Intelligent temporal causal inference framework for wastewater treatment plant nitrogen removal: Multi-stage spurious causal elimination

Abstract

Wastewater treatment operations require transparent, interpretable models for regulatory compliance and safety, yet the intricate mechanisms involved in biological nitrogen removal present significant challenges for achieving interpretable mechanistic understanding. To address this, this study proposes AquaCausal, a novel hybrid causal inference framework that integrates the time-aware Peter & Clark Momentary Conditional Independence (PCMCI) algorithm, deep learning, and a multi-stage mechanism to eliminate spurious causal relationships. A perturbed simulation dataset was generated and validated using a calibrated wastewater treatment plant (WWTP) model, establishing a benchmark for causal discovery. The framework systematically refines causal relationships through L1-regularized Granger causality testing, permutation feature importance analysis, and a four-dimensional robustness assessment. This process reduced initial potential causal relationships by 73 %, ultimately identifying 25 core causal relationships with high confidence. The derived causal network quantified key time-lagged dependencies, establishing an interpretable, data-driven representation of nitrogen removal mechanisms for optimizing treatment processes and facilitating adaptive intelligent control.