Uncertainty-Aware Transformer-Based Remaining Range Estimation for Heavy-Duty Electric Vehicles

Electrification of heavy duty on-road trucks for regional freight is a viable option for fleets to lower their carbon intensity and reduce operation and maintenance costs. However, there is considerable uncertainty in projecting their daily range because highly variable mass, among other factors, confounds battery state of charge prediction algorithms. Previous work by the authors proposed an electric vehicle range prediction model based on two parallel recurrent neural networks—one using mean-variance estimation to output a predicted mean and variance, and one using bounded interval estimation to provide bounds on the state of charge required to complete a trip. This dual RNN approach resulted in an estimate of remaining range and error bands of state of charge over the route. The previous work was limited because it did not incorporate driving conditions, like road type and ambient temperature, that affect driver behavior and, consequently, energy consumption. This work includes embedding additional contextual information about the road network and environmental conditions to narrow the error bands, which leads to more accurate state of charge prediction. The road-aware RNN algorithm is trained and tested on time series data collected from electric heavy duty trucks over 30 months. The improved algorithm shows that the bands of uncertainty in the prediction of the remaining range can be improved by 10% compared to the previous model, thereby increasing confidence in daily operation by lowering the likelihood of unexpected battery depletion. Further, interpretation of the analysis shows that the algorithm can adapt to route geometry and traffic patterns, giving fleet managers a more robust basis to plan charging schedules under daily route variability. As a result, this refined approach for contextual road-awareness accelerates the adoption of electric heavy-duty trucks by mitigating range anxiety and improving operational planning.