This letter presents an innovative power amplifier (PA) behavioral model (BM) method valid for a range of different ambient temperatures and input power levels. This work presents a novel input image layer for a real-valued time-delay convolutional neural network (RVTDCNN). This image layer uses preprocessed ambient temperature and dissipated power. The preprocessed temperature and power as well as the present samples are placed in a central position inside the image layer. This maximizes the number of convolution operations that they are included in thereby magnifying the importance of these inputs in the feature maps. The newly proposed method delivers, in comparable conditions, a normalized mean square error (NMSE) improvement of over 3 dB compared to a previously published method.