CN103425743A - Steam pipe network prediction system based on Bayesian neural network algorithm - Google Patents

Abstract

The invention discloses a steam pipe network prediction system based on a Bayesian neural network algorithm. The steam pipe network prediction system based on the Bayesian neural network algorithm mainly comprises a relational data base, a data collection module, a data display module and a Bayesian neural network prediction module. The data display module is arranged on an engineer station, the Bayesian neural network prediction module is arranged on an application server, the relational data base is arranged on a relational data base server, and the data collection module is arranged on a real-time data base. The relational data base is a data communication medium between the data display module and the Bayesian neural network prediction module, the Bayesian neural network prediction module writes results of the Bayesian neural network prediction module into the relational data base, and the data display module reads and displays the results from the relational data base. The steam pipe network prediction system based on the Bayesian neural network algorithm has the advantages of achieving rapid solution, ensuring precision of a calculation model and being capable of meeting requirements of a process technology, and the calculation results are in fit with actual operation conditions.

Description

Steam pipe system prognoses system based on the Bayesian neural network algorithm

Technical field

The present invention relates to the health-care massage device field, specifically, specially refer to a kind of Bayesian neural network algorithm and the steam pipe system prognoses system based on this algorithm.

Background technology

Steam is one of iron and steel enterprise's important energy source medium.In large-scale associating iron and steel enterprise, vapour system is the complex object with characteristics such as large dead time, large inertia, non-linear, Multivariable Coupling, variable elements, be in particular in that polydispersion user, multi-steam source, multiple pressure power grade (in, low pressure), multi-state change (season, production decision), the operational administrative personnel lack the Accurate Prediction of pipe network operation work information and effectively control.In the face of so complicated operation conditions, managerial personnel's great majority or adopt the management and running modes of " in response to formula ", rely on and produce for many years the experience command system operation accumulated, and emptying occurs often, the situation such as the use that degrades, and causes great waste.Must cause like this lacking effective pipe network operation PREDICTIVE CONTROL means.

Bayesian neural network is a kind of data digging method that Bayes principle is combined with neural network, it is incorporated into Bayes principle in the weights study of neural network, the posterior probability of weights of take is the optimization aim function, ask for the weights of neural network by the posterior probability that maximizes weights, thereby set up complete neural network model.Neural network based on bayes method is widely used in the nonlinear system model Research on Identification, and has obtained good actual effect.

Summary of the invention

The present invention adopts the artificial neuron algorithm based on Bayesian network to carry out forecasting research to the steam pipe system operating condition, predicts the outcome and can be used for the pipe network scheduling controlling, for optimizing pipe network operation, provides decision support.Carry out neural network training based on the Bayesian network algorithm, can improve the generalization ability of network.For avoiding the training of crossing of network, introduced the penalty term that means the complicated network structure in the objective function of network, so that can be by reduce the complicacy of network structure in the training optimizing process.

Technical matters solved by the invention can realize by the following technical solutions:

Steam pipe system prognoses system based on the Bayesian neural network algorithm is characterized in that: comprise real-time data base server, application server, relational database server and engineer station;

Described real-time data base server has data acquisition module and real-time data base, and data acquisition module is for the collection site data and it is stored in to real-time data base, and real-time data base is for providing field data to the Bayesian neural network prediction module;

Described application server has the Bayesian neural network prediction module, application server is connected with the real-time data base server, described Bayesian neural network prediction module adopts the Bayesian neural network algorithm, for calculate operating condition model and the data of steam pipe system according to field data;

Described relational database server has relational database, and the port of described relational database is connected with the engineer station with application server respectively, the operating condition model and the data that for storing the Bayesian neural network prediction module, calculate;

Described engineer station has data disaply moudle, and this data disaply moudle is for reading and show operating condition model and the data that relational database is stored; Described data disaply moudle has the data-interface part, for to the Bayesian neural network prediction module, inputting data.

A kind of algorithm of the Bayesian neural network for the steam pipe system prognoses system, it is characterized in that: described algorithm flow is as follows:

1) calculate the error function J of Bayesian neural network _D:

$J_D=\frac{1}{2}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{(d_{\mathrm{nk}}-o_{\mathrm{nk}})}^2$

In formula, d _NkFor the actual output of network, o _NkFor desired output, k is number of samples, the output quantity number that n is neural network;

2) calculate power attenuation term J _W:

$J_W=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}\underset{i=1}{\overset{w}{\mathrm{\Σ}}}{w}_i^2$

W is weight vector, and w is the weights number, is w _iBe i weights;

3) by the total error function definition be F (w):

F(W)=αJ _w+βJ _D

In formula, α, β is super parameter, for controlling the distribution form of weights and threshold value; The size of super parameter is determining the training objective of Bayesian neural network, if α<<β, lay particular emphasis on and reduce training error, but the possibility over-fitting; If α>>β, lay particular emphasis on restriction weights scale, but possible error is larger;

4) calculate super parameter alpha, the posterior probability of β distributes:

