(688d) Data Driven Neural Netowrk Based Models of Ro Desalination Plant Operation and Ro Membrane Performance
A machine learning modeling approach was investigated as a means of developing data-driven models for describing RO desalination plant performance (e.g., flux and salt rejection) and for evaluating the suitability of RO membranes for potable water applications requiring the rejection of a broad range of organic compounds. The concept of plant memory (incorporated to enable description of process fluctuations), along with basic input process operating parameters, formed the basis input variables for describing plant behavior. Cognitive neural network model training was carried out with the normalized permeate flux and salt passage for various model architectures and memory time intervals. Model results demonstrated that plant performance could be described to a reasonable level of accuracy with respect to both permeate flux and salt passage with a plant memory time interval. Forecasting of plant performance was shown to be feasible and with good accuracy with a reasonable memory time interval to enable practical process control.
The current approach is providing the basis for developing and incorporating neural network data-driven models in a control strategy and early-warning system of the deterioration of RO plant performance. In this regard, the passage of organics through RO membranes is particularly critical for applications that involve RO membranes in water treatment plants. Neural network models can be effective in generating Quantitative Structure-Property Relations (QSPR) for the organic passage (P), sorption (S) and rejection (R) using the most relevant set of molecular descriptors. In the present work, the approach was demonstrated based on an experimental data set of fifty organics with four different RO membranes. A number of feature selection methods were employed. Pre-screening was carried out, with Principal Components Analysis and SOM of the chemical domain for the study chemicals, as defined by chemical descriptors, to identify the applicability domain and chemical similarities. The QSPR models predicted organic passage, rejection, and sorption within the range of the standard deviations of measurements for the experimental data set of fifty compounds. The application for the approach for compounds of interest, for which experimental data were not available, demonstrated reasonable mass balance closures. The implications of the above approaches will be discussed with respect to the development of a comprehensive tool for assessing RO plant performance in RO water treatment processes.