Deep Long Short-term Memory (LSTM) Network with Sliding-window Approach in Urban Thermal Analysis

The visualization of urban thermal analysis is a prevalent topic in the establishment of smart cities. With the popularity of mobile devices, large volumes of data about people and their locations are gradually accumulating at mobile base stations. In this study, we used such call details records (CDR) and long short-term memory (LSTM) networks—a kind of recurrent neural network (RNN) —to predict the future traffic of a base station. By implementing gate mechanism, the LSTM can solve the problem of exploding and vanishing gradients of ordinary RNNs. We use a sliding-window approach to transform the problem of time series forecasting into a supervised learning problem. Then, we use the proposed deep LSTM network to model the traffic of base stations, which enables the prediction of future traffic and the generation of the heat map of a city. The method we presented can decrease the root mean square error (RMSE) of the predicted access time down to 23.34 minutes per hour per base station.