Time-aware evaluation of metaheuristic-optimized support vector regression for permeate conductivity in hybrid nanofiltration–reverse osmosis desalination

Abstract Machine learning models of desalination plants are often tested on random splits of time-ordered records, which can overstate their accuracy. This study evaluated support vector regression (SVR) tuned by Particle Swarm Optimization (PSO), the Genetic Algorithm (GA), Bayesian Optimization (BO), and Harris Hawks Optimization (HHO) for predicting permeate conductivity in a trihybrid nanofiltration–reverse osmosis pilot plant. Seventy-three published observations of five operating variables, covering about 55 days, were used. All preprocessing and tuning relied on training data only. The models were tested under random, blocked, and chronological validation and compared with default SVR, multiple linear regression, random forest, the training mean, and persistence. Persistence, which uses the last measured value, gave the lowest testing error in every scheme (RMSE 26.6–34.0 µS/cm). No tuned model showed a statistically supported improvement over default SVR. Moving from random to chronological testing raised the RMSE of the tuned models from 44 to 47 to 61–69 µS/cm. In the chronological test, the tuned models over-predicted permeate conductivity by 33–39 µS/cm on average, consistent with sensitivity to temperature values outside the training range. Permutation feature importance (PFI) changed with the validation scheme, and a residual-based uncertainty model underestimated the testing errors. Within the high-pressure period, the five inputs showed weak linear correlations with permeate conductivity, and none of the tested models extracted useful predictive skill from them. Hyperparameter optimization therefore did not improve forward prediction for the dataset. Time-aware validation and simple reference models should be standard in the evaluation of data-driven models for membrane plants. Additional process variables and prospective validation are needed before such models can support real-time operation.

Authors

Institutions

Publication Details

Journal
Journal of King Saud University - Engineering Sciences
Published
2026-11-12
DOI
https://doi.org/10.1007/s44444-026-00135-w
Primary Topic
Membrane Separation Technologies
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Time-aware evaluation of metaheuristic-optimized support vector regression for permeate conductivity in hybrid nanofiltration–reverse osmosis desalination

Ali Aldrees, Salisu Dan’azumi, Sani Isah Abba, Abdulhayat M. Jibrin
Journal of King Saud University - Engineering Sciences
Membrane Separation Technologies
article

Time-aware evaluation of metaheuristic-optimized support vector regression for permeate conductivity in hybrid nanofiltration–reverse osmosis desalination

Ali Aldrees, Salisu Dan’azumi, Sani Isah Abba, Abdulhayat M. Jibrin
article en

Abstract

Abstract Machine learning models of desalination plants are often tested on random splits of time-ordered records, which can overstate their accuracy. This study evaluated support vector regression (SVR) tuned by Particle Swarm Optimization (PSO), the Genetic Algorithm (GA), Bayesian Optimization (BO), and Harris Hawks Optimization (HHO) for predicting permeate conductivity in a trihybrid nanofiltration–reverse osmosis pilot plant. Seventy-three published observations of five operating variables, covering about 55 days, were used. All preprocessing and tuning relied on training data only. The models were tested under random, blocked, and chronological validation and compared with default SVR, multiple linear regression, random forest, the training mean, and persistence. Persistence, which uses the last measured value, gave the lowest testing error in every scheme (RMSE 26.6–34.0 µS/cm). No tuned model showed a statistically supported improvement over default SVR. Moving from random to chronological testing raised the RMSE of the tuned models from 44 to 47 to 61–69 µS/cm. In the chronological test, the tuned models over-predicted permeate conductivity by 33–39 µS/cm on average, consistent with sensitivity to temperature values outside the training range. Permutation feature importance (PFI) changed with the validation scheme, and a residual-based uncertainty model underestimated the testing errors. Within the high-pressure period, the five inputs showed weak linear correlations with permeate conductivity, and none of the tested models extracted useful predictive skill from them. Hyperparameter optimization therefore did not improve forward prediction for the dataset. Time-aware validation and simple reference models should be standard in the evaluation of data-driven models for membrane plants. Additional process variables and prospective validation are needed before such models can support real-time operation.

Journal of King Saud University - Engineering SciencesVol. 38(8)
King Fahd University of Petroleum and Minerals (SA), Prince Mohammad bin Fahd University (SA), Prince Sattam Bin Abdulaziz University (SA)
Prince Sattam bin Abdulaziz University
Clean water and sanitation
Membrane Separation Technologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon