Path loss is important for the unmanned aerial vehicle (UAV) placement, trajectory optimization, and power allocation in UAV-aided communications. By considering both the factors of path delay and reflection angle (RA) in the non-line-of-sight (NLoS) paths, a new machine learning-based delay–angle-joint path loss prediction method for UAV mmWave channels is proposed. The one-input back propagation based neural network (BPNN) and two-input BPNN are built to predict the path power for line-of-sight (LoS) case and NLoS case, respectively. Meanwhile, a data acquisition method is developed to obtain massive data set for training the BPNN. According to the geometric information of digital map, a calculation method for path delay and RA is also proposed to drive the BPNN. The proposed method is simulated and analyzed based on ray-tracing (RT) simulated data under a typical urban scenario at 28 GHz. The 2D relationship of power–delay for the LoS case and 3D relationship of power–delay–RA for the NLoS case are obtained through the trained BPNN, which are well consistent with the validation set of RT data and outperform the traditional methods, i.e., 3GPP model and exponential model. Moreover, it’s found that the path power rapidly decreases when RA is 65–75 degrees under the simulation scenario, which could be an important reference for transceiver placement and route planning to reduce the impact of NLoS paths in the UAV channel.