Hierarchical patch-based forecasting and transfer learning enable a long-horizon, scalable model for reverse osmosis in a wastewater reuse facility

초록

Accurate long-term forecasting of permeate conductivity is essential for stable operation and compliance with quality standards in wastewater reuse reverse osmosis (RO). However, the prediction is complicated by irregular cleaning-in-place (CIP) events and nonstationary operating conditions. This study developed a hierarchical patch-based forecasting framework that combines a hierarchical dependency mixer (HDMixer) with staged transfer learning for long-horizon permeate conductivity prediction. The target facility is a full-scale brackish water RO plant with a capacity of 90,000 m3/d, operating 12 trains across two parallel lines. An operational time series from 2020 to 2025 comprising 101 variables and 543,448 observations was used. CIP cycles were inferred from operation logs and refined using permeate flow rate and conductivity signatures. Cycle-aligned variablelength patches were generated to retain cycle-specific dynamics without imposing a fixed window. HDMixer consistently outperformed the Transformer and long short-term memory baselines across all forecast horizons from 30 min to 12 h, achieving an R2 of 0.990 at the shortest horizon and maintaining an R2 of 0.849 at the longest. Post-hoc interpretation using conditional Shapley additive explanations indicated that brine flow rate and recovery contributed most to the forecasts. Despite a statistically significant line-to-line domain shift, staged transfer learning recovered most of the performance gains by updating only the prediction head, normalization layer, and input projection. Overall, the proposed framework enables scalable, long-horizon forecasting with physically interpretable explanations and rapid cross-line adaptation for operational decision-making in wastewater reuse reverse osmosis.

키워드