Long-term real-time network traffic flow prediction using LSTM recurrent neural network

Real-time traffic volume prediction is vital in proactive network management, and many forecasting models have been proposed to address this. However, most are unable to fully use the information in traffic data to generate efficient and accurate traffic predictions for a longer term. Wei Cai explores predicting multistep, real-time traffic volume using many-to-one LSTM and many-to-many LSTM.

Real-time traffic volume prediction plays a vital role in proactive network management, and many forecasting models have been proposed to address this issue. However, most of them suffer from the inability to fully use the rich information in traffic data to generate efficient and accurate traffic predictions for a longer term (i.e., seven-day predictions at a five-minute interval). Wei Cai explores predicting multistep, real-time traffic volume using two types of long short-term memory (LSTM) networks: many-to-one LSTM and many-to-many LSTM by creating a flexible ensemble forecasting system that combines numbers of neural network and predictions out of interpolation. Considering the large number of data points in the current dataset, in order to address one of the typical concerns with recurrent neural network (RNN) models with respect to a longer training time, it was necessary to sample on the original dataset by each 12 data points (about an hour), and training was applied on the sampled data. Then 11 data points using interpolation between each pair of two sampled data points was done to effectively evaluate the model performance, with similar methods applied to generate a longer term of predictions at a five-minute time interval. Experimental results demonstrate that the proposed approach can effectively deal with the changing traffic pattern and show good performance in generating multistep predictions. Based on experimental comparisons, many-to-many appears to be more appropriate for sequence predictions where multiple input time steps are required in order to predict a sequence of output time steps.