This paper focuses on millimeter wave (mmWave)channel prediction by machine learning (ML) methods. PreviousML-based mmWave channel predictors have limitations on re-quirements of the amount of training data, model generalizationability, robustness to noise, etc. In this paper, we propose aCNN model with a novel feature selection strategy for mmWavechannel prediction. Automatic hyperparameter tuning (AHT)algorithms are embedded in the training process to iterativelyoptimize the predictive performance of the proposed CNN. Thediversification strategy is leveraged to enhance the robustness ofthe AHT procedure against different communication scenarios. To improve the generalization ability of the prediction model, theinput features are designed to capture the correlation betweenthe physical environment and the channel characteristics. Inparallel, the Cartesian coordinates of the transmitter (Tx) and re-ceiver (Rx) are transformed into polar ones to reduce the model’ssensitivity to coordinate noise. Numerical results demonstrate theeffectiveness of the proposed CNN model in predicting mmWavechannel characteristics in various communication scenarios.