The lithology prediction of tight reservoirs is an important basic task in petroleum exploration, because lithology data are essential for formation correlation and sedimentary simulation etc. Therefore, how to obtain reliable lithology information has gradually become a hot focus on earth science research. Some important logging data especially lithology data in old oilfields have been lost due to aging equipments and wells. Newly developed production wells require huge cost in manpower and material on drilling and coring. These have brought great difficulties to the development of oilfields. Therefore a new method for lithology prediction based on date deep learning neural network was established . The logging data of oilfields could could be directly used as the input variates of neural network to get the lithology data of corresponding reservoirs without artificial pretreatments. By constructing a fully connected neural network, the tight reservoirs in Ordos basin were taken taken as the research objects to identify 5 kinds of lithologies and compare their real values. The results showed that this method didn’t need to build interpretation models and use complex calculation process, which can also be able to obtain satisfactory adaptability and prediction accuracy. The accuracy of the above lithology identification achieved to 71%, which can meet the requirement of the commercial application (70%). Therefore, this method can replace the current traditional method.