Membrane Flux Prediction In A Disc Tube Reverse Osmosis Using Radial Basis Function Neural Network

Because of the superiority in stabilizing effluent quality,Reverse osmosis technology is widely used in landfill leachate which has high ammonia nitrogen concentration and large fluctuation of water quality and quantity.For both designers and managers,the performance of landfill leachate reverse osmosis separation is the focus of their attention.The factors that affect the separation performance of reverse osmosis or the flux of water fil m are often complex,and it’s hard to describe effectively with traditional theoretical models in actual engineering.In view of this problem,this paper studied the main factors affecting the flux of reverse osmosis membrane and the prediction ability of radial basis function artificial(RBF)neural network to reverse osmosis membrane flux.And a two level dish type reverse osmosis(DTRO)system of leachate in a municipal solid waste landfill site in Southwest China is taken as the research object.The main results are as follows:(1)In the two level DTRO system,by controlling the operating pressure difference and influent pH,the test verified that they had a strong correlation with the flux of water production.And the correlation coefficient R~2 was 0.987 and 0.864,respectively.The test also analyzed the relationship between membrane pressure difference and running time in the two cleaning cycles,there’s little change in inlet conductance and continuous increase in inlet conductance.It’s verified that the influent conductance and the operation time have an important relationship with the reverse osmosis flux.(2)By controlling the operating pressure of a DTRO system in a short time,this paper found that the flux of water production increased with the increase of transmembrane pressure after the initial 7 min.But in the next 37 min,although the transmembrane pressure continued to rise,the flux of water production did not change.This paper holds that with the increase of operating pressure,the solute concentration of the raw material side increases,resulting in concentration polarization phenomenon which made the increase of local resistance and osmotic pressure counteract the increased mechanical pressure.(3)Based on the theory of irreversible thermodynamics,a theoretical model of two stage DTRO membrane flux was established by regression analysis.Although the theoretical model can better fit the training data,and the adjusted correlation coefficient in the first and second level systems were 0.973 and 0.976,respectively.But the prediction results for test data were not ideal.The average relative error and the maximum relative error of the two level system were obviously inferior to the RBF neural network model.(4)This paper selected five indexes which include time,conductance,temperature,pH and pressure difference,and established prediction models for two level DTRO membrane flux by using three representative RBF neural networks algorithms which include K-means clustering,orthogonal least squares(OLS)and resource allocation(RAN)algorithm.The test results showed that the K-means algorithm and the RAN algorithm have certain advantages over the OLS algorithm in terms of prediction performance and training time,and OLS algorithm and RAN algorithm can find smaller network structure.From the perspective of practical engineering applications,RAN neural network which has the characteristics of dynamic learning can overcome the inadequately collected information in the process of operation and constantly adapt to the changes in the operating environment.