Forecasting Torsional Resonance in Electric Vehicles by Learning a Quantile Regressor

Electric vehicles are prone to torsional resonance when traveling across rough terrain, which can subject the drive shaft to undue mechanical stress. Conventional time-series prediction methods often fall short in predicting such resonance phenomena accurately because they primarily concentrate on the central tendencies of a distribution and neglect the tails. In this study, we propose a probabilistic forecasting framework employing quantile regression to predict multiple percentile levels for the amplitude of torsional resonances. To assess the efficacy of our framework, we constructed a specialized dataset comprising 2,000 simulator-generated vibration signals that are representative of realistic driving conditions on uneven roads. Our findings substantiate that the proposed quantile-based forecasting model can predict long-term resonance with considerable accuracy compared to conventional methods. Additionally, our results reveal that temporal convolutional networks augmented with seasonal and trend decomposition achieved superior performance across a selection of five baseline models applied to our dataset.