During offline optimization, real fitness evaluations cannot be obtained. Therefore, building high-quality surrogates based on the limited amount of offline data is crucial. Many ensemble surrogates are built by combining a number of models that are trained based on different subsampled data, and these weak models without characteristics lack interpretation. In this paper, we propose an adaptive differential evolution based on a dual-scale surrogate ensemble model, termed DSEDE. The ensemble model consists of a fine scale surrogate model and a global scale surrogate model, which approximates fine landscape information and the overall trend landscape, respectively. Adaptive differential evolution is used to search for the global optimum of the offline surrogate model. Comparing four state-of-the-art offline data-driven optimization algorithms, the results show that the proposed algorithm is efficient and effective in solving offline optimization problems. DSEDE costs the least time in optimization, especially for high-dimensional problems.