Combined Particle Swarm Optimization and Reinforcement Learning for Water Level Control in a Reservoir

This article focuses on the research and advancement of an optimal system for the automatic regulation of the water level in a reservoir to eliminate flooding in the area where it is located. For example, in this article, the regulation of the level in the Mariselu Reservoir from the dam in Bistrita-Nasaud County, Romania, was considered as a practical application. Industrial PID controller tuning provides robust and stable solutions; however, the controller parameters may require frequent tuning owing to uncertainties and changes in operating conditions. Considering this inconvenience, an adaptive adjustment of the PID controller parameters is necessary, combining various parameter optimization methods, namely reinforcement learning and Particle Swarm Optimization. A new optimization method was developed that uses a mathematical equation to guide the Particle Swarm Optimization method, which in essence enhances the fitness function of reinforcement learning, thus obtaining a control system that combines the advantages of the two methods and minimizes their disadvantages. The method was tested by simulation using MATLAB and Python, obtaining very good results, after which it was implemented, which successfully prevented floods in the area where it was placed. This optimal automation system for dams should be implemented and adapted for several dams in Romania.