Naslov
Fouling detection in industrial heat exchanger using recurrent LSTM neural network

Autori
Rimac, Nikola ; Ujević Andrijić Željka ; Cvjetojević Vladimir

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
27th Croatian Meeting of Chemists and Chemical Engineers, BOOK OF ABSTRACTS / Marković, Dean ; Meštrović, Ernest ; Namjesnik, Danijel ; Tomašić, Vesna - Zagreb : Hrvatsko kemijsko društvo, 2021

Skup
27. hrvatski skup kemičara i kemijskih inženjera (27HSKIKI)

Mjesto i datum
Veli Lošinj, Hrvatska, 05.10.2021. - 08.10.2021

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Neural network, LSTM, Fouling, Process modeling, Heat exchanger
(Neuronske mreše, LSTM, Nastajanje naslaga, Modeliranje procesa, Izmjenjivač topline)

Sažetak
There is a need of continuous detection of fouling formation due to its side effects to the process. Fouling formation in heat exchangers is one of major problems in oil refinery which significantly impacts mechanical and thermal characteristics of heat exchangers. Corrosion and increased pressure drop, as well as lower heat transfer are one of main fouling side effects which lead to production losses and additional energy costs. Also, real plant processes operate in dynamic conditions with frequent process regime changes making fundamental models difficult to develop, therefore require more complex methods. Machine learning models which do not give physical insight into the process mechanisms but describe input-output variable relations can be developed and used. In this work, model for fouling detection based on Long-Short Term Memory recurrent neural network is presented. LSTM recurrent neural network is capable learning order dependance in sequential data like time dependent data continuously acquired from industrial heat exchanger. Number of steps into data history and its impact on model results was examined. Various optimizers and activation functions of LSTM network were examined during model training and testing to give optimal results. Model is developed in Python programming language using Keras library on continuously acquired data from industrial heat exchanger. Model results show its possibility to determine fouling in examined oil refinery heat exchanger.

Izvorni jezik
Engleski

Znanstvena područja
Kemijsko inženjerstvo

POVEZANOST RADA