Optimizing sustainable desalination plants with advanced ML-based uncertainty analysis

Reliable, cutting-edge environmental research and innovative solutions to critical environmental challenges are essential for a sustainable future. The desalination industry is increasingly interested in predicting the performance of hybrid nanofiltration/reverse osmosis (NF/RO) systems for treating very saline brine to recover valuable resources. This study aimed to optimize the effectiveness of a Long-Short-Term Memory (LSTM) model by applying two distinct metaheuristic optimization techniques: Genetic Algorithm (GA) and Crow Search Algorithm (CSA). The focus was on leveraging these algorithms for uncertainty analysis and forecasting in hybrid NF/RO desalination processes within Saudi Arabia. Furthermore, Particle Swarm Optimization (PSO) was utilized to determine the most suitable data models, consisting of three and four parameters, for this modelling purpose. Statistical tests based on analysis of variance (ANOVA) test (ANOVA F-test and the Welch F-test), covariance analysis (covariance matrix, t-statistics, and p-values), and pairwise Granger causality tests were conducted. Statistical methods and visual techniques were employed to assess and compare the precision of the LSTM model when integrated with the metaheuristic algorithms (LSTM-GA and LSTM-CSA) against the standalone LSTM model. The results demonstrated that the metaheuristic algorithms provided higher accuracy than the standalone LSTM model in both data models. In hybrid NF/RO desalination prediction, the LSTM-GA (MAE=0.13), LSTM-CSA (MAE=0.14), and LSTM (0.20) models achieved increasingly higher accuracy. Besides, the Monte Carlo method was utilized to evaluate the uncertainty linked with the predictions. The findings indicated that the LSTM model integrated with the GA (LSTM-GA) demonstrated the slightest uncertainty in its predictions.