$p(\mathrm{\&alpha;},\mathrm{\&beta;}|D,H)=\frac{p\left(D\right|\mathrm{\&alpha;},\mathrm{\&beta;},H)p(\mathrm{\&alpha;},\mathrm{\&beta;}|H)}{p\left(D\right|H)}$

In formula, and p (α, β | H) be super parameter alpha, the prior probability of β, p (D|H) is normalized factor, and p (D| α, β, H) is likelihood function, and D is the training sample sum, and H is network parameter;

5) respectively to super parameter alpha, β asks local derviation, the value of super parameter when going out to have maximum significance;

${\mathrm{\&alpha;}}_{\mathrm{MP}}=\frac{\mathrm{\&gamma;}}{{2J}_w\left(w_{\mathrm{MP}}\right)},$ ${\mathrm{\&beta;}}_{\mathrm{MP}}=\frac{N-\mathrm{\&gamma;}}{2J_D\left(w_{\mathrm{MP}}\right)};$

In formula, γ can reduce the number of parameters of performance index function in neural network, and its span is 0～w; w _MPFor F (W) corresponding weights and sets of threshold values while getting minimum value;

6) judge whether F (W) value restrains, if do not restrain, does not carry out new round iterative, until optimum solution occurs, obtains having the Bayesian neural network model of maximum significance.

Compared with prior art, beneficial effect of the present invention is as follows:

The Bayesian neural network model is applied in the forecast of steam pipe system operating pressure, the model prediction result shows the speed of convergence at network, all be better than traditional BP neural network in precision of prediction and accuracy, realize steam pipe system prediction and calculation fast and accurately, more contributed to the operation and management to steam pipe system.

The accompanying drawing explanation

The hardware schematic diagram that Fig. 1 is steam pipe system prognoses system of the present invention.

The process flow diagram that Fig. 2 is Bayesian neural network algorithm of the present invention.

Embodiment

For technological means, creation characteristic that the present invention is realized, reach purpose and effect is easy to understand, below in conjunction with embodiment, further set forth the present invention.

Referring to Fig. 1, the steam pipe system prognoses system based on the Bayesian neural network algorithm of the present invention, mainly comprise real-time data base server, application server, relational database server and engineer station.

Wherein, relational database server is connected with application server with the engineer station, and application server, except with relational database server is connected, also is connected with the engineer station with real-time data base, keeps exchanges data between the three.Application module comprises relational database, data acquisition module, data result display module, Bayesian neural network prediction module.The data result display module is deployed in the engineer station, and the Bayesian neural network prediction module is deployed in application server, and relational database is deployed in relational database server, and data acquisition module is deployed in real-time data base.

Relational database is the data communication medium between display module and predictor computation module.Prediction module writes relational database by result of calculation, and display module is read and shown from relational database.

The real-time data base server has data acquisition module and real-time data base, and data acquisition module is for the collection site data and it is stored in to real-time data base, and real-time data base is for providing field data to the Bayesian neural network prediction module; The Bayesian neural network prediction module adopts the Bayesian neural network algorithm, for calculate operating condition model and the data of steam pipe system according to field data.The port of relational database is connected with the engineer station with application server respectively, the operating condition model and the data that for storing the Bayesian neural network prediction module, calculate.Data disaply moudle is for reading and show operating condition model and the data that relational database is stored; Described data disaply moudle has the data-interface part, for to the Bayesian neural network prediction module, inputting data.

Referring to Fig. 2; The specific algorithm of Bayesian neural network prediction module is as follows:

Generally, the performance function of neural network adopts the square error function.Note neural network model training sample D is the training sample sum, and W is network weight (threshold value) vector.Under the condition of given network structure H and weight vector W, the error function J of network _DGet the quadratic sum of error:

$J_D=\frac{1}{2}\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{(d_{\mathrm{nk}}-o_{\mathrm{nk}})}^2---\left(1\right)$

D in above formula _NkFor the actual output of network, n is number of samples, the output quantity number that k is neural network.

Regularization method commonly used is to add power attenuation term J after error function _W:

$J_W=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{w}_i^2---\left(2\right)$

So in formula, w is the network weight sum. the total error function may be defined as:

F(W)=αJ _W+βJ _D (3)

α in formula, β is super parameter, is controlling the distribution form of other parameters (power and threshold value).The size of super parameter is determining the training objective of neural network, if α<<β, lay particular emphasis on and reduce training error, but the possibility over-fitting; If α>>β, lay particular emphasis on limiting network weights scale, but possible error is larger.In actual applications, need compromise to consider, the minimization objective function is when reducing the network training error, reduces the complicacy of network structure.For regularization method, difficult point is determining of super parameter, and this paper adopts bayes method to realize the selection of super parameter.

