Adaptive reinforcement learning for energy-efficient high-recovery closed-circuit reverse osmosis

Area of Science:

• Water treatment technologies
• Artificial intelligence in engineering
• Process control and optimization

Background:

• Closed-circuit reverse osmosis (CCRO) offers high water recovery but faces control challenges due to semi-batch operations.
• Traditional rule-based control systems struggle to adapt to dynamic operational conditions.

Purpose of the Study:

• To develop and evaluate a reinforcement learning (RL) framework for optimizing CCRO operation.
• To compare the performance of RL-based control against static rule-based controllers.

Main Methods:

• A data-calibrated dynamic CCRO simulator was integrated with a proximal policy optimization RL agent.
• The RL agent was trained across 24 environmental settings, simulating over 10 million steps.
• Performance was evaluated across 375 scenarios, assessing specific energy consumption and recovery rates.

Main Results:

• The RL agent achieved a mean specific energy consumption of 0.489 kWh/m³ and a mean recovery rate of 95.5%.
• RL control outperformed static rule-based control by 13.14% in energy savings and 3.92% in water recovery.
• The simulator accurately reproduced plant behavior with low RMSE for key parameters.

Conclusions:

• The RL framework offers a promising, adaptive control strategy for CCRO systems.
• This approach enhances operational efficiency and water recovery, particularly for decentralized plants.
• Explainable AI and edge computing analyses confirm the framework's feasibility and interpretability.