Selective catalytic reduction (SCR) systems in coal-fired power plants must maintain low NOx emissions during increasingly frequent load changes. Variations in flue gas temperature and flow complicate ammonia-injection control and can cause NOx overshoot or excessive NH₃ slip. This study evaluates an implicit generalized predictive self-tuning controller using a coupled dynamic model of a 660 MW ultra-supercritical coal-fired power plant and its SCR system. The controller combines recursive least-squares identification with generalized predictive control (GPC) and is compared with proportional–integral–derivative (PID) control between 50% and 75% turbine heat acceptance (THA), at load-cycling rates of 0.5–2.0% $P_{\mathrm{e}0}$ min⁻¹. GPC improves NOx set-point tracking and reduces NH₃ slip over the conditions examined. During loading-down, the maximum outlet NOx concentrations are 48.43 mg m⁻³ with GPC and 65.78 mg m⁻³ with PID. During loading-up at 1.0% and 2.0% $P_{\mathrm{e}0}$ min⁻¹, GPC reduces the cumulative NH₃-slip index by 46.52% and 75.56%, respectively. The identified model coefficients vary more strongly at higher ramp rates, while the loading-down response also depends on the transient SCR inlet temperature. These results indicate that online model adaptation can improve ammonia-injection control during load-cycling.