Hybrid Deep Learning and First-Principles Model for Predictive Control of Solar Direct Steam Generation
Linear Fresnel Reflector (LFR) systems for Direct Steam Generation (DSG) exhibit strong nonlinear and transient dynamics due to diurnal solar variability, making real-time control challenging. To address this, a Gated Recurrent Unit (GRU)-based deep neural network surrogate replaces the computationally intensive first-principles LFR model and is embedded within a Nonlinear Model Predictive Control (NMPC) framework for fast prediction. Trained using scheduled sampling, the surrogate achieves prediction fits of 87.39% for steam quality and 90.55% for LFR outlet pressure. The GRU-based LFR model is integrated with a first-principles Steam Drum (SD) model in a centralized NMPC scheme to regulate steam quality and drum water level. Closed-loop simulations under varying solar insolation with cloud cover demonstrate effective tracking of time-varying setpoints, validating the proposed hybrid predictive control approach.