Electric Vehicle Charging Profile Forecasting Using Hybrid Models

Technical Summary: Electric Vehicle Charging Profile Forecasting Using Hybrid Models

Problem Statement
The rapid expansion of Electric Vehicle (EV) adoption and Fast Charging Stations (FCS) necessitates accurate forecasting of power demand. While literature extensively covers aggregated demand forecasting, the prediction of individual EV charging profiles remains underexplored. Accurate individual profiles are critical for battery-aware scheduling, granular updates to station-level forecasts, and estimating departure times. However, existing approaches face challenges regarding the lack of granular public datasets, dependency on proprietary Battery Management Systems (BMS), and the computational demands of deep learning models. Furthermore, a significant gap exists in "time transposition"—the transformation of a power-versus-State-of-Charge (SoC) profile into a power-versus-time curve. This transformation relies on information (arrival/departure SoC, departure time, battery capacity) that is often unknown or estimated prior to connection, potentially jeopardizing forecast accuracy.

Methodology
The authors propose a hybrid, lightweight forecasting tool comprising two distinct instances to estimate charging profiles before and during the charging process. The model utilizes a public dataset of 1,878 Level 3 (L3) FCS charging sessions from the École Polytechnique Fédérale de Lausanne (DESL), augmented with ambient temperature data.

1. Unconnected Estimation (Pre-Connection):

   • Objective: Generate an initial guess of the charging profile (Power vs. SoC) before the EV connects.
   • Features: The model uses battery capacity ($C$) and ambient temperature at arrival ($T_{ta}$).
   • Capacity Estimation: Since actual capacity may be unknown, a univariate Gaussian Mixture Model (GMM) with K-fold Cross-Validation (CV) is employed to estimate $C$ based on historical distributions.
   • Profile Prediction: A Random Forest (RF) regressor predicts the full charging profile ($CP_{ta}$) mapping SoC to Power. The training data is constructed via SoC-based interpolation of historical sessions, extended via constant extrapolation to ensure a fixed-length representation.
   • SoC/Time Forecasting: To handle time transposition, GMM-based models (optimized by hour of day) predict arrival and departure SoC ($SoC_{ta}$, $SoC_{td}$).

2. Connected Estimation (Post-Connection):

   • Objective: Continuously refine the profile during the charging session using a rolling-horizon framework.
   • Method: A lightweight statistical approach based on Euclidean distances is used. At each time step $t_i$, the model constructs an input vector containing the battery capacity and the realized power history up to $t_i$. It compares this against a matrix of historical realizations to find the closest match ($j^*$) and updates the forecast using the corresponding historical profile segment.

3. Time Transposition:

   • The model converts the SoC-based profile into a time-based curve using a simple EV surrogate model. This process is evaluated under varying levels of information availability (known vs. estimated $C$, $SoC_{ta}$, and $SoC_{td}$).

Key Contributions

1. Hybrid Forecasting Tool: The paper outlines a comprehensive tool combining a supervised learning approach (RF) for initial profile estimation and a statistical distance-based method for real-time refinement.
2. Accuracy Quantification: The study quantifies the forecasting accuracy of the model, demonstrating improvements through sequential updates.
3. Information Impact Analysis: The authors perform an ablation-like analysis to identify the specific impact of information availability (capacity, arrival SoC, departure SoC) on the time transposition of the charging profile.

Results

• Forecasting Performance: The sequential refinement significantly improves accuracy. The median Envelope-weighted Mean Absolute Error (EMAE) for the unconnected forecast (with estimated capacity) is approximately 40%. After the connected updates, the median EMAE drops to roughly 16% by the 10th iteration. Beyond 10 iterations, the error stabilizes or slightly increases, suggesting an optimal stopping point for refinement.
• Capacity Estimation: The GMM-based capacity estimator successfully replicates the distribution of real battery capacities, though the error range can be wide, impacting initial forecasts.
• Time Transposition Scenarios: Eight scenarios were tested regarding the availability of $C$, $SoC_{ta}$, and $SoC_{td}$.
  • Capacity ($C$): Errors in capacity estimation lead to systematic scaling errors in the power curve and significantly affect the bias of charging time and energy metrics. Knowing $C$ shifts the distribution toward symmetry.
  • SoC Information: Errors in $SoC_{ta}$ and $SoC_{td}$ primarily affect the temporal alignment (dispersion) of the power curve rather than the scaling.
  • Combined Impact: When all variables are known, the median EMAE for the time-transposed profile drops to around 45%. The study notes that errors compound, but capacity is the dominant factor for bias, while SoC information primarily influences the spread of the error distribution.
  • Bias Observation: The model exhibits a tendency to underestimate charging time (positive time error) across most scenarios.

Significance and Claims
The paper claims to fill a gap in the literature regarding individual EV charging profile forecasting and, crucially, the analysis of time transposition under uncertain information conditions. The authors assert that their hybrid method provides a computationally efficient solution suitable for rolling-horizon applications, avoiding the heavy computational costs of deep learning architectures that require frequent retraining.

The study highlights that while individual profile forecasting is underexplored, it offers tangible benefits for grid management and battery-aware scheduling. However, the authors modestly note that the benefits of such granular forecasting can be jeopardized by the variable extent of available information. The results suggest that accurate knowledge of battery capacity is paramount for reducing systematic bias, whereas SoC information is critical for reducing dispersion in time and energy estimates. Future work is planned to integrate wider databases and benchmark the model against other state-of-the-art baselines.