Reinforcement Learning-Based Control for Organic Rankine Cycles Using GP-NARX Models

The Organic Rankine Cycle (ORC) system is vital for waste-heat recovery and power generation, holding great promise for enhancing energy efficiency. However, its complex nonlinear dynamics and multiple uncertainties pose challenges to accurate modeling via traditional mechanistic analysis. Additionally, traditional Proportional-Integral (PI) controllers fall short in response speed, overshoot, and stability. To tackle these problems, this paper proposes an innovative framework. For modeling, the Gaussian Process Nonlinear Autoregressive with Exogenous Inputs (GP-NARX) model is used. It builds a dynamic model from historical data, effectively capturing time-dependent relationships without complex mechanistic analysis. In control, the PI controller serves as a baseline. Due to its limitations, a Soft Actor-Critic-Based reinforcement learning controller is designed. It can learn superior control strategies, reducing overshoot, shortening adjustment time, and suppressing disturbances. This paper provides new approaches for ORC system modeling and control, significant for improving performance and efficiency.