This article proposes a novel practical energy harvesting (EH) model-assisted deep learning framework for intelligent channel tracking. Specifically, a multiantenna wireless system is considered for energy beamforming in a nonlinear model-based EH scenario. Deep autoencoder technique is utilized for learning the channel characteristics due to nonconvexity of the channel estimation optimization problem. The performance evaluation is validated in low signal-to-noise ratio regimes, thereby providing key optimal design insights. Numerical results depict an overall performance enhancement as compared with existing benchmarks.