Millimeter-Wave Received Power Prediction from Time-Series Images Using Deep Learning

Deep learning is applied to predict received power in a 60 GHz band propagation model from time-series images. Both three-dimensional (3D) convolutional neural network (CNN) and the combination with long short-term memory (CNN+LSTM) are used to construct predictive models. The CNN and CNN+LSTM models are evaluated to predict the received power up to 2 seconds ahead with root-mean-square errors of 2.81 dB and 2.27 dB, respectively. In addition, Kirchhoff approximation (KA) is applied as an efficient data generator to calculate the received power analytically for the corresponding propagation model. The calculated received power values are compared with measured results to validate the formulation.