(199a) Machine Learning Modeling of Intermittent Operation of Reverse Osmosis Wellhead Water Treatment and Desalination Systems Via Transfer Learning and Long Short-Term Memory

Machine learning models have been increasingly explored to describe the operation of Reverse Osmosis (RO) water purification and desalination systems. However, RO membranes design and operational modes are not standardized and thus differences exist among field deployed RO plants with respect to plant operational protocol and configuration. Given the increasing interest in the deployment of distributed inland water treatment/desalination systems, it is of particular interest to explore of system operation, via model development via Deep Artificial Neural Networks (ANNs) whereby a model from a given desalination system can be leveraged to accelerate the development of an operational model for a similar operating system. In particular, the above approaches can provide significant advantage at the early stages of system commissioning/operation, and/or when a given system is relocated to treat another source water and data for model training is likely to be insufficient in the early stages of system commissioning. To address the above challenges, Transfer Learning (TL) methodologies can be utilized, whereby knowledge obtained from a unique ANN is transferred to the remaining nets reducing the ANN design time and improving the ANN model accuracy. Accordingly, the present work introduces a modeling approach that integrates TL, deep learning (DL), and a similarity-based long short-term memory (S-LSTM), with transfer learning technique to describe the unsteady-state operation of different (and geographically separated) field RO systems treating impaired groundwater. The Maximum Mean Discrepancy (MMD) and Kullback-Leibler divergence were utilized to extract prior similarity knowledge as physical constraints for weight updating (the weights for data from the different systems), which guided the process of training the S-LSTM model. In the modeling process, Transfer AdaBoost (TrAdaboost) approach was combined with S-LSTM to utilize the learned features (weights of the links within S-LSTM model trained based on data for about 200 days of system operation) according to the pairwise similarity computed among model attributes. The above learned features (i.e., body of the existing S-LSTM model without model head) were then reused in the new S-LSTM model of the same structural configuration for a new RO system, where data availability was for a period of two weeks, to optimize the model weights. The above approach resulted in significant improvements in forecasting accuracy and stability owing to the implemented TL strategies combined with S-LSTM. TL from one system to another proved to be highly effective even though operating conditions (e.g., applied pressure, permeate flux, salt passage and temperature) for the two RO systems were different. The forecasting performance of the model for the new system, with weights optimized based on an unseen operational dataset of a few weeks, was up to R² of 0.88 and average absolute relative error (AARE) of 5.7% compared with R² of 0.61 and AARE of 19.3% for S-LSTM model based on merely the new system’s limited dataset. The TL approach can be useful for early sensor fault detection for newly deployed systems, product quality control, and optimization of system operation in its early deployment stage, by relying on data-driven models from existing systems of similar configurations thus overcoming the constraints of limited data availability for the new systems.