Uncertainty-aware machine learning maps the real range of electric cargo vans

Electric cargo vehicles promise cleaner urban logistics, but their usefulness hinges on one deceptively simple question: how far can the battery actually take the vehicle? A new study published in Energy Informatics tackles that question not with a single-point prediction, but with seven machine learning frameworks that quantify how uncertain their own estimates of energy consumption and remaining driving range really are. Using real-world driving data from a small electric cargo vehicle, the researchers show that the choice of uncertainty quantification method involves genuine trade-offs between prediction accuracy, interval coverage and interval width, and that no single framework wins on every criterion.

The work, carried out by Mrugank Gandhi and Archana Y. Chaudhari, addresses a persistent weakness in electric vehicle prediction systems. Deterministic models output one number for energy use or range, but real driving is shaped by shifting traffic, weather, road gradients, driver behaviour and battery characteristics. The authors argue that single-point predictions may not be reliable enough for applications such as intelligent energy management, charging optimisation and route planning, and that poorly calibrated or overconfident predictions can leave drivers with inaccurate range estimates and heightened range anxiety, the fear that the battery might run out before reaching the destination or the next charging point.

To compare uncertainty methods fairly, the study used experimental data collected from a Musoshi Pop-Up Mini electric cargo vehicle driven on a 2-kilometre trajectory on the campus of Eskisehir Osmangazi University in Turkey. Data were captured at one-second intervals across 35 runs at three target speeds of 15, 25 and 35 km/h, in loaded and unloaded conditions, during both forward and return trips, and across two seasons, with 12 trips logged in summer and 23 in winter. The route includes different slopes and real driving conditions, although the authors note it is semi-controlled and does not reflect the full complexity of urban logistics environments.

Preprocessing was substantial. Measurements were synchronised to uniform one-second spacing, missing values were filled by linear interpolation, and physically inconsistent records were filtered out, including speeds below 0 or above 50 km/h and state-of-charge values outside the 0 to 100 percent range. Slopes were computed from altitude differences and horizontal distances between successive GPS points, with altitude data drawn from the Google Elevation API and slope values limited to between minus 20 and plus 20 percent. Each roughly 2-kilometre trip was divided into 50-metre segments, and segment-level features were averaged: slope, speed, acceleration, vehicle mass, outdoor temperature, battery pack voltage and motor temperature. Energy consumption per segment was derived from the change in battery state of charge, using a battery capacity of 15.6 kWh, with negative values indicating energy recovered through regenerative braking.

The seven frameworks combined three base models with different uncertainty mechanisms: Support Vector Regression with residual bootstrap, conformal prediction and adaptive conformal prediction; Random Forest variance estimation; and LightGBM with quantile regression, conformal quantile regression and adaptive conformal quantile regression. Conformal methods offer distribution-free coverage guarantees, bootstrap and forest variance capture predictive variability, and quantile methods estimate intervals directly. The adaptive variants trained an additional LightGBM regressor to predict absolute residuals of the base model, allowing interval widths to adjust to operating conditions. Hyperparameters were tuned with the Optuna framework using five-fold GroupKFold cross-validation, and trip-level grouping via GroupShuffleSplit kept samples from the same trip from leaking across training, calibration and test partitions. All conformal frameworks were calibrated to a nominal 95 percent prediction interval coverage.

On deterministic accuracy, the SVR Residual Bootstrap framework led the field, with a root mean square error of 2.358, a mean absolute error of 1.742 and a coefficient of determination of 0.893. It also recorded the best Willmott’s Index at 0.970, Skill Score at 0.893 and Explained Variance Score at 0.895. A Friedman test across all seven frameworks found statistically significant differences at the 0.05 level, and the Wilcoxon signed-rank test confirmed that SVR Residual Bootstrap significantly outperformed the SVR Conformal framework. By contrast, SVR Conformal and SVR Adaptive Conformal showed lower deterministic performance, with a Willmott’s Index of 0.924, a Skill Score of 0.756 and the highest Absolute Percentage Bias at 21.19 percent, reflecting a compromise between point accuracy and reliable uncertainty estimation.

