The Prediction Model for Road Slope of Electric Vehicles Based on Stacking Framework of Deep Learning

Abstract

Accurate real-time information on road slopes and the capacity to forecast future moment gradient values are critical for the vehicle control, stability, and driving comfort. Thus, this study proposes a stacking model method for road slope estimation of electric vehicles. Gated Circulation unit (GRU), Convolutional Neural Network (CNN), and CNN-GRU are used as the base classifiers, and Multilayer Perceptron (MLP) is used as a meta-classifier. The vehicle dynamics equations are examined to select the appropriate parameters to feed into the base classifier for training. The meta-classifier is trained using the estimated results from the basic classifier. The current slope values are estimated by slicing the training set by data sampling time and windowing the training data set to predict the future slope values in 2s, 3s, and 4s. Road experiments are conducted, and error indicators are selected for evaluation. The stacking model is compared with each base classifier, Adaptive Kalman filter, Recursive Least Squares with Forgetting Factor and Back Propagation Neural Network for estimating the current moment slope, and it is verified that the stacking model can better estimate the current slope value and outperform the conventional algorithm. Comparing the stacking model with the predicted results of each base classifier for future time slope prediction shows that the stacking model is more accurate at predicting the slope values in the short future time.