Organic Rankine cycles (ORCs) convert low-temperature waste heat into electricity, offering a potential route to improving energy utilization in microreactor systems. However, heat-source fluctuations can drive evaporator outlet superheat away from its target, posing challenges for transient control. This study adopts a deep reinforcement learning method for working-fluid flow control, combining offline policy learning with online pump-speed adjustment to improve superheat responses to heat-source disturbances. Using a dynamic R134a ORC simulation model, a soft actor–critic (SAC) control scheme is developed with expert-demonstration-assisted training and safety constraints. Its performance is compared with that of a tuned proportional–integral (PI) controller under identical heat-source flow steps and pump-speed and rate limits. Preliminary simulations show that, when the heat-source flow increases from 1.50 to 3.00 kg/s, SAC reduces the superheat settling time from 55 to 26 s and the positive overshoot by 71.4% relative to PI. When the flow decreases from 2.18 to 1.50 kg/s, SAC reduces the maximum superheat undershoot by 50.1% and maintains superheat within the prescribed band of 5 ± 0.5 K throughout the response. These results demonstrate advantages in settling time and positive overshoot for the evaluated control scheme under the tested heat-source step disturbances. The findings provide preliminary simulation evidence for ORC transient control targeting microreactor waste heat recovery.

microreactor,Organic Rankine Cycle,deep reinforcement learning,superheat control