The uncertainty results inverted part of that picture. Random Forest variance estimation achieved the highest coverage with a Prediction Interval Coverage Probability of 0.97, but at the cost of wide intervals with a Mean Prediction Interval Width of 14.98. LightGBM Conformal Quantile Regression reached a coverage of 0.96 with the widest intervals at 19.46. SVR Conformal delivered a coverage of 0.95 with a moderate interval width of 15, which the authors highlight as consistent and stable uncertainty estimation suited to real-world applications, while SVR Adaptive Conformal reached 0.94. LightGBM Adaptive Conformal Quantile Regression produced narrower intervals of 13.49 with a coverage of 0.88, and the authors describe it as better balancing coverage reliability and interval efficiency than standard CQR. LightGBM Quantile Regression gave the sharper intervals of 8.40 but a low coverage of 0.74, indicating underestimated uncertainty, and SVR Residual Bootstrap had the narrowest intervals at 6.06 with moderate coverage of 0.84.

A coverage-width criterion reinforced the pattern. Random Forest variance, at 0.3557, and SVR Conformal, at 0.3561, recorded essentially equal and the lowest values, indicating a strong balance between interval sharpness and coverage reliability, while LightGBM Quantile Regression produced the highest value at 1.8280, reflecting insufficient coverage despite narrow intervals. The authors conclude that narrower intervals came with lower coverage reliability and that methods with higher coverage tended to produce wider intervals, an inherent contradiction in uncertainty-aware prediction systems. They suggest that higher-coverage methods suit safety-critical applications where reliability matters most, while sharper, lower-coverage methods suit applications where predictive precision is prioritised.

Computational efficiency also favoured deployment. SVR Conformal had the smallest inference time at 0.0105 seconds, followed by LightGBM Conformal Quantile Regression at 0.0115 seconds, while Random Forest variance estimation was slowest at 0.0572 seconds because predictions must be aggregated across many decision trees, 176 in the optimised ensemble. All frameworks averaged inference times below 0.06 seconds, which the authors say demonstrates computational feasibility for real-time electric vehicle applications. The remaining driving range itself was estimated analytically from the segment-level energy predictions, propagating the lower and upper confidence bounds through state-of-charge and usable-energy calculations, with a minimum allowable state of charge set at 5 percent for operational safety.

The authors are candid about limitations. The study relies on a single cargo L5 vehicle model, limited seasonal and operating conditions, and a short 2-kilometre semi-controlled campus route, which may affect generalisability to other vehicle types, battery technologies, longer routes, traffic settings and climates. Conformal and quantile-based methods also depend on calibration quality and data distribution consistency, which can suffer under extreme distributional shifts. Future work, they write, will explore deep learning and sequential models such as LSTM networks, temporal convolutional networks, transformers and Bayesian deep learning, test the framework on larger datasets from different EV platforms, examine the effect of state-of-charge sensor uncertainty on range estimation, and develop online adaptive conformal systems that can handle real-time distributional shifts. In practical terms, the framework could help drivers and fleet managers obtain confidence-aware energy and range predictions for route planning, charging decisions and energy management, turning raw predictions into estimates that come with an honest measure of how much to trust them.

Subject of Research: Uncertainty quantification methods for estimating energy consumption and driving range in electric cargo vehicles using real-world driving data.

Article Title: Machine learning-based uncertainty quantification for energy consumption and driving range estimation in electric cargo vehicles

Article References: Gandhi, M., & Chaudhari, A. Y. (2026). Machine learning-based uncertainty quantification for energy consumption and driving range estimation in electric cargo vehicles. Energy Informatics, 9(1), Article 79. https://doi.org/10.1186/s42162-026-00682-1

Image Credits: AI Generated

DOI: 10.1186/s42162-026-00682-1

Keywords: electric vehicles, uncertainty quantification, machine learning, driving range, energy consumption, conformal prediction, quantile regression, prediction intervals, range anxiety, cargo EV, Support Vector Regression, LightGBM