On Electric Vehicle Energy Demand Forecasting and the Effect of Federated Learning

Technical Deep Dive: On Electric Vehicle Energy Demand Forecasting and the Effect of Federated Learning

Overview

This paper presents a comprehensive comparative study of well-established time series forecasting methods for the task of Electric Vehicle Energy Demand Forecasting (EDF). The authors evaluate the accuracy, privacy, and energy consumption trade-offs of centralized and federated learning approaches across four real-world EVSE datasets.

Problem & Context

The wide adoption of electric vehicles (EVs) and distributed energy resources is reshaping power grid operations. Accurate EDF is critical for managing grid stability and sustainability, enabling utility providers to anticipate demand, optimize resource allocation, and implement proactive measures. However, EDF is challenging due to factors like varying user routines, weather conditions, and unknown battery states.

The authors formulate the EDF problem as an intermittent and lumpy time series forecasting task, and examine statistical, machine learning, and deep learning methods to address it, including:

• ARIMA, SARIMA, SARIMAX
• XGBoost
• LSTM, GRU, BiLSTM, BiGRU

They also extend their analysis to the Federated Learning (FL) paradigm, which enables collaborative model training across distributed entities without sharing raw data.

Methodology

The authors propose a modular framework for training EDF models under both centralized and federated settings:

1. Data preprocessing: Converting the raw EVSE transaction data into time series format, handling missing values, outliers, and normalization.
2. Centralized model training: Training the statistical, ML, and DL models on the aggregated dataset.
3. Federated model training: Partitioning the dataset into EVSE hubs, and training the FL-compatible models (XGBoost, LSTM, GRU, BiLSTM, BiGRU) across the distributed clients.

The federated learning workflow uses the FedProx and FedXGBllr algorithms for aggregating local model updates.

Data & Experimental Setup

The authors use four real-world EVSE datasets, including three public datasets (Dundee, Palo Alto, Boulder) and one proprietary dataset (FEUP):

• Temporal range: 2011-07-29 to 2025-06-20 (3444 days)
• Total transactions: 2.4M
• EVSEs: 700+
• Locations: 130+
• EVSE hubs (for federated learning): 20+

They evaluate the models using seven performance metrics: MASE, SMAPE, MAAPE, WAPE, RMSE, MAE, R^2.

Results

Centralized Learning:

• XGBoost consistently outperforms the statistical (ARIMA, SARIMA, SARIMAX) and RNN-based (LSTM, GRU, BiLSTM, BiGRU) models across the datasets.
• XGBoost achieves MASE scores as low as 0.14, SMAPE below 0.16, and R^2 up to 0.42.
• RNNs show promise in capturing complex, non-linear patterns, but struggle with the intermittent and lumpy nature of the data.

Federated Learning:

• The optimal FL model varies across datasets: FedBiGRU for Dundee, FedProxXGB for FEUP, FedBiGRU for Boulder, FedBiLSTM for Palo Alto.
• Federated models narrow the performance gap with their centralized counterparts on some datasets (FEUP, Palo Alto), but trail significantly on others (Dundee, Boulder).
• The federated XGBoost models (FedAvgXGB, FedProxXGB) remain inferior to the centralized XGBoost, indicating the challenge of approximating a high-performing centralized model under federated constraints.

Energy Consumption & Emissions:

• Federated learning can incur up to 2.7 orders of magnitude higher energy consumption compared to centralized training, depending on the model and configuration.
• Reducing the number of local epochs per FL round ("light" configuration) can lower the energy overhead by 80.6% on average, with a modest impact on forecasting accuracy.
• The annual carbon emissions overhead of the "light" FedEDF configuration is estimated at 0.11 g CO2e, making it a more environmentally responsible solution.

Interpretation

The results highlight the importance of model selection in privacy-preserving EDF deployments:

• Centralized XGBoost remains the dominant model across all datasets, showcasing the challenge of approximating its performance under federated constraints.
• Federated neural networks (BiGRU, BiLSTM) can narrow the accuracy gap on some datasets by better capturing local sequence dynamics, but their performance remains inconsistent.
• A practical strategy is to use federated XGBoost when data is homogeneous, and switch to federated bi-directional RNNs for highly heterogeneous datasets.

The energy consumption analysis reveals the significant carbon footprint of federated learning, highlighting the need for careful configuration and efficient algorithms to minimize the environmental impact.

Limitations & Uncertainties

• The study focuses on a single forecasting horizon (next 12-hour interval), and does not explore the models' performance on longer-term predictions.
• The impact of additional features (e.g., weather, grid conditions) on forecasting accuracy is not investigated.
• The energy consumption analysis is limited to the training phase and does not consider the inference costs.

What Comes Next

The authors propose several future research directions:

• Incorporating additional feature engineering techniques to capture spatial and behavioral factors.
• Exploring alternative FL aggregation strategies that exploit the EVSE topology.
• Investigating methods to further reduce the energy consumption and carbon emissions of federated learning.