Design of the PID Controller for Hydro-turbines Based on Optimization Algorithms

• Published in: International journal of control, automation, and systems Vol. 18; no. 7; pp. 1758 - 1770
• Main Authors: Perng, Jau-Woei, Kuo, Yi-Chang, Lu, Kuan-Chung
• Format: Journal Article
• Language: English
• Published: Bucheon / Seoul Institute of Control, Robotics and Systems and The Korean Institute of Electrical Engineers 01.07.2020; Springer Nature B.V; 제어·로봇·시스템학회
• Subjects: Algorithms; Control; Control systems design; Damping; Engineering; Errors; Frequency deviation; Frequency response; Fuzzy systems; Genetic algorithms; Hydraulic turbines; Machine learning; Mathematical analysis; Mechatronics; Multiple objective analysis; Neural networks; Particle swarm optimization; Proportional integral derivative; Radial basis function; Regular Papers; Response time; Robotics; Transfer functions; 제어계측공학; integral of time-multiplied-squared-error; integral absolute error; integral gain; multiple objective particle swarm optimization; neural network; Bees; reinforcement learning; genetic algorithm; rise time; deep learning; radial basis function; frequency sensitivity; integral square-error; integral of time multiplied by absolute error; proportional gain
• DOI: 10.1007/s12555-019-0254-7
• ISSN: 1598-6446

In this study, multiple objective particle swarm optimization (MOPSO), genetic algorithm, bees, and reinforcement learning (RL) are used to calculate the rise time (tr), integral square-error, integral of time-multiplied-squared-error, integral absolute error, and integral of time multiplied by absolute error of the system transfer function and then we use a fuzzy algorithm on MOPSO, GA, bees, and RL based on the frequency sensitivity margin of a water turbine governor to optimize the proportional gain (kp) and integral gain (ki) and calculate the relative collapsing frequency response values. The MOPSO algorithm returned the optimal result. The radial basis function (RBF) neural network curve is obtained from the MOPSO algorithm with three variables (i.e., kp, ki, kd = 0.6 and grid frequency deviations values), and finally we identify and predict three variable values near the RBF neural network curve through deep learning. The result of the grid frequency deviation is close to 0, and the gain response time is better for damping the frequency oscillations in different operating conditions.