Deep Learning Based Channel Prediction at 2-26 GHz Band using Long Short-Term Memory Network

We report a method of predicting variations in path loss using long short-term memory (LSTM) as deep learning. The training data and validation data are path loss data measured in Kanagawa, Japan, and the measurement frequencies are in the 2.2 GHz, 4.7 GHz, and 26.4 GHz frequency bands. The median data of the path loss after 1 second was predicted using 100 points of fast fading data obtained about every 0.1 seconds. The median data was derived using fast fading data of 100 points (about 10 seconds). Utilizing the prediction method using LSTM, the root-mean-square error (RMSE) for the validation data was about 2.2 dB at 2.2 GHz, about 2.1 dB at 4.7 GHz, and about 2.4 dB at 26.4 GHz. The prediction errors were improved by 1 dB or more than predictions using the latest observed values.