Known according to Bayes's derivation [4,5], super parameter alpha, the posterior probability of β is distributed as:

$p(\mathrm{\&alpha;},\mathrm{\&beta;}|D,H)=\frac{p\left(D\right|\mathrm{\&alpha;},\mathrm{\&beta;},H)p(\mathrm{\&alpha;},\mathrm{\&beta;}|H)}{p\left(D\right|H)}---\left(4\right)$

In formula, and p (α, β | H) be the prior probability of super parameter, p (D|H) is normalized factor, and P (D| α, β, H) is likelihood function.While solving, to α, β asks respectively local derviation, the value of super parameter when going out to have maximum significance:

${\mathrm{\&alpha;}}_{\mathrm{MP}}=\frac{\mathrm{\&gamma;}}{{2J}_w\left(w_{\mathrm{MP}}\right)}---\left(5\right)$

${\mathrm{\&beta;}}_{\mathrm{MP}}=\frac{N-\mathrm{\&gamma;}}{2J_D\left(w_{\mathrm{MP}}\right)}$

In formula, γ can reduce the number of parameters of performance index function in network, and its span is 0～W.W _MPCorresponding weights and sets of threshold values while for J, getting minimum value.After obtaining super parameter alpha with maximum significance, β, again judge whether the J value restrains, do not restrain and carry out new round iterative, until optimum solution occurs, obtain having the neural network model of maximum significance.

Pipe networks prediction model calculation procedure is as follows:

Step 1: with the less super parameter of first value initialization and network connection weight;

Step 2; Adopt the Levenberg-Marquardt optimized algorithm once to train network, make target function type (3) minimum;

Step 3: calculating parameter γ _g, verification msg predicated error and significance;

Step 4: according to formula (5), (6) adjust super parameter; Repeating step two is to step 4 until convergence.

Above demonstration and described ultimate principle of the present invention and principal character and advantage of the present invention.The technician of the industry should understand; the present invention is not restricted to the described embodiments; that in above-described embodiment and instructions, describes just illustrates principle of the present invention; without departing from the spirit and scope of the present invention; the present invention also has various changes and modifications, and these changes and improvements all fall in the claimed scope of the invention.The claimed scope of the present invention is defined by appending claims and equivalent thereof.

Claims (2)

1. the steam pipe system prognoses system based on the Bayesian neural network algorithm, is characterized in that: comprise real-time data base server, application server, relational database server and engineer station;

Described real-time data base server has data acquisition module and real-time data base, and data acquisition module is for the collection site data and it is stored in to real-time data base, and real-time data base is for providing field data to the Bayesian neural network prediction module;

Described application server has the Bayesian neural network prediction module, application server is connected with the real-time data base server, described Bayesian neural network prediction module adopts the Bayesian neural network algorithm, for calculate operating condition model and the data of steam pipe system according to field data;

Described relational database server has relational database, and the port of described relational database is connected with the engineer station with application server respectively, the operating condition model and the data that for storing the Bayesian neural network prediction module, calculate;

Described engineer station has data disaply moudle, and this data disaply moudle is for reading and show operating condition model and the data that relational database is stored; Described data disaply moudle has the data-interface part, for to the Bayesian neural network prediction module, inputting data.

2. the algorithm of the Bayesian neural network for the steam pipe system prognoses system, it is characterized in that: described algorithm flow is as follows:

1) calculate the error function J of Bayesian neural network _D:

$J_D=\frac{1}{2}\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{(d_{\mathrm{nk}}-o_{\mathrm{nk}})}^2$

In formula, d _NkFor the actual output of network, o _NkFor desired output, k is number of samples, the output quantity number that n is neural network;

2) calculate power attenuation term J _W:

$J_W=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{w}_i^2$

W is weight vector, and w is the weights number, w _iBe i weights;

3) by the total error function definition be F (w):

F(W)=αJ _W+βJ _D

In formula, α, β is super parameter, for controlling the distribution form of weights and threshold value; The size of super parameter is determining the training objective of Bayesian neural network, if α<<β, lay particular emphasis on and reduce training error, but the possibility over-fitting; If α>>β, lay particular emphasis on restriction weights scale, but possible error is larger;

4) calculate super parameter alpha, the posterior probability of β distributes:

$p(\mathrm{\&alpha;},\mathrm{\&beta;}|D,H)=\frac{p\left(D\right|\mathrm{\&alpha;},\mathrm{\&beta;},H)p(\mathrm{\&alpha;},\mathrm{\&beta;}|H)}{p\left(D\right|H)}$

In formula, and p (α, β | H) be super parameter alpha, the prior probability of β, p (D|H) is normalized factor, and p (D| α, β, H) is likelihood function, and D is the training sample sum, and H is network parameter;

5) respectively to super parameter alpha, β asks local derviation, the value of super parameter when going out to have maximum significance;

${\mathrm{\&alpha;}}_{\mathrm{MP}}=\frac{\mathrm{\&gamma;}}{{2J}_w\left(w_{\mathrm{MP}}\right)},$ ${\mathrm{\&beta;}}_{\mathrm{MP}}=\frac{N-\mathrm{\&gamma;}}{2J_D\left(w_{\mathrm{MP}}\right)};$

In formula, γ can reduce the number of parameters of performance index function in neural network, and its span is 0～w; w _MPFor F (W) corresponding weights and sets of threshold values while getting minimum value;

6) judge whether F (W) value restrains, if do not restrain, does not carry out new round iterative, until optimum solution occurs, obtains having the Bayesian neural network model of maximum significance.

Images