Application of Reinforcement Learning Algorithm for Automation of Canal Structuresâ

• Published in: Irrigation and drainage 2015, Vol.64 (1), p.77-84
• Main Authors: Shahverdi, Kazem, Mohammad Javad Monem
• Format: Article
• Language: English
• Subjects: algorithms; artificial intelligence; automation; industry; irrigation canals; learning; mathematical models; researchers; strength (mechanics)
• ISSN: 1531-0353; 1531-0361

Automation is a main option for improving the flexibility of irrigation canals. Control systems are required for this task. The main element of the control systems is a control algorithm. There are different control algorithms which have been developed. Investigation of the developed algorithms showed they have some limitations in terms of accuracy and time. These limitations make the researchers develop and introduce new and better control algorithms. Reinforcement learning (RL), which is a branch of artificial intelligence systems, is a powerful control algorithm successfully applied in robot control and in industry. In this research, a mathematical model of the RL algorithm was developed to control water depth upstream of check structures. RL learns the map of water depth‐check structures adjustment by maximizing a reward function. To test the RL algorithm, a mathematical model of a flume was used. Different scenarios of inflow increase and decrease were simulated. System response time (SRT), maximum absolute error (MAE), and integral of absolute magnitude of error (IAE) indicators were used to evaluate RL. Results were obtained and analysed. Low depth variations and performance indicators were shown. The maximum values of SRT, MAE and IAE were obtained as 252 s, 3.07% and 0.152%, respectively. Results showed the RL algorithm is a beneficial control system and can be applied in irrigation canals.