An RNN-Based Adaptive Hybrid Time Series Forecasting Model for Driving Data Prediction

In this paper, we propose a hybrid time series forecasting model, named as the Adaptive Multivariate Exponential Smoothing-Recurrent Neural Networks (AMES-RNN), which enables accurate prediction for time series data with non-seasonal and additive trend characteristics. The AMES-RNN follows a hybrid approach in which each of the statistical and deep learning models predicts particular time series components and then merges their output. To enhance prediction performance, the optimal smoothing coefficients of the Exponential Smoothing (ES) model are estimated and updated online. Here, the coefficient estimation is performed through a deep learning-based regression model, and a method for training the regression model is presented. In addition, the prediction model utilizes future-implying information as additional input if available in order to improve prediction accuracy. The effectiveness of the proposed model was validated through multistep forecast tests using vehicle driving data that has non-seasonal and additive trend characteristics. The results show that the prediction accuracy of the proposed model was improved at least 23.0% compared to those of the existing prediction model. Additionally, we demonstrated that AMES-RNN requires low computational resources, making it feasible to perform online predictions.