The Level-of-Detail (LOD) 2 building models defined in CityGML are used widely in three-dimensional (3D) city applications. Many of these applications demand valid solid-based geometry (closed 2-manifold), which is crucial for analytical and computational purposes. However, this condition is often violated in practice because of the way LOD2 models are constructed and exchanged. Examples of the resulting errors include missing surfaces, intersecting building parts, and superfluous interior geometry. In this study, we present a heuristic shrink-wrapping algorithm for reconstructing valid solid-based LOD2 buildings by repairing and generalizing invalid input models. A single building model is first decomposed as intersection-free and reassembled by constrained tetrahedralization. The bounding membrane is then shrunk by incrementally carving the selected boundary tetrahedra and wrapping the expected shape of the building. In the algorithm, combinations of heuristics are proposed to guide the carving process. Topological and geometrical constraints are proposed to ensure the validity and exactness of the output model. The semantics of the input geometry are preserved and missing semantics are deduced based on pragmatic rules. We evaluated the performance of the algorithm using 3D building models, including CityGML datasets. The results showed that our method achieved state-of-the-art performance at repairing 3D building models.