Study On Predication Model Of Di-n-Butyl Phthalate Degradation Based On BP Neural Network In The Anaerobic/Anoxic/Oxic Wastewater Treatment Process

With the large application of global plastics, phthalates have become one of the most common pollutants and its'impact on water bodies becomes gradually serious. Therefore, the most common phthalates-dibutyl phthalate (DBP) was selected as a target pollutant in this paper and AAO (anaerobic-anoxic-aerobic) activated sludge treatment process was used to determine the optimum conditions of DBP removing. In this study, the BP neural network predication model of DBP was estabilished which took BP neural network as theory basis and the optimization and simulation of model were accomplished.By starting the reactor and optimizing the process conditions of AAO system, the results showed that after 15 days'culuture, the removal rates of all conventional indicators and DBP were more than 80%, and the quality of effluent was good, the system operated stable. When HRT=8h, SRT=15d, the removal rate of effluent COD reached more than 90%, the removal rate of ammonia was 97%, the removal rate of DBP was 95.62%, and the optimum system condition was HRT=18h, SRT=15d.At different HRT, SRT conditions, the removal rate of DBP all reached 90% in system; the DBP removal in aerobic tank>the removal in anoxic tank>the removal in anaerobic tank, and the DBP removal in aerobic tank took the total removal accounted for 55-65%; the DBP absorpted by activated sludge in each tank took the total DBP in mixture content accounted for more than 78%; in the system, the degradation of DBP mainly was: about 3% discharged with the effluent, 25% accumulated by the system, 2% discharged with the sludge and 70% of biodegradation.According to the linear relationship of DBP biodegradation kinetic model: y ? Kn x?n, the degradation kinetics fitting formulas in anaerobic, anoxic and aerobic conditions were achieved and the Kbio, KS,ηvalues under various conditions were calculated. The results showed that theηvalues were 0.56, 0.85 and 1.00 in the anaerobic, anoxic and aerobic conditions respectively, which were higher than the recommended values of ASM2 (the recommended values ofηFe andηNO3 were 0.1 and 0.6 seperately). The degradation rate constant K and substrate saturation coefficient KS in anaerobic, anoxic and aerobic conditions were 37.31, 56.18, 66.22 and 603.6, 960.06, 1076.47 respectively and gradually increased; the model predicted results of effluent DBP concentrations in each pool, the predicting results were consistented with the experimental results and the relative error was less than 10%.The selected 75 sets of data were processed by cluster analysis and principal component analysis methods, 5 sets of data which had a negative impact on the model predication were removed, the remaining 70 sets of experimental data were used to optimize the BP neural network, the number of layers of neural network model was finally determined as three, and its'structure was (6, 16, 1); the transfer function from input layer to hidden layer was tansig function, while the transfer function from hidden layer to output layer was pureline function; the training function was trainlm; the learning rate was 0.03 and the momentum factor was 0.8.After the parameters of the model was determined, the selected 55 groups data from 70 groups data were took as the training samples, 15 groups were took as the test samples and to train the model, only through 11 epochs, the training goal was reached. Using the established BP neural network model to simulate, the simulation results showed that the network output and actual output were very close, the RMSE, MSE and MAPE of network were very small and the values were 0.35012, 0.12258 and 1.6966 respectively, and R was 0.99382, while the absolute relative errors ranged from 0.071% to 4.4898%; the RMSE, MSE and MAPE of network in forecasting were also relatively small and were 1.0231, 1.0467, 5.6253 seperately, and R was 0.88199, while the relative error rate ranged from 0.1328% to 11.077%, which showed the predicted results of the BP network were good.Compared the esatabilshed BP neural network model with the dynamic model, the results showed that the prediction error of BP neural network smaller than the kinetic model and the BP neural network prediction model more accurate than the kinetic model.