High-temperature reactors with pebble bed modules (HTR-PM) are designed to operate over wide conditions, which presents challenges to model-based optimization and control. The main difficulty lies in the fact that an accurate model applicable to the full operating region is usually unavailable. In order to implement significant load changes for HTR-PM, iterative strategies are developed to deal with plant-model mismatch. The model, embedded with a parameter mapping from the plant status to the model characteristics, is determined not only by parameter updates on a refined subset which ensures reliable estimation but also by parameter sensitivities to enhance the ability on prediction. On the basis of such a model, we design algorithms centered on trust region concepts, so that input corrections allowed at each iteration are determined adaptively through rigorous optimization. Moreover, the underlying model is updated adaptively to maintain a sufficiently accurate approximation to the plant. The proposed optimization strategy is illustrated with the load change problem of HTR-PM, where both reactors undergo load changes synchronously.