Abstract:  With the Web’s rapid development，the demands of low latency and fast response become increasingly urgent over the past few decades.In order to achieve this goal，the prefetching techniques are widely used，where documents are prefetched into caches in advance.Using prefetching techniques，we can avoid network congestion and raise access efficiency.Therefore，an effective prediction model is very essentialin the prefetching technique.Considering the necessities of high accuracy rate and practicability，we use the Prediction by Partial Match(PPM) suffix tree as a fundamental model to predict web pages.We point out some deficiencies on the side of neglect of users’ differences and the metric simplification in current cache－prefetching work.Then we present a multi－parameter web prediction model with a self－adaptation adjustment based on the user hierarchy.The main contents are listed as follows：First，we propose a user classification model based on the history access log in this paper.User behaviors are analyzed and user permutation distribution can be acquired.Then our model classifies users into different categories according to the user contribution degree distribution.The users with different contribution degree account ought to own different weights.In addition，for the users with very low contribution，we align their access web sequences and clusters them.Secondly，a method that sets the node objective function with the multi－parameter effecting is presented to construct the prediction model.The objective function involved with multiple parameters is constructed with elements related to cache replace strategies as the page accessing heat and the user classification accumulation based on the accessing frequency.And we regard the node with maximum value as one owns the strongest predictive ability.We also establish an adjustment mechanism when the prediction tree is working.So the model can learn continuously and adjust dynamically.Finally，we compare our model with several existing models through experiments.Our model has better performance on the prediction accuracy and the cache hit ratio，and we can get better results by adjusting model parameters.