Similarity-based retrieval of semantic graphs is a core task of Process-Oriented Case-Based Reasoning (POCBR) with applications in real-world scenarios, e.g., in smart manufacturing. The involved similarity computation is usually complex and time-consuming, as it requires some kind of inexact graph matching. To tackle these problems, we present an approach to modeling similarity measures based on embedding semantic graphs via Graph Neural Network (GNNs). Therefore, we first examine how arbitrary semantic graphs, including node and edge types and their knowledge-rich semantic annotations, can be encoded in a numeric format that is usable by GNNs. Given this, the architecture of two generic graph embedding models from the literature is adapted to enable their usage as a similarity measure for similarity-based retrieval. Thereby, one of the two models is more optimized towards fast similarity prediction, while the other model is optimized towards knowledge-intensive, more expressive predictions. The evaluation examines the quality and performance of these models in preselecting retrieval candidates and in approximating the ground-truth similarities of a graph-matching-based similarity measure for two semantic graph domains. The results show the great potential of the approach for use in a retrieval scenario, either as a preselection model or as an approximation of a graph similarity measure.