Temporal fusion transformer model for predicting differential pressure in reverse osmosis process

Reverse osmosis (RO) is an advanced water treatment technology that effectively removes a broad spectrum of pollutants from water. A critical aspect in assessing the integrity of RO membranes and maintaining filtration systems is the differential pressure (DP). Conventional methods for predicting DP, which often depend on sensors or programmable logic controllers, encounter limitations due to the complexity of process conditions and variability in operational data. This study seeks to improve DP prediction in industrial RO processes through the application of deep learning models. We implemented a state-of-the-art temporal fusion transformer (TFT) model that effectively differentiates between static and dynamic variables. The TFT-based model demonstrated superior performance with an R² value exceeding 0.9813, significantly outperforming the long short-term memory (LSTM) model, which achieved an R² value >0.9364. This enhancement in prediction accuracy indicates that transformer-based algorithms, by concentrating on key features, can surpass more complex neural networks in regression tasks. Notably, the TFT model adeptly managed static variables—typically problematic for time-series models—alongside dynamic variables. The effectiveness of the model in incorporating static inputs, such as process numbers and cleaning injection status, was confirmed by R² values of 0.9813 with the static encoder and 0.8980 without it. Furthermore, we evaluated the reliability of the model by examining the relative importance of input features through an attention map. The adaptability and interpretability of this approach confer substantial benefits, enhancing energy efficiency and operational performance in various industrial settings.