Comparative Study of Forecasting Schemes for IoT Device Traffic in Machine-to-Machine Communication

Abstract

We present a comparative study of Autoregressive Integrated Moving Average (ARIMA) Multi-Layer Perceptron (MLP) 1-Dimensional Convolutional Neural Network (1-D CNN) and Long-Short Term Memory (LSTM) models on the problem of forecasting the traffic generation patterns of individual Internet of Things (IoT) devices in Machine-to-Machine (M2M) communication. We classify IoT traffic into four classes: Fixed-Bit Periodic (FBP) Variable-Bit Periodic (VBP) Fixed-Bit Aperiodic (FBA) and Variable-Bit Aperiodic (VBA). We show that LSTM outperforms all of the other models significantly in the symmetric Mean Absolute Percentage Error (sMAPE) measure for devices in the VBP class in our simulations. Furthermore we show that LSTM has almost the same performance in this metric for the FBA class as MLP and 1-D CNN. While the training time per IoT device is the highest for LSTM all of the forecasting models have reasonable training times for practical implementation. Our results suggest an architecture in which an IoT Gateway predicts the future traffic of IoT devices in the FBP VBP and FBA classes and pre-allocates the uplink wireless channel for these classes in advance in order to alleviate the Massive Access Problem of M2M communication.