In order to overcome the defects of high complexity and probability of falling into the local minimum occurred in the present neural network (NN) based predistortion methods, a new predistortion method based on real-valued focused time-delay fully connected cascade neural network (RVFTDFCCNN) trained by the joint algorithm of orthogonal differential evolution krill herd (ODEKH) and Neuron-by-Neuron (NBN) was proposed. Both the predistorter and the inverse estimator in the predistortion system were modeled by RVFTDFCCNN. The initial parameters of RVFTDFCCNN were first obtained by the ODEKH algorithm with strong global search ability. Then the RVFTDFCCNN was further trained by the Neuron-by-Neuron (NBN) algorithm. According to the chain rule of complex function partial derivative, the calculation of the Jacobian matrix in the NBN algorithm was optimized at two levels. In the simulation, the DTMB wideband signal was used as the input signal of the predistortion system. The simulation results showed that, with the same requirement of the training error and the generalization error, the computation amount of the NBN algorithm in training RVFTDFCCNN was considerably reduced than that in SHL neural network. The ODEKH algorithm converged faster than the traditional krill herd algorithm. The joint ODEKH-NBN training algorithm had higher training accuracy than the Levenberg-Marquardt(LM) training algorithm by one amplitude order, and outperformed the latter algorithm by 2dB in terms of adjacent channel power ratio (ACPR) of the HPA output signal after predistortion. Thus it was concluded that the proposed predistortion method owned low complexity and good predistortion performance.
