Estimation of Hydrogen Production in a Laboratory Scale Anaerobic Reactor Using Machine Learning

Biohydrogen production through dark fermentation is influenced by the interaction of hydraulic, organic, thermal, and biological conditions, making the online monitoring of continuous anaerobic reactors challenging. This study presents a hybrid virtual-sensing framework that combines CatBoost machine learning with a simplified mass-balance model derived from the Anaerobic Digestion Model (ADM1) to estimate the performance of a 2.1 L laboratory-scale expanded granular sludge bed (EGSB) reactor fed with cheese whey. The reactor was operated for 280 days, including a 49-day start-up and acclimation period followed by 231 days of experimental operation under hydraulic retention times from 4 to 8 h and temperatures between 35 and 55 °C. The experimentally determined hydrogen yield coefficient (YH2) was used as the prediction target for the CatBoost model. Model inputs included pH, hydraulic retention time, organic loading rate, temperature, and influent chemical oxygen demand (COD), together with a five-day historical window to capture temporal dynamics. On an independent chronological test set including the final 45 days of operation, CatBoost achieved an R2 of 0.89, an RMSE of 0.066, an MAE of 0.047, and a MAPE of 11.9%. The predicted hydrogen yield coefficient was subsequently incorporated into the simplified ADM1-based model, yielding an R2 of 0.87 for effluent COD prediction and an R2 of 0.88 for daily hydrogen production. The effluent substrate concentration was estimated independently of YH2, whereas the incoming substrate concentration was used in the calculation of the hydrogen yield coefficient. The proposed framework provides a cost-effective approach for monitoring continuous biohydrogen production without requiring routine measurements of volatile fatty acids (VFAs), CO2, or CH4.