BP神经网络模型在表层沉积物及其非残渣态组分吸附双酚A研究中的应用

应用沉积物吸附双酚A(BPA)BP神经网络模型,模拟了松花江表层沉积物的不同泥水比、非残渣态组分(有机质、铁氧化物、锰氧化物)和BPA初始浓度对BPA吸附量的影响。所建BP神经网络模型相关系数R²为0.9665,校正集均方差(MSE_c)、验证集均方差(MSE_v)和预测集均方差(MSE_p)分别为0.0068、0.0596和0.1285;利用遗传算法优化估算了基于BP神经网络模型的沉积物吸附BPA的最大吸附量,优化值与实验值的相对偏差为0.96%~8.21%。此外,利用BP神经网络模型预测了沉积物非残渣态组分(有机质、铁氧化物、锰氧化物)质量百分比及摩尔含量变化与BPA吸附量的关系,经分析可知,铁氧化物和有机质对沉积物吸附BPA起着促进作用,沉积物非残渣态组分吸附BPA的相对贡献(K)为K_Fe>K_OMs>K_Mn,即沉积物中铁氧化物是BPA的主要吸附位,而Mn氧化物则对沉积物吸附BPA起着抑制作用。

Application of BP Neural Network to the Adsorption of Bisphenol A onto Surfacial Sediments and Their Non-residual Fractions

BP neural network model, which could be thought of as being related to artificial intelligence, machine learning, parallel processing and statistics, is being used more and more common. In the present study, a three layer BP neural network model of bisphenol A (BPA) adsorption onto the surfacial sediments (SSs) sampled from Songhua River in Jilin Province had been established to simulate the influence of several different factors, such as solution to solid ratio, contents of non-residual fractions (organic matters, Fe oxides and Mn oxides), and initial concentration of BPA, on the adsorption capacity of BPA. The correlation coefficient (R²) of the established BP neural network model was 0.966 5, which was larger than 0.8. The mean square error of the calibration set (MSE_c), the root mean square error of validation set (MSE_v), and the mean square error of the predication set (MSE_p) was 0.006 8, 0.059 6, and 0.128 5, respectively. The maximum adsorption of BPA adsorbed onto SSs collected from Songhua River was estimated and calculated by genetic algorithms (GA). GA is a search technique used in computing to find exact or approximate solutions to optimization and search problems and it is categorized as global search heuristics, on the basis of the established BP neural network model. The optimized results of the maximum capacity of BPA absorbed onto SSs (without treatment, H₂O₂ extraction, NH₂OH·HCl extraction, and (NH₄)₂C₂O₄ extraction) were 0.532 mg/g, 0.502 mg/g, 0.917 mg/kg and 0.8992 mg/g, and under the same conditions, the experimental values were 0.542 mg/g, 0.445 mg/g, 1.081 mg/g and 0.836 mg/g, respectively. The relative errors between the optimal values by GA and the experimental ones were in the range of 0.96%~8.21%. The amount of BPA adsorbed onto SSs was predicted using the established BP neural network model as a function of non-residual fractions including organic matters, Fe oxides and Mn oxides on a mass or a molar base, respectively. The predicted results of the maximum BPA adsorption capacity on a mass base showed that there has been a general uptrend of the BPA adsorption with the increase of Fe oxides and organic matters and a general downtrend with increase in Mn oxides. Meanwhile, the maximum capacity of BPA adsorbed onto SSs indicated the same results on a molar base with the results obtained on a mass base. The relative contributions of the BPA maximum adsorption onto SSs, expressed by the ratio of the mass of adsorbed BPA to the contents of non-residual fractions, were calculated as follows: K_Fe=0.002 8, K_OMs=0.000 2 and K_Mn=-0.031 8, respectively. It could be inferred that both of the Fe oxides and organic matters have positive effect on the BPA adsorbed by SSs, while Mn oxides inhibited the adsorption of BPA onto SSs. Hence, the contributions of the non-residual fractions (including organic matters, Fe oxides and Mn oxides) onto SSs to the maximum adsorption of BPA followed the order as: K_Fe>K_OMs>K_Mn. The fact that Fe oxides was confirmed as the main binding site for BPA adsorption onto SSs was demonstrated through the mechanism analysis through the established BP neural network model, yet the reasons why adverse effect of Mn oxides on the adsorption of BPA onto SSs should be further studied.