(469a) Process Modeling of Intermittent Wellhead RO Water Treatment Operation Via Integration of Self-Organizing Maps and Long Short-Term Memory Recurrent Neural Network (RNN)

Membrane-based water treatment and desalination processes are suitable for deployment at various levels of capacity from small home-based systems to systems suitable for municipal, industrial and even agricultural water supplies. Deployment of such water systems, particularly for distributed high recovery wellhead water purification in small communities that are confronted with impaired potable groundwater supplies, where 24/7 of operator availability is infeasible, requires self-adaptive operation that responds to temporally variable environmental and water quality conditions, in addition to fluctuating water use patterns. The operation of such systems is intermittent, and the product water must meet regulatory safe drinking water requirements, whereby removal of the contaminants of concern must ensure that the product water must be well below the regulatory maximum contaminant level (MCL). In order to develop a robust operational strategy of high recovery wellhead water treatment via reverse osmosis (RO), integrated with feed pretreatment and post-treatment, process models are needed that account for actual behavior of system components (pumps, actuators, sensors, membranes) under temporally variable conditions. Accordingly, in the present work a process model was developed for a uniquely designed high recovery wellhead RO water treatment system operating in a remote agricultural community.

Model development was based on a year of operational time-series data and utilized to evaluate system performance and operational models with respect to nitrate and salt passage, as well as membrane permeability. A forecasting approach, accompanied by a comprehensive exploratory analysis was undertaken making use of long short-term memory (LSTM) recurrent neural network (RNN) architecture with and without an attention coefficient. Performance of the RO system was evaluated for different modes of system operation that included flushing, permeate production of safe product water, startup, and shutdown. A total of twenty quantitative model attributes were utilized (including, for example, pressures, flow rates, temperature, conductivity and nitrate concentrations) and their attribute significance was determined for model building. The LSTM model in which an attention coefficient was used had excellent performance for the different modes of operation (i.e., R² value of ~0.95 for both nitrate permeate and salt passage). In order to address the tradeoff between model complexity and robustness, a minimal set of the LSTM model attributes was identified and ranked via Spearman’s correlation coefficient and feed forward feature selection (FFFS) approaches. Subsequently, a supervised learning approach based on self-organizing maps (SOM) was applied to discover similarities (or dissimilarities) among the information provided by the different model attributes and thus ensure that pertinent information is provided for model input. It is noted that the assessment of attribute significance for the LSTM models for permeate nitrate and salt passage was improved via supervised SOM analysis that incorporated class information in the input data. The above was accomplished by inserting additional class components in the input vector to indicate the class-affiliation of the data. The class component increased the clustering tendency of the samples belonging to the same class and expelled samples of foreign classes on the map; thus, stretching the SOM separation between given classes, and optimizing the filtering/extracting attributes of higher significance per class. It is stressed that in deriving the process models, an upper regulatory nitrate level was set given the concern in ensuring that the model would provide accurate prediction of permeate product quality with respect to both nitrate and salt passage. In summary, the machine learning approach developed in this study demonstrated multitude of benefits including: (a) determining deviation of nitrate from the baseline value, (b) forecasting performance of a real RO system with respect to nitrate permeate and salt passage, (c) assessing the correlation between nitrate permeate and salt passage for a real system, and (d) providing a foundation for model-based control of the treatment system for nitrate and salt removal.