Performance prediction and power regulation of organic Rankine cycle turbine with deep learning

Real-time acquisition of the aerodynamic characteristics of turbines under off-design conditions is critically important for organic Rankine cycle (ORC) systems. Deep learning methods present a promising approach to address this challenge. In this paper, a hybrid deep neural network for turbine aerodynamic prediction (Turbo-HDNN) model is proposed to achieve continuous and accurate prediction from operating parameters to physical field distributions, and further to aerodynamic performance parameters, along with a sensitivity analysis. Additionally, an improved Turbo-HDNN model, Turbo-IHDNN, is developed to predict physical fields and flow loss characteristics on typical blade cross-sections. Finally, aiming at different system requirements, fast power regulation is accomplished based on the Turbo-HDNN model and gradient-based optimization algorithms. The results demonstrate that it takes only 1 ms to generate predicted physical fields, power and efficiency for a single sample. The computational speed reaches six orders of magnitude faster than computational fluid dynamics, with superior accuracy compared to conventional surrogate models. Sensitivity analysis indicates that the power is most sensitive to inlet pressure, while efficiency is most sensitive to rotation speed, with strong coupling among multiple design variables. The Turbo-IHDNN model can accurately reconstruct two-dimensional physical fields with a maximum error of only 1.6% for the pressure loss coefficient at the rotor outlet. For four power regulation tasks, it can rapidly and robustly complete operating parameter optimization within 100 iterations while ensuring the efficiency remains above 94.65%. This study provides a novel feasible approach for aerodynamic analysis of turbines under off-design conditions and power regulation of ORC systems.