Data-driven models for performance prediction in a microalgae-based wastewater treatment process
Year: 2026
Type: preprint
Abstract: Microalgae-based wastewater treatment has the potential to advance tertiary phosphorus recovery while producing biomass for resource valorization. The development of reliable near-term prediction models can help reduce system upsets, support proactive process control, and optimize resource utilization. In this study, Light Gradient-Boosting Machine (LGBM), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) models were comparatively evaluated for predicting effluent ammonium, nitrate, and orthophosphate concentrations and biomass dynamics in a full-scale (568 m 3 ·d -1 ) microalgal system using 25 months of operational and microalgal community structure data. LSTM consistently achieved the highest prediction accuracy while effectively capturing system upset events. SHapley Additive exPlanations (SHAP) analysis provided interpretable insights into feature importance across both feature and temporal dimensions, with SHAP-guided feature reduction substantially reducing computational requirements without compromising predictive performance. Benchmarking against the mechanistic process model further identified the complementary application niches of mechanistic and data-driven approaches, providing practical guidance for model selection.
Source: ChemRxiv
Authors: Ga-Yeong Kim, JooMan Kim, Jeremy S. Guest
Institution: University of Illinois Urbana-Champaign
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213984044