In order to improve the ability of transformer differential protection to identify inrush current and internal fault current, a new method based on generalized regression neural network (GRNN) is proposed. First use the full wave Fourier algorithm calculated differential current characteristics as training samples, then use the cross validation method to find out spread parameter’s optimal value of the GRNN neural network, meanwhile the best input and output value of the training sample are calculated. A neural network for identifying inrush current is constructed by these parameters, the simulation results show that: the GRNN neural network has a good convergence, fast calculation and prediction output precision is very high, can accurate, effective and rapid identification of inrush current and internal fault current.