(659c) Dynamic Data-Driven Models of Reverse Osmosis Desalination Systems

The control and optimization of reverse osmosis (RO) desalination plants can benefit from the use of plant-specific process models that account for feed conditions (salinity, temperature and pH) as well as operational deviation of various plant components (e.g., pumps, valves). In this regard, data-driven models, developed using machine-learning techniques, can be particularly useful for assessing plant performance, operational deterioration (e.g., due to membrane fouling and failure of plant components), as well as forecasting of plant performance based on historical operational data. Accordingly, the present study presents a systematic approach for the development of data-driven models of RO membrane plant operation based on real-time data. The approach was developed based on data acquired from a novel fully-automated and remotely controlled mini-mobile-modular (M3) RO pilot plant designed and built at UCLA. Operational data with the M3 plant were obtained over a wide range of process conditions (e.g., product water recovery, feed salinity, temperature, pH, feed pressure and cross flow velocity). The M3 RO plant rapidly and automatically generated (aided by its embedded computer controlled system), sets of time-series data for prescribed and randomly set cyclical process conditions (including the effect of process perturbations) under both dynamic (trajectory toward control set points) and steady state operations. The acquired plant data (recorded at high frequency) were utilized to develop data-driven models based on Self-Organizing-Maps (SOM) and Support Vector Machines (SVM). Plant sensors (e.g., salinity and flow rate monitors) can fail or malfunction (even if for short periods) and as was also encountered in the present plant. Therefore, various data filters were applied along with assessment of data consistency (based on overall and salt mass balances) for fault detection and isolation (e.g., identification of malfunctioning plant sensors). Subsequently, optimal data-driven plant models were developed using the appropriate feed conditions (as input variables) to predict the transitions between different states of plant operation after changes in the manipulated variables, thereby providing accurate estimates of permeate and retentate conditions (flow-rate and conductivity). An important element in the model development was a variable-length data segment (“window”) that was introduced in order to include information of most recent trends of plant operation. The current approach demonstrates that dynamic data-driven RO plant models can be developed at a reasonable level of accuracy, and can be used in tandem with deterministic models to describe the complete domain of plant operability under both steady and unsteady state conditions.