Abstract: 　The deep nitrogen removal efficiency of the denitrifying biological filter (DNBF) was accurately estimated under different external carbon sources and C/N conditions by establishing a DNBF deep nitrogen removal prediction model based on support vector regression (SVR) and BP neural network (BPNN), which was further optimized using an evolutionary algorithm. The model was trained and its generalization ability was verified using experimental data from DNBF, with the optimal prediction model determined based on evaluation parameters. Results showed that SVR had better prediction performance for TN removal rate (R²=0.904) compared to BPNN (R²=0.876). By applying differential evolution algorithm (DE)-SVR and elite retention Genetic algorithm (EGA)-BPNN optimized by evolutionary algorithm, improvements of 1.5% and 11.5% were achieved respectively, compared to SVR and BPNN models. EGA-BPNN demonstrated significantly better predictions for TN removal rate, NO^－₂-N mass concentration, and NO^－₃-N mass concentration with R² values of 0.991, 0.971, and 0.926 respectively when compared to other models; indicating that the structure and model parameters of the neural network were effectively optimized through the evolutionary algorithm resulting in improved model performance. Additionally, EGA-BPNN exhibited good generalization ability as evidenced by R² values of 0.969 for TN mass concentration, 0.980 for NO^－₂-N mass concentration, 0.974 for NO^－₃-N mass concentration, and 0.864 for COD mass concentration, thus confirming its effectiveness in predicting nitrogen removal efficiency of DNBF under different external carbon source strategies.

Key words: denitrifying biofilters, external carbon sources, support vector regression, back-propagation neural network, evolutionary algorithm