In this contribution, we propose an algorithm for replacing non-linear process simulation integrated in multi-level optimization of an energy system superstructure with surrogate models. With our approach, we demonstrate that surrogate models are a valid tool to replace simulation problems in multi-stage optimization frameworks and enable the improvement of their computational performance. Furthermore, we want to show that the quality of the results is not penalized and flexibility is provided to the optimization. It is desired to keep the amount of labeled data samples needed to create the surrogate model to a minimum, since their creation is computationally expensive. In our algorithm, sampling methods are used to create an initial set of data points in the input domain in the decision variables of the simulation model to be replaced. ANNs are trained on the initial training set. Using Dropout as a Bayesian approximation for quantifying the uncertainty of a prediction, the predictions can be qualified. New data points are continuously labelled and added to the training set based on the achieved prediction quality, until a minimum quality of the model is met. When applied in the optimization superstructure, the ANN can only be used when the prediction quality for the given data point is satisfying. Integrating these surrogate models in an optimization framework of an energy system will allow to only access the computationally expensive simulation when the quality of the prediction of the surrogate model is not sufficient. Simultaneously, a continuous improvement of the surrogate model will be achieved by using the created simulation results to parallelly refine the surrogate model by adding the created data points to the training set and therefore improve the model's validity range. It is found that the methodology of continuously adding the data points based on the prediction uncertainty improves the quality of the surrogate model. Initial results indicate that when applied in the optimization framework, the suggested methodology holds potential to improve computational time and flexibility.