This paper summarizes theories and methods utilizing historical observation data presented by Chinese meteorologists in numerical weather prediction (NWP).Not only real present evolution data is discussed but also discussed the problem utilizing historical observation data that accumulated ten years.The anthors point out that it should change the way NWP is put forward and review the process of this change from initial condition problem to evolution problem and to again inverse problem.Based on the real problem characteristic,three idealized cases were classified:(1) both the numerical model and data is exactly credible but there are parameters without observation;(2) the data is credible but there is error in model;(3) the numerical model is credible but data has error.So called the four-dimensional variational data assimilation is really belongs to the third case above,but it only utilize present evolution data and didn't utilize historical data.In real world,because both the model and data are not exactly credible,it should change forecast problem to information problem.In this way,we can assess the future weather probility by utilizing both historical observation data and physical law of atmospheric process evolution.