Organic Rankine cycle (ORC) has been widely deployed for waste-heat recovery; however

their strong nonlinearity

pronounced time delays

and multivariable coupling make control-oriented optimization highly dependent on efficient dynamic simulation models. Although physics-based mechanistic models provide strong interpretability and high accuracy

their computational burden and limited real-time capability hinder their use in online optimization and intelligent control. To address this issue

this study proposes a data-driven modeling approach based on LSTM (long short-term memory) that maintains high fidelity while substantially improving the computational efficiency of dynamic models for complex thermal systems

thereby enabling real-time prediction and operational optimization for on-site units. The results show that the developed model achieves high-accuracy prediction of ORC variables (R2＞0.97) and improves simulation efficiency by approximately 138-fold compared with conventional mechanistic models. Furthermore

random multi-condition optimization studies quantify the dominant effect of split ratio on net power output

with a maximum net-power gain approaching 450 kW. Based on these findings

an online optimization framework of “heat-source identification-dynamic prediction-real-time optimization” is established

providing a transferable methodological basis for intelligent optimal control of ORC and other complex thermodynamic cycle systems.