Causal Structure Analysis of THMs Influencing Factors in Water Treatment Processes Using Bayesian Network

This study applied a Bayesian Network (BN) model to identify the causal factors influencing the formation of trihalomethanes (THMs) during the chlorination process in water treatment plants. While conventional multiple linear regression (MLR) explains THMs concentrations based on linear correlations among variables, it is limited in representing probabilistic dependencies and directional relationships. In contrast, the BN approach allows probabilistic representation of dependencies and directional relationships through structural and parametric learning. Using twenty years (2003–2022) of operational data from the A Water Treatment Plant in the Han River basin, four structure learning algorithms—Hill-Climbing, Tabu Search, Grow–Shrink, and Max–Min Hill-Climbing (MMHC)—were compared. Tabu and MMHC yielded the best model fit (BIC=–1914.059). A consensus network combining common arcs across algorithms was constructed, and 500 bootstraps were performed to assess structural robustness. Edges satisfying strength ≥ 0.7 and direction ≥ 0.7 were retained, resulting in a double-robust causal network. The final DAG(Directed Acyclic Graph) identified O_IV (iodine number), O_Pre_CL (pre-chlorination dose rate), R_TEMP (temperature), T_pH(final treated water pH), and T_RCL (residual chlorine) as variables directly associated with THMs, while ammonia nitrogen of raw water(R_NH3) and COD of raw water(R_COD) acted as indirect (grandparent) nodes through O_Pre_CL and T_pH, respectively. Standardized coefficients (β) showed negative effects for O_IV (–0.449) and T_pH (–0.164), and positive effects for R_TEMP (+0.398), T_RCL (+0.153), and O_Pre_CL (+0.011). These results indicate that THMs formation is primarily governed by operational factors such as chlorine dosage, temperature, pH, and residual chlorine. The BN thus serves as a probabilistic causal model with both predictive accuracy and explanatory power, providing a theoretical foundation for data-driven decision-making in disinfection by-product control and process optimization in water treatment systems.