(94b) Methods for AI-Enabled Water Treatment

Water treatment aims to ensure the quality and sustainability of water resources, which is critical to drinking water safety, fishing, agriculture, and leisure activities. Methods for treatment of water can be classified as drinking water treatment and wastewater treatment. For drinking water treatment, water works commonly use coagulation, flocculation, sedimentation, and filtration to reduce turbidity, total organic carbon (TOC), disinfectant by-products (DBP), etc. It is also important to control the lead and copper levels during water distribution due to the corrosion of water distribution pipes. For wastewater treatment, wastewater treatment plants (WWTP) use anoxic reactions and aerobic reactions to improve the biochemical oxygen demand (BOD) and reduce total suspended solids (TSS), total dissolved solids (TDS), total nitrogen (TN) of the effluent, etc.

The water treatment process is a complex dynamic process that can be regarded as an adaptive control problem. For example, the influent water properties may be impacted by anthropogenic pollutants and seasonal shifts, resulting in different alkalinities, temperatures, pH, and dissolved oxygen (DO) concentrations at different times and requiring different degrees of treatment. Over-treatment may cause corrosion of water pipes due to DBP, and additional material cost and power consumption; while sudden events such as massive rainfalls and flash floods can overload the system and cause large changes in influent water properties, requiring quick reaction in treatment.

Two aspects in water treatment are important: sensing and control. This work surveys recent advancements in applying artificial intelligence (AI) technology (especially the deep learning models) to both aspects, and envisions new opportunities to apply AI models. Specifically, a system for water treatment often consists of cascaded subsystems with feedback control, where the various water quality variables in the influent and effluent of each subsystem can be monitored by sensors or sampled using laboratory analytical tests. For continuous monitoring and early reaction, soft-sensing is advocated by recent works to forecast future variable values based on historical readings. Models such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and even Transformers have been applied and shown promising results in forecasting accuracy. As for feedback control, deep learning models have been proposed to approximate the ideal control law represented as non-linear dynamic equations. When simulation is available, deep reinforcement learning can also be applied to learn a good control policy. AI models can be trained either with simulated data or real historical data. Benchmark simulation model (BSM) No. 1 and No. 2 are examples where simulated data can be obtained for AI model training and testing, while real data can be collected at water works or WWTPs during their operation, or even through a mobile lab such as the one owned and operated by the Birmingham Water Works Board which allows for testing out various new treatment solutions (e.g., new materials) at various water sources (lakes and rivers) by pumping water. Finally, federated learning is a promising direction to collectively learn AI models at various water works, WWTPs, and sensors deployed at different places in lakes and rivers. Federated learning utilizes the data at various sites to better train an AI model without the need of direct data sharing among the sites.