High-precision prediction of unionized hydrogen sulfide generation based on limited datasets and its impact on anaerobic digestion of sulfate-rich wastewater

Abstract

Data-driven models can simulate the complex anaerobic digestion (AD) process and predict the production of unionized hydrogen sulfide (H2S), thus improving methane (CH4) production by alleviating the unionized H2S inhibition. However, due to the limitations of model structures and small datasets, traditional data-driven models cannot accurately simulate the production of sulfides and their morphological differentiation process. In this study, by integrating method of deep neural network (DNN) model and random standard deviation sampling (RSDS), a method called RSDS-DNN was established to simulate the generation of unionized H2S in the AD process. The virtual data was generated by the RSDS method based on the mean value and standard calculated from the raw dataset collected from the experiment. Compared with the results of the artificial neural network (ANN) models trained with the raw dataset (the lowest MSE of 464.02 and the lowest MAPE of 29.00% for testing dataset), the DNN model trained with a virtual dataset showed obvious advantages. The DNN model with 2 hidden layers and 100 hidden neurons in each layer trained with a virtual dataset with 1060 samples achieved the best performance with a mean R-square (R2) of 0.978, the minimal MAE of 87.80, and the minimal MAPE of 9.21% for the testing dataset. The Integrated RSDS-DNN model is a potential approach to assist the reducing of unionized H2S production and enhance CH4 production in the AD process.