The energy landscape paving (ELP) method is a class of heuristic global optimization algorithm based on Monte Carlo sampling.By incorporating generation of initial conformation based on greedy strategy, conformation update mechanism based on pull moves and some heuristic off-trap strategies into the improved ELP method, we propose a new version of ELP, called ELP-pull moves. We test ELP-pull moves on both two-dimensional (2D) and three-dimensional (3D) hydrophobic-hydrophilic (HP) protein folding model. For ten 2D benchmark sequences of length ranging from 20 to 100, the proposed algorithm finds the lowest energies so far. Within the achieved results, the algorithm converges more rapidly and efficiently than the previous methods. For all ten 3D sequences with length 64, the ELP-pull moves method finds new lower energies within comparable computational times. The numerical results demonstrate our algorithm is a powerful method to study lattice protein folding model.