The neural network approach for estimation of heat transfer coefficient in heat exchangers considering the fouling formation dynamic

• Published in: Thermal science and engineering progress 2024-06, Vol.51, p.102615, Article 102615
• Main Authors: Ilyunin, Oleg, Bezsonov, Oleksandr, Rudenko, Sergiy, Serdiuk, Nataliia, Udovenko, Serhii, Kapustenko, Petro, Plankovskyy, Sergiy, Arsenyeva, Olga
• Format: Article
• Language: English
• Subjects: Energy efficiency; Fouling prediction; Heat transfer; Plate heat exchangers
• ISSN: 2451-9049; 2451-9049
• DOI: 10.1016/j.tsep.2024.102615

•The fouling in PHE for thin juice preheat train is observed.•The fuzzy logic technique was applied to expand the set of training parameters for ANN models.•The FFNN model for prompt fouling estimation makes the prognose with 99%•The LSTM model for long-term fouling estimation makes the prognose with 89% Routine maintenance for plate heat exchanger (PHE) cleaning improves the effectiveness of heat exchange network operation. Until recently, complex mathematical modelling was used to predict the value of the heat transfer coefficient after a certain period of operation of the heat exchanger, as well as the point in time when the coefficient reached the allowable limit. The applied mathematical tools included the systems of differential equations, matrices of heuristic coefficients, which needed a lot of computer resources. This paper offers an artificial neural network (ANN) technique for forecasting the following values: heat transfer coefficient at any time points during the operating period of PHEs; the time point, when the heat transfer coefficient reaches its lower permitted value. In this method, ANN uses the fuzzy logic techniques to expand the set of training parameters for the model, working with data of industrial measurements and data obtained from the mathematical modelling of the process. It allowed to train the developed feed-forward neural network (FFNN) with the coefficient of determination R2 equal to 0.99 and can predict the thermal resistance in PHE based on measurement data. To adequately predict the time-point to reach the limiting value of the heat transfer coefficient, it was proposed a recurrent neural network (RNN) with a hidden layer of long short-term memory (LSTM), where R2 value came up to 0.89.