Selective Windows Autoregressive Model for Temporal IoT Forecasting

Temporal Internet of things (IoT) data is ubiquitous. Many highly accurate prediction models have been proposed in this area, such as Long-Short Term Memory (LSTM), Autoregressive Moving Average Model (ARIMA), and Rolling Window Regression. However, all of these models employ the direct-previous window of data or all previous data in the training process; therefore, training data may include various data patterns irrelevant to the current design that will reduce the overall prediction accuracy. In this paper, we propose to look for the previous historical data for a pattern that is close to the current one of the data being processed and then to employ the next window of data in the regression process. Then we used the Support Vector Regression with Radial Basis Function (RBF) kernel to train our model. The proposed model increases the predicted data’s overall accuracy because of the high relevancy between the latest data and the extracted pattern. The implemented methodology is compared to other famous prediction models, such as ARIMA and the rolling window model. Our model outperformed other models with a 9.91 Mean Square Error (MSE) value compared with 12.02, 18.79 for ARIMA and rolling window, respectively.