Improved one-step prediction accuracy does not necessarily lead to better closed-loop control in the presence of plant-model mismatch. A learning-based model-correction framework is proposed for constrained model predictive control and evaluated on a synthetic numerical benchmark inspired by pressurized water reactor dynamics. The benchmark incorporates point kinetics, fuel- and coolant-temperature reactivity feedback, partial measurements, state estimation, and control-rod actuation. Instead of directly learning a conventional one-step prediction residual, the method learns the discrepancy between the nonlinear system transition and the transition predicted by the nominal control model. This target focuses learning on the model discrepancy relevant to control decisions rather than on the propagated state-estimation-error term. The correction is applied only at the first prediction step, limiting the propagation of learning errors through the prediction horizon. Its magnitude is gradually reduced as the estimated fuel temperature approaches its upper constraint. The method was evaluated independently under combined and isolated model perturbations using paired operating scenarios. The evaluation showed substantially improved aggregate power-tracking performance with lower overall control-input variation and no increase in observed constraint violations. The improvement was not uniform across all operating conditions. Overall, the results indicate that learning a control-relevant model correction and applying it only at the first prediction step can improve closed-loop performance under plant-model mismatch.

model predictive control,model uncertainty,plant-model mismatch,learning-based control,load-following control,nuclear reactor control