(422b) Neural Network Approach for Modeling the Performance of Ro Membrane Processes

One of the more serious problems encountered in Reverse Osmosis water treatment plants is the occurrence of membrane fouling, which limits process efficiency (separation performances, water flux, salt rejection) and shortens membrane life-time. To accurately quantify and effectively control the adverse impact of membrane fouling and scaling, it is desirable to be able to predict the development of membrane fouling There are no general deterministic models available for predicting the development of fouling in full-scale reverse osmosis plants. The major obstacles in developing such predictive models are the complexity of feed composition, the inability to realistically quantify the real-time variability of feed fouling propensity and lack of understanding of the interplay of various fouling mechanisms as well as the precise role of membrane surface properties and membrane interaction with various foulants and fouling precursors. In order to develop a practical method of describing full-plant performance dynamics, Artificial Neural Networks based models can be developed using actual plant data. Using such models the deviation of process set points and upsets can be identified in advance so as to enable effective process control. In the present study a neural network (NN) based Reverse Osmosis model was developed for predicting process performance and the time evolution of membrane fouling. Using real-time measurements for process variables as well as feed water quality, from a full-scale RO plant, a framework was developed for building neural network based RO models that to evaluate the onset of fouling via the development of soft sensors for anticipating process upsets. The NN output variables were selected so as to ensure that they reveal clear information regarding RO process performance that could subsequently be used for the development of neural network-based soft sensors. In order to avoid redundancy of variables and deterioration of NN performance (e.g., due to unnecessary input parameters) of noise) variable selection was carried out using a number of advanced algorithms. In addition, the spectrum of relevant NN architectures were evaluated relative to the selection of the best set of model input variables in order to optimize the NN structure. Results of the study revealed that the NN-RO models are able to capture changes in RO process performance. Moreover, preliminary results are encouraging in demonstrating that the NN based RO models can be successfully used for interpolation, as well as for reasonable forecasting of process time evolution which is believed to be sufficient and for practical RO process control applications. The implications of the current approach for the development of optimization and control strategies and for forecasting of process upsets due to fouling development will be discussed and demonstrated via a number of examples.