We study a multiple-input single-output energy harvesting (EH) system, where the receiver is capable of harvesting energy from an energy source with a deterministic energy profile. Assuming the practical assumptions of imperfect channel state information (CSI) and limited-rate feedback channel, the minimum mean square error and the random vector quantized (RVQ) schemes are resorted to implement the CSI estimation and feedback, respectively. Considering the characteristic of the EH system, a power and time allocation strategy is proposed for the CSI training and feedback to maximize an upper bound of the downlink data rate. Using the optimization theory, the optimal resources allocation and the RVQ beamforming vector are obtained together with the optimal subset of training antennas. Numerical results verify the tightness of the upper bound and significant advantages of the proposed scheme over other approaches.