In present， the measured location of the near-horizontal directional drlling trajectory is lag behind the bit， and the actural parameters of the delayed area can not be obtained in time， thus artificial prediction should be made for the next trajectory adjustion. In order to decreased the human facts and improve the accuracy of the prediction， a forecasting model is established based on BP neural network which is used for controlling underground directional drilling trajectory in tunnel. The model is a four-layer BP neural network， and it chooses 11 input parameters and 2 output parameters which are changed from 13 borehole space and trajectory controlling parameters from 12m before MWD including dip angles and azimuths etc. The parameters of the net forecasting model is obtained using 502 groups of training data from 6 boreholes in different mining areas. Then the forecasting results of the 12 groups of test data are compared with that of the artificial experience from 24 technicians. The results show that the mean absolute error of the downhole space parameters i.e. dip angle and azimuth are only 0.51° and 0.68° predicted inrespectively by the logsig activation function and the double-hidden-layer BP neural network which has the point structure of 9×6， and the prediction error obeys normal distribution. The accuracy prediction results derived from the BP neural network model is 35% lower than that from the technicians who work more than 5 years， and the effect from the field application is satisfied which meets the needs of drilling trajectory control. The research offers theoratical and practical base for the intelligent directional drilling work.