Integrating machine learning algorithm with sewer process model to realize swift prediction and real-time control of H2S pollution in sewer systems

Abstract

The frequent occurrence of safety incidents in sewer systems due to the emergency toxicity of hydrogen sulfide (H2S) necessitate timely and efficient prediction, early warning and real-time control. However, various factors influencing H2S generation and emission leads to a substantial computational burden for the existing dynamic sewer process models and fails to timely control the H2S exposure risk. The present study proposed a swift prediction model (SPM) that combined the validated dynamic sewer process model (the biofilm-initiated sewer process model, BISM) with a high-speed machine learning algorithm (MLA), achieving accurately and swiftly predict the dissolved sulfide (DS) concentration and H2S concentration in a specific sewer network. Based on Gradient Boosting Decision Tree-based SPM, the simulated concentrations of DS and H2S are 1.95 mg S/L and 214 ppm, respectively, which are closely to the field-measured values of 1.82 mg S/L and 219 ppm. Notably, SPM achieved a computation time of less than 0.3 sec, and a significant improvement over BISM (> 5000 sec) for the same task. Moreover, the real-time and dynamic dosing scheme facilitated by SPM outperformed the conventional constant dosing scheme provided by dynamic sewer process model, which significantly improved the H2S control completion rate from 69% to 100%, and achieved a significant reduction in chemical dosage. In conclusion, the integration of dynamic sewer process models with MLA addresses the inadequacy of monitoring data for MLA training, and thus pursues swift prediction of H2S generation and emission, and achieving real-time, effective, and economic control of H2S in complex sewer networks.