"3"^^ . "Kolovratn\u00EDk, Michal" . "Rekurentn\u00ED neuronov\u00E1 s\u00ED\u0165 s modulem zpracov\u00E1n\u00ED dat a s kvadratick\u00FDm a kubick\u00FDm neuronem"@cs . . . . . "P(FR-TI1/538)" . "[D1622E2A4BA0]" . "TptRNN2011" . "2"^^ . . "Rekurentn\u00ED neuronov\u00E1 s\u00ED\u0165 s modulem zpracov\u00E1n\u00ED dat a s kvadratick\u00FDm a kubick\u00FDm neuronem" . . "Ing. Ji\u0159\u00ED Pliska,I & C Energo a.s. ; \u0158editel technick\u00E9ho rozvoje ; Pra\u017Esk\u00E1 684/49, 674 01 T\u0159eb\u00ED\u010D tel.: +420 568 893 111 fax: +420 568 893 999, tel.: +420 568 893 300, email: jpliska@ic-energo.eu" . "226363" . . "Rekurentn\u00ED neuronov\u00E1 s\u00ED\u0165 s modulem zpracov\u00E1n\u00ED dat a s kvadratick\u00FDm a kubick\u00FDm neuronem"@cs . "http://users.fs.cvut.cz/ivo.bukovsky/Research_and_Applications/projects/FRTI1538/indexSW1.htm" . . "Nonconventional dynamic neural network for measurement validation and real-time prediction of NOx emissions of pulverized firing boiler at the coal powder powerplant M\u011Bln\u00EDk 1. As a preprocessing module of input data to the network, the program includes also signal preprocessing modules including a correlation analysis module, mutual information module, input variable reselection and grouping modules, and dimensionality reduction by principal component analysis module for input reconfigurations. The particular design of the programmed network and the signal preprocessing techniques avoids the need for computationally heavy optimization techniques (training), that would not be suitable for real time retraining otherwise. The implemented training algorithm is a modification of the back-propagation through time that is suitable for real time retraining to handle nonstationarity of the burning process as well as individual failures of measured variables. The designed neural network uses nonconventional higher-order neural units (options for linear, quadratic, or cubic neural units). The programmed neural network does not use measured O2 as an input parameter, which was required and it is a unique solution worldwide to our best knowledge. The program is implemented in Matlab, and the resulting codes are Matlab codes and converted executables."@en . . "Rekurentn\u00ED neuronov\u00E1 s\u00ED\u0165 s modulem zpracov\u00E1n\u00ED dat a s kvadratick\u00FDm a kubick\u00FDm neuronem" . "NOx prediction; data validation; neural network"@en . "Program je uva\u017Eov\u00E1n jako sou\u010D\u00E1st komplexn\u00ED dod\u00E1vky \u0159e\u0161en\u00ED v celkov\u00E9 cen\u011B p\u0159ekra\u010Duj\u00EDc\u00ED 1 000 000 K\u010D,- . Odhad d\u00EDl\u010D\u00ED ceny programu je 300 000,- K\u010D. Spoluvlastn\u00EDkem SW je I. & C. Energo a.s., T\u0159eb\u00ED\u010D;" . "RIV/68407700:21220/11:00192337!RIV13-MPO-21220___" . . "Recurrent neural network with data processing modul and quadratic and cubic neural unit"@en . "Bukovsk\u00FD, Ivo" . "Recurrent neural network with data processing modul and quadratic and cubic neural unit"@en . . "Byla naprogramov\u00E1na nekonven\u010D\u00ED dynamick\u00E1 neuronov\u00E1 s\u00ED\u0165 pro validaci a predikci emis\u00ED NOX pr\u00E1\u0161kov\u00E9ho kotle vysok\u00E9ho v\u00FDkonu pro elektr\u00E1rnu M\u011Bln\u00EDk 1. Jako vstupn\u00ED modul je naprogramov\u00E1n korela\u010Dn\u00ED modul vstupn\u00EDch dat, modul v\u00FDpo\u010Dtu vz\u00E1jemn\u00E9 informace, v\u00FDb\u011Brov\u00FD modul vstup\u016F a seskupovac\u00ED modul, a modul pro redukci po\u010Dtu vstup\u016F metodou %22principal component analysis%22 a pro jejich rekonfiguraci. Konkr\u00E9tn\u011B vyvinut\u00FD a naprogramovan\u00FD n\u00E1vrh s\u00EDt\u011B v\u010Detn\u011B p\u0159edzpracov\u00E1n\u00ED dat umo\u017E\u0148uje efektivn\u00ED