ}}}{\{d_k-f\[\underset{j=0}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_{jk}f\left(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_{ij}{x}_i\right)\]\}}^2---\left(2\right)$

In formula, f is each layer neuron excitation function；ω_jkFor the weight between hidden layer neuron and output layer neuron； v_ijFor the weight between input layer and hidden layer neuron.

Adjusting weights purpose is to make output result constantly approach expected results, and general employing error gradient descent algorithm adjusts Weight:

${\mathrm{\Δ\ω}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\ω}}_{jk}}---\left(3\right)$

${\mathrm{\Δ\υ}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\υ}}_{jk}}$

△ ω in formula_jkFor the weight adjustment amount between hidden layer neuron and output layer neuron；△v_ikFor input layer god Weight adjustment amount between unit and hidden layer neuron；η ∈ (0,1) is proportional control factor.

Step 4, data export: show that the level of tail water number list of optimum, training period receive assorted efficiency factor, inspection Phase receives assorted efficiency factor, probative term maximum error and probative term mean error.

Comparison of computational results such as following table:

It is 99.86% that the training period of error backpropagation algorithm receives assorted efficiency factor, and probative term is received assorted efficiency factor and is 99.82%, probative term maximum error 0.32m, mean error 0.10m；Jacking method maximum error 0.74m, mean error 0.33m； Jacking amount is 0.11m more than 1m time error back-propagation algorithm mean error, jacking method mean error 0.40m, inventive algorithm Error less.

The above embodiments are only the preferred technical solution of the present invention, and are not construed as the restriction for the present invention, this Shen Embodiment in please and the feature in embodiment, can mutual combination in any in the case of not conflicting.The protection model of the present invention Enclose the technical scheme should recorded with claim, including the equivalent of technical characteristic in the technical scheme that claim is recorded Scheme is protection domain.Equivalent the most in this range is improved, also within protection scope of the present invention.

Claims (4)

1. consider mining under reservoir water level computational methods for many rivers jacking impact, it is characterized in that comprising the following steps:

One, tributary flow and the reservoir storage outflow of downstream jacking are made as mode input predictor, mining under reservoir water level Export for model, utilize the neural computing level of tail water of error backpropagation algorithm；

Two, the tentative calculation of early stage impact is considered: choose all factors of front n period as mode input predictor, select simultaneously Take the tributary flow of the downstream jacking of current time and reservoir storage outflow as mode input predictor, repeatedly try Calculate, find the result that error is minimum；

Three, parameter tentative calculation: different hidden neuron number, maximum iteration time, training speed and training precision numerical value is set, Carry out tentative calculation repeatedly, the result that error identifying is minimum；

Four, data output: show that the level of tail water number list of optimum, training period are received assorted efficiency factor, probative term and receive assorted Efficiency factor, probative term maximum error, probative term mean error.

A kind of mining under reservoir water level computational methods considering the jacking impact of many rivers the most according to claim 1, its feature It is that tentative calculation process is:

1) forward-propagating input information, forward-propagating input information process in, the data of input from input layer by excitation letter Number is successively propagated to hidden layer, output layer；The internodal connection of each two all represents one for being added by this connection signal Weights, referred to as weight；

2) reversely feedforward calculates error, if the output valve of output layer is relatively big with differing of desired value, result of calculation is the most pre- Phase, then error is pushed away to input layer is the most counter by output layer, adjust connection weight between each layer simultaneously, re-start forward fortune Calculate.By the continuous adjustment of weight, network is made to develop towards the direction of energy desired output, when output result reaches desired prediction During precision, network calculations stops.

3., according to a kind of mining under reservoir water level computational methods considering the jacking impact of many rivers described in claim 2, it is characterized in that Concrete operation process is: set, and in 3 layers of error backpropagation algorithm network, input information is

X=(x₁, x₂..., x_i..., x_n)^T, i=1,2 ..., n,

Wherein x_iFor inputting the i-th element of information；

Output layer output information is O=(o₁, o₂..., o_k..., o_l)^T, k=1,2 ..., l,

Wherein o_kKth element for output layer output information；

Desired output information is d=(d₁, d₂..., d_k..., d_l)^T, k=1,2 ..., l,

Wherein d_kKth element for desired output information；

Weight square V between input layer and hidden layer represents, V=(V₁, V₂..., V_j..., V_m)^T, j=1,2 ..., m,

Wherein column vector V_jFor the weight vector that hidden layer jth row neuron is corresponding；

Weight matrix W between hidden layer and output layer represents,

W=(W₁, W₂..., W_k... W_l)^T, k=1,2 ..., l,

Wherein column vector W_kFor the weight vector that output layer kth row neuron is corresponding；

When exporting result and being inconsistent with expectation, there is calculating error E:

$E=\frac{1}{2}{(d-O)}^2---\left(1\right)$

After expansion:

$E=\frac{1}{2}\underset{k=1}{\overset{l}{\Σ}}{\{d_k-f\[\underset{j=0}{\overset{m}{\Σ}}{\mathrm{\ω}}_{jk}f\left(\underset{i=0}{\overset{n}{\Σ}}{\mathrm{\υ}}_{ij}{x}_i\right)\]\}}^2---\left(2\right)$

In formula, f is each layer neuron excitation function；

ω_jkFor the weight between hidden layer neuron and output layer neuron；

v_ijFor the weight between input layer and hidden layer neuron；

Adjusting weights purpose is to make output result constantly approach expected results, employing error gradient descent algorithm adjustment weight:

${\mathrm{\Δ\ω}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\ω}}_{jk}}---\left(3\right)$

${\mathrm{\Δ\υ}}_{jk}=-\mathrm{\η}\frac{\∂E}{\∂{\mathrm{\υ}}_{jk}}$

△ ω in formula_jkFor the weight adjustment amount between hidden layer neuron and output layer neuron；

△v_ikFor the weight adjustment amount between input layer and hidden layer neuron；

η ∈ (0,1) is proportional control factor.

A kind of mining under reservoir water level computational methods considering the jacking impact of many rivers the most according to claim 1, its feature Being: in step 3 to set hidden neuron number as 10, maximum iteration time is 10000 times, and training speed is 0.1, training essence When degree is 0.0000004, thus setup algorithm error is minimum.