Abstract: To mitigate bed-temperature fluctuations caused by thermal inertia, fuel delay, and air-coal coupling during load changes in a circulating fluidized bed boiler, a low-order dynamic gray-box digital twin was developed. Recursive least squares updates the bed-temperature model online, while a proximal policy optimization agent coordinates coal feed, primary air, and secondary air. A safety layer limits action magnitude and rate. Parameterized simulations show that the adaptive twin reduces bed-temperature prediction RMSE by 17.9% relative to the static twin. Under a 345–400 t/h load step, the safety layer reduces action violations from 56 to 0, at the cost of degraded tracking performance. The closed loop therefore enables an explicit tradeoff between action safety and tracking performance.