Deep Reinforcement Learning-Based Intelligent Water Level Control: From Simulation to Embedded Implementation.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Embedded Systems

Background:

• Traditional water level control systems often struggle with dynamic and nonlinear conditions.
• Implementing advanced control algorithms on resource-constrained hardware presents significant challenges.

Purpose of the Study:

• To design, simulate, and implement an intelligent water level control system using Deep Reinforcement Learning (DRL).
• To validate the real-time performance and adaptability of a DRL-based controller on a low-cost embedded platform.

Main Methods:

• Utilized the Deep Deterministic Policy Gradient (DDPG) algorithm for training actor-critic neural networks in a MATLAB simulation.
• Deployed the optimized control policy onto an Arduino Uno microcontroller for real-time embedded implementation.
• Evaluated controller performance against external disturbances and sensor noise.

Main Results:

• The DRL-based controller achieved a steady-state error of <0.05 cm and 16% overshoot in physical implementation.
• Demonstrated a 22% improvement in tracking accuracy compared to conventional Proportional-Integral-Derivative (PID) control.
• Successfully adapted to external disturbances and sensor noise in real-time.

Conclusions:

• Deep Reinforcement Learning, specifically DDPG, is feasible for real-time intelligent water management on low-cost embedded systems.
• The proposed architecture is suitable for Internet of Things (IoT)-based water management, smart agriculture, and distributed sensor networks.
• This work highlights the novelty of deploying DRL on resource-constrained microcontrollers for robust control applications.