Research on prediction method of hydrogen sulfide concentration in anaerobic pool of vegetable pickling industry...

Anaerobic pools in the vegetable pickling industry are typical confined spaces where hydrogen sulfide (H₂S) readily accumulates, leading to frequent fatal poisoning. Traditional single-point detection methods fail to capture the three-dimensional stratified distribution of H₂S, and existing machine learning prediction models mainly focus on municipal sewage systems, lacking systematic research tailored to the unique structural, operational, and environmental characteristics of anaerobic pools in the pickling industry. To address this critical gap, this study proposes an interpretable multi-feature fusion prediction method that integrates structural, geometric, environmental, sewage state, and detection position features. Systematic on-site detection was conducted on 46 anaerobic pools from 38 enterprises across three major pickling regions in China, resulting in a high-quality multi-feature dataset comprising 99 sample groups covering diverse production scales and operational scenarios. After standardized feature encoding and normalization to eliminate dimensional differences, the prediction performance of seven classical machine learning algorithms was systematically compared. The results indicate that the XGBoost model achieved the best overall performance, with a coefficient of determination (R²) of 0.94. Its mean absolute error was reduced by 32% to 66% compared with the other six models. Feature importance analysis revealed that the pool closure status and sewage agitation state were the two dominant factors affecting H₂S accumulation, followed by the internal temperature, wind speed, and humidity. This study makes two key contributions: (1) it establishes a transferable five-dimensional feature representation system for gas distribution prediction in confined spaces, overcoming the limitation of traditional models that rely solely on single-category environmental parameters; (2) it enables accurate three-dimensional prediction of H₂S concentrations in unmeasured areas using only a few detection points, providing a low-cost and high-efficiency technical solution for pre-operation risk assessment and daily safety management. The proposed method can effectively support the optimization of the monitoring point layout and the formulation of targeted prevention measures, contributing to the reduction of H₂S poisoning accidents in the vegetable pickling industry.