Skip to main content

Perception and Planning for Intelligent Energy Management in Electric Vehicles

Electric Vehicle Range Estimation

What is the actual range of an electric vehicle, given a certain state-of-charge of the battery? Which destinations can the driver reach?

One of the most important factors influencing the diffusion of electric vehicles is the so-called “range anxiety”, meaning the difficulty, for the driver, to correctly estimate, knowing the battery level, if the vehicle will be able to reach a specific destination. The factors that affect the energy consumption depend both on the driving habits and the static or dynamic characteristics of the route. This research intends to provide methods and solutions for an accurate modelling of these factors and for an efficient prediction of the vehicle’s attainable range.


Driver-Specific Estimation of Electrical Vehicle Range

We approach the estimation of an electric vehicle range as a sequential decision making problem, making use of policy evaluation on a Markov decision process to represent the energy consumption in reaching a specific destination . The driver’s behaviour is analysed in terms of route preferences and acceleration profile, without any need of user intervention, and encapsulated in a life-long learning system, to allow an incremental adaptation of the energy consumption model to the specific driver.


Probabilistic Attainability Maps

A probabilistic attainability map shows, in real time, the confidence level with which every destination in the map can be reached by the driver. The confidence levels are estimated through an efficient implementation of a linear regression model, which takes into account environmental factors, the current battery level, as well as the behaviour and preferences of the driver. A life-long learning strategy of these features and their impact on vehicle range predictions guarantees a high level of flexibility and a continuous refinement of the model. Tests on an all-electric Nissan Leaf, both for the collection of the data and the evaluation of the estimated confidence bounds, show prediction accuracy in line with other state-of-the-art solutions.