Electric Vehicle Charging Profile Forecasting Using Hybrid Models
Riccardo Ramaschi, Mario Paolone, Sonia Leva · 2026
Read Electric Vehicle Charging Profile Forecasting Using Hybrid Models by Riccardo Ramaschi, Mario Paolone, Sonia Leva (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-18443.txt.
Introduction
Riccardo Ramaschi, Mario Paolone, and Sonia Leva propose a hybrid, lightweight method in a 2026 paper to predict EV charging profiles before and during charging, at both single-charger and fleet levels. Single-charger curves remain little studied, yet battery-aware planning could use them to refine fleet forecasts and departure-time guesses. Tested on multiple cars from a public dataset, the method also gauges how varying data access shifts the time-shifted charging curve.
Catalog record
| CREATOR | Riccardo Ramaschi, Mario Paolone, Sonia Leva |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 21970 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.18443
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Electric Vehicle Charging Profile Forecasting Using Hybrid Models Riccardo Ramaschi, Mario Paolone, Sonia Leva Abstract Electric Vehicle (EV) fast charging stations require forecasting techniques both at the single charger level and aggregated level. While for the latter several models exist, forecasting individual EV charging profiles is still underexplored in literature. However, such methods may be potentially used by battery-aware scheduling, leading to a more granular update of the charging station aggregated forecast and provide a more accurate estimation of EVs departure times. Nonetheless, the variable extent of available information in time and in different settings could jeopardize these benefits. For this reason, we propose a hybrid and lightweight method to estimate the EV charging profile before and during the charging process. Besides evaluating this method on multiple EVs from a public dataset, we also assess the impact of different level of information in the time transposition of the charging profile. I Introduction Global Electric Vehicle (EV) penetration is constantly increasing, recording a 25% additional sales in 2024 with respect to 2023 [ 9 ] . This growth is supported by the parallel expansion of the charging infrastructure, which experienced a 28% year-on-year increase in charging points in 2025 [ 7 ] in the European Union, most of which are part of DC Fast Charging Stations (FCS). This kind of infrastructure, that experienced approximately a 50% increase in 2025, is mainly equipped with charging points with at least 50 kW charging rate. Since the aggregated effect of EV demand in FCS is considered of utmost importance for power distribution grids’ energy [ 3 ] and congestion [ 12 ] management, the majority of literature focus on the aggregated power demand. In fact, researchers often model the EV demand as a single time series that is fed into statistical methods leveraging historical endogenous and exogenous data [ 13 ] . On the contrary, forecasting each individual EV charging profile is underexplored in literature, despite the potential benefits of such information. In fact, predicting the EV charging profile may be used by battery-aware power scheduling [ 20 ] , to provide a more granular update of the charging station aggregated forecast [ 15 ] and to estimate the individual EVs departure times [ 17 ] . Despite some works exist on battery charging profile prediction, few focus on the EV domain and on its unique characteristics and constraints, i.e. the lack of granular and public datasets [ 4 ] and the dependency of the charging profile to the proprietary battery management system [ 21 ] . Among these, most are limited by private datasets [ 11 ] , a small set of different EV manufacturer [ 17 ] and power-constrained charging sessions, where the maximum charging rate is limited by the capabilities of the charging station [ 15 , 20 ] . To the best of the author knowledge, two works address the charging profile estimation within FCS focusing on the EV required power, thus avoiding power-constrained charging sessions. Li et al. [ 11 ] proposed a deep learning framework that initiates the charging profile estimation and further refines it throughout the charging session. Besides relying on a private dataset, the work represents a strong contribution; however, the model employs architectures that could be computationally demanding if retraining are needed over a rolling horizon. Simolin et al. [ 18 ] modeled charging profiles of heterogeneous EVs, from a public dataset [ 5 ] of a Level 3 (L3) FCS, through a simple polynomial model that still overcame existing baselines. In this work, we present and assess the performance of a comprehensive charging profile forecasting tool and we also plan to fill a gap in the charging profile time transposition. This consists in the transformation of the EV power vs State-of-Charge (SoC) into the power vs time curve. In fact, while both [ 11 ] and [ 18 ] considers the information needed to perform this transposition known, this is not always the case and it may impact the process [ 19 ] . Therefore, we also identify the impact of the information stream between the EV and the charging station in the power vs time curve definition, e.g. arrival/departure SoC, arrival/departure time and battery capacity. Our contribution is threefold. First, we outline the forecasting tool. Second, we quantify the forecasting accuracy of this model. Third, we perform an ablation-like analysis on the availability of the abovementioned features. II Methodology The individual EV charging profile forecasting tool is made of two main instances. The first one produces an initial guess on the charging profile prior to connection (Section II-A ). At this step, we assume known the arrival time t_{a} , either through a booking system or through real-time sensing. The second one provides a continuous refinement after connection (Section II-B ). At this step, considering a L3 FCS [ 10 ] , the infrastructure has access to the battery capacity, the realized charging profile \mathcal{H}_{t_{i}}=\{(SoC_{t},P_{t})\}_{t=t_{a}}^{t_{i-1}} , the user-defined departure SoC SoC_{t_{d}} , while the departure time t_{d} is inferred. The overall model and its time transposition are described in Section II-C . Fig. 1 : Overall model for charging profile estimation and its time transposition. In bold the target of the different models, in italic the potentially unknown information. II-A Unconnected EV Following the indications of [ 14 ] on ambient temperature impact at EV arrival time on charging profiles, we include T_{t_{a}} as a feature in our unconnected forecasting model, together with the EV battery capacity that we indicate as C . The objective is to estimate the charging profile, defined as the mapping between the SoC and the charging power over the full range \text{SoC}\in[0,100]\% , with a 1% resolution, that we define as \mathcal{CP}_{t_{a}}=\{P(SoC)\}_{SoC=0\%}^{100\%} . We first construct a database of historical sessions through SoC-based interpolation of the charging profiles. Since not all sessions span between 0 and 100%, we handle missing data using constant extrapolation at the boundaries. That is, we extend the profile using its first and last available values, if needed. This results in a fixed-length representation of the training dataset. We then formulate the forecasting task as a supervised learning problem, employing a Random Forest (RF) regressor targeting \mathcal{CP}_{t_{a}} . II-A 1 Capacity estimation C could be either the actual value, if promptly communicated, or an estimation. Nonetheless, we generalize by proposing a capacity estimation through the Gaussian Mixture Model (GMM), a probabilistic and parametric method used for clustering and forecast. Inspired by the work in [ 8 ] , where GMM is used as an EV forecaster over the same dataset [ 5 ] , we employ an univariate GMM with K-fold Cross Validation (CV) [ 16 ] to identify the most suitable model. While in testing phase, the identified model generates a random sample from the learned probability distribution. II-B Connected EV Once connected, we build a rolling-horizon framework to update the charging profile with a time granularity \Delta t , so that we continuously refine the unconnected first guess and the following estimations. For this purpose, to exploit the historical realizations of other EVs and with the goal of developing a lightweight solution suitable for rolling-horizon application, we propose a trivial statistical method