Online Learning Models for Vehicle Usage Prediction During COVID-19

TL;DR

This work investigates online learning for predicting the first daily BEV departure time and trip distance using a COVID-19-era fleet dataset. It compares multiple online models (QR, QKNN, QARF, MCNN) plus a historical average baseline, and augments predictions with uncertainty quantification via prediction intervals. The study introduces a per-vehicle online framework, selects well-behaving vehicles through clustering tendencies, and applies feature engineering and progressive validation to assess performance. Results show modest overall gains over baselines, with QR excelling at departure-time accuracy and QARF/QKNN/MCNN delivering better distance predictions and interval estimates, highlighting challenges in forecasting vehicle usage under pandemic-driven behavioral shifts.

Abstract

Today, there is an ongoing transition to more sustainable transportation, for which an essential part is the switch from combustion engine vehicles to battery electric vehicles (BEVs). BEVs have many advantages from a sustainability perspective, but issues such as limited driving range and long recharge times slow down the transition from combustion engines. One way to mitigate these issues is by performing battery thermal preconditioning, which increases the energy efficiency of the battery. However, to optimally perform battery thermal preconditioning, the vehicle usage pattern needs to be known, i.e., how and when the vehicle will be used. This study attempts to predict the departure time and distance of the first drive each day using online machine learning models. The online machine learning models are trained and evaluated on historical driving data collected from a fleet of BEVs during the COVID-19 pandemic. Additionally, the prediction models are extended to quantify the uncertainty of their predictions, which can be used to decide whether the prediction should be used or dismissed. Based on our results, the best-performing prediction models yield an aggregated mean absolute error of 2.75 hours when predicting departure time and 13.37 km when predicting trip distance.

Figures (6)

• Figure 1: Overview of our process for online learning model evaluation.
• Figure 2: Distribution of trips in BEV data set, with respect to driving distance and departure time.
• Figure 3: The mae per vehicle when predicting driving distance. The cars are sorted by the mae of the baseline (mean)
• Figure 4: The mae per vehicle when predicting departure time. The cars are sorted by the mae of the baseline (mean)
• Figure 5: Examples from a single vehicle on how the mae changes as more drives are observed when predicting driving distance and departure time.