多参数影响下污水总氮浓度预测最优方法研究

Optimal method for predicting the total nitrogen concentration of wastewater under the influence of multiple parameters

污水总氮(TN)深度脱除是当前我国污水处理领域的重大科技需求.TN的去除受到多种环境及操作条件的影响,开发多参数条件下稳健的TN浓度预测方法是降低污水厂能耗、实现智能化控制的重要前提.针对以上问题,以某实际污水处理厂反硝化深床滤池为例,采用BP神经网络(BP)、量子遗传算法优化的BP神经网络(QGA_BP)、改进的QGA_BP和支持向量回归机(SVR),在进水流量和碳源投加量等13种变量条件下,对滤池出水TN进行了模拟预测.共选取147组数据,其中130组用于出水水质和工艺参数的拟合模拟,17组用于结果验证.将总氮实测值依次与BP,QGA_BP和改进的QGA_BP神经网络以及SVR预测结果进行对比,相关系数R2依次增大,分别为0.221,0.275,0.826和0.951,即预测值与实测值之间的拟合度逐渐升高.SVR克服了神经网络预测误差较大的问题,对多参数影响下TN浓度的预测具有较高的准确性和稳定性,用其替代常用的神经网络算法具有明显的优势.

The deep removal of total nitrogen（TN）is a major scientific and technological demand in the field of wastewater treatment in China nowadays.Multiple parameters-affected TN removal desires a robust prediction of TN concentration in order for the energy consumption reduction and intelligent control of the treating system.In respect of the issues above,taking the denitrification deep bed filter of a real wastewater treatment plant as an example,this paper applied four types of predicting methods containing the Back Propagation（BP）neural network,quantum genetic algorithm BP neural network（QGABP）,improved QGABP neural network and support vector regression machine（SVR）,to simulate effluent TN concentration under a total of 13 parameters of inlet flow,carbon source addition,and so on.A total of 147 groups of data were selected,among which 130 groups for the simulation of effluent water quality and process parameters,and 17 groups for results verification.Measured values of TN were compared with predicted results from BP,QGABP,improved QGABP and SVR,respectively.The fitting degree between the predicted value and the measured value increased gradually,with the correlation coefficients（R2）of 0.221,0.275,0.826 and 0.951 for each method,respectively.SVR algorithm overcomes the relatively big error by neural network and has a high accuracy and stability for TN concentration prediction under the influence of multiple parameters,exhibiting a good alternative to neural network algorithms.