Good driving practices, encompassing actions like maintaining smooth acceleration, sustaining a consistent speed, and avoiding aggressive maneuvers, can yield several benefits. These practices enhance energy efficiency, reduce accident risks, and significantly lower maintenance costs. Consequently, the presence of a system capable of providing actionable insights to promote such driving behavior is crucial.
Addressing this need, the Drive-GPT model is introduced, representing an AI-based generative pre-trained transformer. Within this study, the transformative potential of deep learning networks, specifically based on transformers, is showcased in capturing the typical driving patterns exhibited by individuals in diverse road, traffic, weather, and vehicle health scenarios. The model's training dataset comprises an extensive 90 million data points from multivariate time series originating from telematics systems in 100 vehicles traversing eight distinct Indian cities over a six-month span.
These pre-trained models offer substantial utility for downstream applications, including the computation of driving scores, generation of driving recommendations, and the classification of driving behavior as either proficient or suboptimal. The performance evaluation on test data indicates commendable results, with a coefficient of determination (R-squared) of 0.98 and a root mean square error (RMSE) of 0.0346. Furthermore, a discernible differentiation emerges in terms of energy efficiency and regenerative braking between good and suboptimal driving behaviors. Notably, this differentiation leads to a notable 25% improvement in energy efficiency and an 18% enhancement in regenerative capabilities.