Robust deep learning model combined with missing input data estimation: Application in a 1000 m3/day high-salinity SWRO plant

Abstract

Predictive models depend heavily on high-quality input data; however, frequent data gaps caused by temporary process shutdowns or sensor failures during the initial implementation of desalination plants present significant challenges. To address these issues, the non-autoregressive multiresolution imputation (NAOMI) technique was applied to fill in missing operational data from high-salinity seawater reverse osmosis (SWRO) systems. Following effective data interpolation using NAOMI, five neural network-based deep learning models were tested to predict three key operational parameters: transmembrane pressure (TMP), energy consumption, and permeate flow rate. Input variables for predicting these parameters included flow rate, temperature, conductivity, pressure, oxidation-reduction potential (ORP), pH, and energy consumption. Applying NAOMI demonstrated significant accuracy improvements across all models, with the long short-term memory (LSTM) model performing best for TMP and energy consumption predictions, achieving root mean square error (RMSE) values of 0.375 and 0.218, respectively. For permeate flow rate prediction, the convolutional neural network-LSTM (CNN-LSTM) model performed optimally, achieving an RMSE of 0.173. Furthermore, Shapley Additive Explanations (SHAP) analysis enhanced model explainability by clarifying the influence of input variables on predictions. Notably, feed conductivity was found to be the most critical factor for TMP prediction, whereas permeate conductivity was the most influential for both energy consumption and permeate flow rate predictions. The study results indicate that integrating advanced data imputation techniques with optimized deep learning models supports effective decision-making and enhances the operational stability of SWRO plants during early development stages.