(534f) Modeling of UF Performance in Pretreatment of Seawater RO Feedwater Using Neural Network with Evolutionary Algorithm and Bayesian Binary Classification

A machine learning approach to describe the dynamics of ultrafiltration (UF) performance in pretreatment of seawater reverse osmosis (RO) feedwater was explored via Ensemble Backpropagation Neural Networks (BPNN) model with Alopex Evolutionary Algorithm (AEA). The model was developed based on data generated for an integrated UF-RO seawater system having a capacity of producing up to about 18,000 gallons/day of permeate product water. The desalination system consisted of three multi-bore inside-out hollow-fiber ultrafiltration elements and three 8 inch high salt rejection RO elements. System operation generated over 13 million data samples collected over a three year period. The desalination unit was operated over a range of conditions generating a total of 180 datasets encompassing both short-term and long-term operational periods. The UF was aver a range of fixed and variable coagulant dosing strategies and backwash of fixed and self-adaptive frequency. UF performance data were also generated from operation with a self-adaptive coagulant controller. Water quality data varied considerably over the operational period that included a storm event which served as a stress test for the generated UF performance model.

The dynamic performance of ultrafiltration was modeled via a data-driven Back Propagation Neural Networks, with Alopex-based evolutionary algorithm (BPNN-AEA). In order to encompass the wide range of raw water quality and operating conditions a family of BPNN-AEA models were developed track the progression of UF membrane hydraulic resistance (during both filtration and backwash), as well as backwash efficiency. The BPNN-AEA models were integrated with a Bayesian classification approach to automatically recognize filtration and backwash modes. To address data diversity challenges (i.e., bias-variance tradeoffs) for a single machine learning model, an AdaBoost ensemble strategy was followed. Model development workflow included preprocessing (including data normalization with respect to the maximum and minimum of the dataset) and exploratory data analysis to remove attributes redundancy followed by establishing attribute significance (among the 22 identified relevant attributes), and the minimum set of required attributes. The initial pool of model attributes was established as the basic process variables that impact UF filtration and backwash as rationalized. Initially, the feed forward feature selection (FFFS) and minimal redundancy maximal relevance (mRMR) were used for identification and ranking of the attributes with respect to model performance. Subsequently, the significance of the attributes was determined iteratively where at each iteration an attribute was sequentially added to the model. Attributes that increased the model performance above a minimum prescribed threshold were retained as relevant significant attributes. The significant model attributes then served to establish a final single ensemble BPNN-AEA model in which weights of the set of ensemble models were aggregated.

Model performance, for UF membrane resistance and backwash efficiency, evaluated over a wide range of operating conditions and coagulant dosing strategies, revealed excellent performance even for cases of temporally +variable water quality (i.e., R2 within the ranges of {0.9 – 0.946} and {0.91 – 0.972} for post-backwash resistance and backwash efficiency, respectively). Model performance was also evaluated during a storm event in which water feed turbidity and Chlorophyll a concentration varied over the range of 1.5-19 turbidity units and 43-142 mg/L, respectively. Predictive performance of the ensemble model was respectable with R2 and RMSE of 0.92 and 0.012, respectively.

In summary, the level of performance attained with the current machine learning modeling approach, which is well suited for describing the dynamics of UF operation, should prove useful for (a) forecasting UF performance due to anticipated variations in water quality, (b) identifying and quantifying deviation of UF performance from intended baseline performance, and (c) providing a basis UF model-based control for self-adaptive operation.