p\u0159etr\u00E9nov\u00E1v\u00E1n\u00ED v re\u00E1ln\u00E9m \u010Dase bez pot\u0159eby \u010Dasov\u011B n\u00E1ro\u010Dn\u00FDch optimaliza\u010Dn\u00EDch algoritm\u016F kter\u00E9 by nebyli pro re\u00E1ln\u011B \u010Dasovou aplikaci vhodn\u00E9. Naprogramovan\u00FD algoritmus u\u010Den\u00ED je varianta metody %22back-propagation through time%22 kter\u00E9 je pro danou neuronovou s\u00ED\u0165 efektivn\u00ED pro p\u0159etr\u00E9nov\u00E1v\u00E1n\u00ED za \u00FA\u010Delem zvl\u00E1dnut\u00ED nestacionarity re\u00E1ln\u00E9ho syst\u00E9mu a \u010Dast\u00FDm v\u00FDpadk\u016Fm \u00FAdaj\u016F m\u011B\u0159en\u00FDch veli\u010Din. Navr\u017Een\u00E1 s\u00ED\u0165 vyu\u017E\u00EDv\u00E1 nekonven\u010Dn\u00ED neuronov\u00E9 jednotky (QNU a CNU). Pro predikci NOx, s\u00ED\u0165 nepot\u0159ebuje na sv\u00E9m vstupu kysl\u00EDk ve spalin\u00E1ch co\u017E bylo po\u017Eadavkem a jedn\u00E1 se o sv\u011Btov\u011B ojedin\u011Bl\u00E9 \u0159e\u0161en\u00ED podle v\u0161ech na\u0161ich poznatk\u016F. S\u00ED\u0165 byla naprogramov\u00E1na v prost\u0159ed\u00ED Matlab a byly vygenerov\u00E1ny i bin\u00E1rn\u00ED k\u00F3dy pro demo verzi. Program byl testov\u00E1n na re\u00E1ln\u00FDch datech elektr\u00E1rny M\u011Bln\u00EDk 1 s dobr\u00FDmi v\u00FDsledky predikce NOx a p\u0159edpokl\u00E1d\u00E1 se jeho vyu\u017Eit\u00ED jako sou\u010D\u00E1st komplex\u00EDho bal\u00EDku \u0159e\u0161en\u00ED dle po\u017Eadavku z\u00E1kazn\u00EDka I. & C. Energo."@cs . . . . . "Byla naprogramov\u00E1na nekonven\u010D\u00ED dynamick\u00E1 neuronov\u00E1 s\u00ED\u0165 pro validaci a predikci emis\u00ED NOX pr\u00E1\u0161kov\u00E9ho kotle vysok\u00E9ho v\u00FDkonu pro elektr\u00E1rnu M\u011Bln\u00EDk 1. Jako vstupn\u00ED modul je naprogramov\u00E1n korela\u010Dn\u00ED modul vstupn\u00EDch dat, modul v\u00FDpo\u010Dtu vz\u00E1jemn\u00E9 informace, v\u00FDb\u011Brov\u00FD modul vstup\u016F a seskupovac\u00ED modul, a modul pro redukci po\u010Dtu vstup\u016F metodou %22principal component analysis%22 a pro jejich rekonfiguraci. Konkr\u00E9tn\u011B vyvinut\u00FD a naprogramovan\u00FD n\u00E1vrh s\u00EDt\u011B v\u010Detn\u011B p\u0159edzpracov\u00E1n\u00ED dat umo\u017E\u0148uje efektivn\u00ED p\u0159etr\u00E9nov\u00E1v\u00E1n\u00ED v re\u00E1ln\u00E9m \u010Dase bez pot\u0159eby \u010Dasov\u011B n\u00E1ro\u010Dn\u00FDch optimaliza\u010Dn\u00EDch algoritm\u016F kter\u00E9 by nebyli pro re\u00E1ln\u011B \u010Dasovou aplikaci vhodn\u00E9. Naprogramovan\u00FD algoritmus u\u010Den\u00ED je varianta metody %22back-propagation through time%22 kter\u00E9 je pro danou neuronovou s\u00ED\u0165 efektivn\u00ED pro p\u0159etr\u00E9nov\u00E1v\u00E1n\u00ED za \u00FA\u010Delem zvl\u00E1dnut\u00ED nestacionarity re\u00E1ln\u00E9ho syst\u00E9mu a \u010Dast\u00FDm v\u00FDpadk\u016Fm \u00FAdaj\u016F m\u011B\u0159en\u00FDch veli\u010Din. Navr\u017Een\u00E1 s\u00ED\u0165 vyu\u017E\u00EDv\u00E1 nekonven\u010Dn\u00ED neuronov\u00E9 jednotky (QNU a CNU). Pro predikci NOx, s\u00ED\u0165 nepot\u0159ebuje na sv\u00E9m vstupu kysl\u00EDk ve spalin\u00E1ch co\u017E bylo po\u017Eadavkem a jedn\u00E1 se o sv\u011Btov\u011B ojedin\u011Bl\u00E9 \u0159e\u0161en\u00ED podle v\u0161ech na\u0161ich poznatk\u016F. S\u00ED\u0165 byla naprogramov\u00E1na v prost\u0159ed\u00ED Matlab a byly vygenerov\u00E1ny i bin\u00E1rn\u00ED k\u00F3dy pro demo verzi. Program byl testov\u00E1n na re\u00E1ln\u00FDch datech elektr\u00E1rny M\u011Bln\u00EDk 1 s dobr\u00FDmi v\u00FDsledky predikce NOx a p\u0159edpokl\u00E1d\u00E1 se jeho vyu\u017Eit\u00ED jako sou\u010D\u00E1st komplex\u00EDho bal\u00EDku \u0159e\u0161en\u00ED dle po\u017Eadavku z\u00E1kazn\u00EDka I. & C. Energo." . . "RIV/68407700:21220/11:00192337" . . "21220" . . "K\u0159ehl\u00EDk, K." . . .