Forecasting Electric Vehicle Charging Station Occupancy: Smarter Mobility Data Challenge

Yvenn Amara-Ouali, Yannig Goude, Nathan Doumèche, Pascal Veyret, Alexis Thomas, Daniel Hebenstreit, Thomas Wedenig, Arthur Satouf, Aymeric Jan, Yannick Deleuze, Paul Berhaut, Sébastien Treguer, Tiphaine Phe-Neau

TL;DR

This paper presents the outcomes and winning methodologies from the Smarter Mobility Data Challenge, focused on predicting electric vehicle (EV) charging station occupancy in Paris. By framing the problem hierarchically and combining gradient-boosted trees (CatBoost/XGBoost) with clever classification reformulations, the top teams successfully navigated real-world time-series challenges like missing data and non-stationarity.

What Problem Does This Paper Address?

As the transition to electric vehicles accelerates, energy providers face a dual challenge: optimizing the electrical grid to handle charging spikes while ensuring drivers can reliably find available charging plugs. Accurate forecasting of charging station occupancy is critical to solving both sides of this equation.

However, real-world EV data is notoriously messy. This paper addresses the core difficulties of time-series forecasting in urban mobility, including high rates of missing data, sudden distribution shifts (e.g., from COVID-19 policy changes or infrastructure updates), and complex spatio-temporal correlations. Crucially, the problem is hierarchical: models must accurately predict occupancy at the individual station level, the neighborhood/area level, and the city-wide global level simultaneously.

Methods

Instead of relying on standard deep learning architectures like LSTMs or Transformers, which struggled with the dataset's short timeframe and non-stationarity, the winning teams leveraged tree-based ensembles and innovative problem reformulations:

  • Regression-as-Classification: Each Belib station features exactly 3 plugs, which can be in one of four states: available, charging, passive (connected but fully charged), or other (broken). To strictly enforce the physical constraint that the sum of all plugs must equal 3, the top teams transformed the multi-task regression problem into a 20-class classification problem, where each class represented a unique, valid combination of plug states.
  • Regressor Chains: For area and global predictions, teams utilized CatBoost regressors arranged in a chain. The prediction of the first plug state was fed as an input feature to predict the second state, capturing the inherent correlations between different plug statuses.
  • Time-Adjusted Loss Functions: To handle a major data distribution shift midway through the dataset, one winning team introduced an exponentially decaying loss function ($L_{exp}$) that gave significantly more weight to recent observations while still utilizing historical data to capture underlying trends.
  • Ensemble Aggregation: The ultimate best-performing model was a weighted aggregation of the top three teams' predictions, optimized via gradient descent using the MLpol algorithm.

Datasets

The challenge centered on an open dataset collected from the Belib network via the Paris Data platform.

  • Scope: 91 charging stations across 4 geographical zones (North, South, East, West) in Paris.
  • Duration: 7 months of data (2020–2021), sampled at 15-minute intervals.
  • Features: Temporal data (date, time of day, day of week) and spatial data (latitude, longitude, zone). Custom features like French national holidays and sine/cosine time transforms were added by participants.

Key Insights and Ideas

  • Embrace Physical Constraints: Standard regression models often predict fractional or impossible plug counts (e.g., predicting 1.5 charging plugs and 2.1 available plugs at a 3-plug station). Converting the target into discrete, valid configuration strings naturally forces the machine learning model to respect physical reality.
  • Contextualizing Missing Data: Rather than treating a sudden spike in missing values as random noise, data analysis revealed a software update on October 22, 2020, that allowed the system to flag previously unrecorded station malfunctions.
Figure 1: Percentage of missing values per day
  • Statistical Regularization: Combining high-variance machine learning models (like XGBoost) with low-variance classical models (like ARIMA) acted as an effective regularizer, smoothing out erratic predictions and improving overall generalization.

Key Results

The performance was evaluated using an $l_1$ norm penalty across all three hierarchical levels (station, area, global).

  • All three winning teams decisively beat the baseline organizer models (a naive median baseline and a standard CatBoost setup).
  • Team Arthur75 achieved the best individual pipeline score of 206.1 using a pure CatBoost and Regressor Chain approach.
  • The Weighted Aggregation of the top three teams outperformed every individual model, achieving the lowest overall error score of 196, driven by a massive performance boost at the individual station level.
Figure 3: Ranking of the competitors

Limitations and Future Work

The primary limitation of this study is the brevity of the dataset; seven months of data is insufficient to cleanly separate long-term urban trends from yearly seasonal effects (such as summer holidays). Furthermore, EV infrastructure and driver habits are evolving rapidly, meaning the dataset represents a historical snapshot.

The authors note that future iterations of this work will integrate live exogenous covariates, specifically real-time traffic data and meteorological conditions, to further refine prediction accuracy.

Summary updated

Abstract

The transport sector is a major contributor to greenhouse gas emissions in Europe. Shifting to electric vehicles (EVs) powered by a low-carbon energy mix would reduce carbon emissions. However, to support the development of electric mobility, a better understanding of EV charging behaviours and more accurate forecasting models are needed. To fill that gap, the Smarter Mobility Data Challenge has focused on the development of forecasting models to predict EV charging station occupancy. This challenge involved analysing a dataset of 91 charging stations across four geographical areas over seven months in 2020-2021. The forecasts were evaluated at three levels of aggregation (individual stations, areas and global) to capture the inherent hierarchical structure of the data. The results highlight the potential of hierarchical forecasting approaches to accurately predict EV charging station occupancy, providing valuable insights for energy providers and EV users alike. This open dataset addresses many real-world challenges associated with time series, such as missing values, non-stationarity and spatio-temporal correlations. Access to the dataset, code and benchmarks are available at https://gitlab.com/smarter-mobility-data-challenge/tutorials to foster future research.
BibTeX
@inproceedings{amaraouali2024forecasting,
  author = {Yvenn Amara-Ouali and Yannig Goude and Nathan Doumèche and Pascal Veyret and Alexis Thomas and Daniel Hebenstreit and Thomas Wedenig and Arthur Satouf and Aymeric Jan and Yannick Deleuze and Paul Berhaut and Sébastien Treguer and Tiphaine Phe-Neau},
  title = {{Forecasting Electric Vehicle Charging Station Occupancy: Smarter Mobility Data Challenge}},
  booktitle = {Accepted by DMLR},
  year = {2024},
  url = {https://openreview.net/forum?id=0wOwrH5fXQ},
}
Figure