The multi-depots vehicle routing problem with time windows （MDVRPTW） was studied. The MDVRPTW model was designed，and an adaptive large neighborhood search algorithm combined with the Gaussian mixture model （GMM） clustering algorithm was proposed. By classifying the customer set before the neighborhood transformation， the initial solution was optimized， and the algorithm efficiency was improved. Six different transformation factors were used. A scoring system was used to enable the algorithm to select appropriate transformation factor adaptively at different stages of the iteration. The rationality of parameter was analyzed and three simulation experiments were designed. The experimental results verifies the efficiency of the algorithm.