Multiple protein structure methods have been proposed for incorporating protein flexibility in molecular docking. One approach for docking ligands onto a rigid receptor is to use an ensemble of multiple rigid structures determined experimentally by X-ray or NMR spectroscopy or generated by numerical simulations. In this work we present an empirical method for generating a wide range of conformational states of a wobbling receptor using restrained Molecular Dynamics simulations (MD) and we propose a partitioning protocol for selecting a few representative conformations of the binding site from restrained MD sampling. Defining a large number of protein structures is computationally expensive when the MD simulations use an explicit solvent representation. For computational efficiency, solvent effect is therefore represented by an ensemble of restraints applied on a subset of specific atoms, using a distance-dependent permittivity function. The parameters used for the restraints and the permittivity are described. Several 100 ns restrained MD simulations are performed using different sets of parameters. In order to optimize the parameters, the results are compared to a 30 ns MD simulation in explicit solvent. Conformational sampling is speeded up by a factor of around 10 -20 when performing restrained MD simulations. A partitioning k-means algorithm is applied to select representative structures of the receptor binding site. The methodology was evaluated on the ligand binding domain of the flexible Peroxysome Proliferator-Activated Receptor-g (PPARg).